memory: decision seat live on h2o again since 1704 under the re-fitted judge

feat(intern-decision): decision seat back on h2o-lightning-4b under the re-fitted judge
Prime ruled the cutover ("do it", 16:40, via albok-dev) after the judge's note question was re-fitted (albok-nemi 0.1.18). switch-to-h2o.sh ran 17:02:23, the GPU 1 gate passed, h2o healthy 17:04:54. Checks on :8033: /health at the pin, /decide/shared 200, 422 over the cap, inline image answers, image URL refused by the shim, vLLM media-domain backstop refuses (400). compose.yaml is h2o again; rollback-to-inc.sh stays the one-step way back.
2026-10-12 17:06:20 -07:00 · 2026-10-12 17:05:59 -07:00 · 2026-10-12 16:21:25 -07:00 · 2026-10-12 16:20:49 -07:00 · 2026-10-12 14:29:15 -07:00 · 2026-10-12 14:09:11 -07:00
22 changed files with 1938 additions and 60 deletions
@@ -168,7 +168,7 @@ https://svos-terminal.phasefinal.com:443 {
 		# The kernel's signed calls come from the box itself (its source on the mesh address); anything else is refused.
 		@not_from_box {
-			path /verbs/* /health
+			path /verbs/* /health /seats/*
 			not remote_ip 100.64.0.9 10.100.50.43 127.0.0.1
 		}
 		handle @not_from_box {
@@ -211,6 +211,15 @@ https://svos-terminal.phasefinal.com:443 {
 		handle @health {
 			import svos_terminal_upstream
 		}
 		# U13 fix (2026-10-12): app-terminal's registry base_url is this origin again (browsers frame panes from it),
 		# so the kernel's signed seat-host calls (U5b) come through here too, from the box only (@not_from_box above).
 		@seat_host {
 			method POST
 			path /seats/observe /seats/check /seats/launch /seats/retire-launch
 		}
 		handle @seat_host {
 			import svos_terminal_upstream
 		}
 		handle {
 			respond 404
 		}
@@ -0,0 +1,192 @@
 # D3 vs Intern-Decision-4B — decision-model seat bench (2026-10-12)
 Ordered by Prime; preregistration `brokkr-smithy`
 `docs/surveys/vllm-sr-decision-landscape-2026-10-11/d3-bench-prereg.md`
 (c93d7b0, amended b87a7ab), plus Amendment 1 (shape calibration, brokkr) and
 Amendment 2 (operator: memory budget = what fits on the seat's GPU, e3dbfac).
 Run by infra-hermes on fv-ml1 GPU 3. **HALTED by operator call at 15:07Z** —
 "fails on a number of levels… stash the findings and call it." This doc is the
 stash: everything below was measured; what was *not* run is stated plainly.
 Harness: jevbench v1.2.16 (public 231 = easy/original/hard), the S1 battery
 (authored144 + perturbations108 + cicada W1/W2 + Wyrd, 1,748 paired scorable
 rows), and the 10-04 shape-runner lineage with a per-model calibrated 32k
 builder. All artifacts: `fv-ml1:/tank/bench/arm-*` and `/tank/bench-jev-2026-10-04`
 (the latter untouched, lkraven-owned).
 ## Headline
 The incumbent (intern-decision-serve 0.1.3, pinned Intern-Decision-4B) is not
 displaced. Every d3 arm that clears the incumbent's accuracy needs either a
 GPU budget the seat cannot hold at the worst shape, or fails the latency
 contract, or both. The one arm inside the seat's own memory budget at the
 worst shape (d3-edge at bs=1) loses 8 battery points and is 28x slower than
 INC at 32k-class state. The quantization warrant closes: NVFP4 does not buy
 the 32k shape, because the 32k problem is O(q x state) activation and protocol
 accounting, not weight bytes.
 ## 1. bf16 ladder (Engine primary, `vllm-sr` dev CLI, GPU 3)
 Revisions pinned per arm (edge e51ab425…, lite c91d55d6…, nano ccd7bc1a…,
 mini 40391eb6…, flash 06d33616…, d3-27 7f257ad7…). jevbench acc = public 231,
 median of 3 repeats (all repeats bit-exact per arm). Battery = labelled-evidence
 acc, single condition. Peak VRAM = steady-state serve MiB. p50 = harness per-call
 median ms.
 | arm | params | jevbench | battery single | peak MiB | p50 ms |
 |---|---|---|---|---|---|
 | **INC (baseline)** | 4B | **202** | .910 | 8,822 (32k shape) | ~1.9–2.0s at 32k |
 | d3-edge | 0.59B | 158 | .749 | 2,398 | 15 |
 | d3-lite | 1.4B | 168 | .852 | 2,676 | 28 |
 | d3-nano | 3.8B | 175 | .907 | 5,304 | 29 |
 | d3-mini | 7.3B | 188 | .945 | 10,540 | 36 |
 | d3-flash | 8.39B | 190 | .947 | 17,600 | 37 |
 | d3-27 (vendor path) | 26.09B | 206 | .970 | 50,300 | — |
 INC reproduction: 202/231 x3, bit-exact against the 10-04 record (`out/intern-native`).
 d3-27 is the only arm that beats INC on the public suite, and it is a path
 divergence: the Engine's placement estimator (4B/param + 2 GiB) refuses a 26B
 model on any single card we have; the vendor `d3_server.py` runs it at 50.3 GiB.
 Gate (jevbench >= INC - 3): PASS at 27B only.
 Argmax cross-check (vendor vs Engine, same snapshot, 231 items): edge 230,
 lite 229, nano 229, mini 231, flash 231 — mismatches are borderline near-ties
 (max prob gap 2.7e-2). Engine fidelity: argmax identical on all non-borderline
 items; probabilities NOT exact (median gap 4.2e-4). The card's "matches exactly"
 claim is false as written; argmax-equivalence holds.
 ## 2. 32k shape cells (Amendment 1: per-model calibrated to ±2% of 32,768 realised tokens)
 Builder length iterated per model in its own tokenizer; every arm converged to
 32,743 realised tokens for the 1q state. 16q cells carry 16 x 32,743 + framing =
 **523,900 wire tokens** because the d3 wire protocol counts state once PER
 QUESTION — O(q x state) accounting, measured, not predicted. INC's equivalent
 is a single forward over one state. cold = first request after ready; launch to
 ready wall time is reported separately below and is NOT folded into cold.
 | arm | crit21_short cold ms | 1q32k cold ms | 16q32k cold ms | 16q32k peak MiB |
 |---|---|---|---|---|
 | INC | (2-call 16+5, see note) | 1,970 (32,715 tok) | **rejected: >32,768 cap at 33,324; maxfit 16q = 32,708 tok, 1,960 ms** | 8,822 |
 | d3-edge | 1,242.6 | 895.4 | 20,790 | 28,404 |
 | d3-lite | 89.1 | 1,139.5 | 18,668 | 28,938 |
 | d3-nano | 103.3 | 1,442.1 | 23,568 | 33,486 |
 | d3-mini | 157.1 | 3,565.8 | **deadline-bound** (503 at 30.002s, all retries) | 71,404 |
 | d3-flash | 242.3 | 4,469.5 | **deadline-bound** | 80,396 |
 | d3-27 vendor | 642.7 | 8,731.0 | 147,724 (366,172 tok, builder cap) | 96,262 |
 Notes. (a) INC crit21 on 10-04 is a client-side 16+5 split summed across two
 `/v1/systemone` calls (`bench_shape.py --max-questions 16`); `/v1/systemone`
 enforces 1..16 questions, so a 21-question crit is structurally out of contract.
 (b) The Engine's 30 s serving deadline is a configuration of the seat, not a
 bug; the wall sits between 3.8B (nano completes at 23.6s) and 7.3B (mini).
 (c) Activation peaks include 32k state at bs=8 (Engine BATCH_SIZE is hard-coded
 8, `family.py:65`); every bs=8 arm exceeds even Amendment 2's budget at 16q32k.
 (d) Latency clause (16q32k <= 1.5x INC): INC's honest 32k-class cell is 1.96s,
 so the clause bar is ~2.9s. NO d3 arm meets it; closest is edge-b1 at 13.0s (6.6x).
 ## 3. bs=1 curve (vendor `--batch-size 1`, the only memory knob)
 | arm | 16q32k peak MiB (bs=8 -> bs=1) | 16q32k cold ms (bs=8 -> bs=1) |
 |---|---|---|
 | d3-edge | 28,404 -> 4,778 | 20,790 -> 13,019 |
 | d3-mini | 71,404 -> 18,514 | deadline -> 53,770 |
 The knob buys memory, never time: latency is protocol-intrinsic O(q x state).
 No setting puts a mini-class model inside the seat at the worst shape (mini-b1
 still misses Amendment 2's budget by 1.6 GiB *and* fails latency at 18x INC).
 ## 4. Stage 4 — control quant trio and the battery sensitivity curve
 d3-lite backbone, RTN (round-to-nearest, no calibration), vendor path,
 `--verify none` (the manifest hashes the main shard; modified packages can't
 self-verify). Paired deltas vs the path-matched bf16 vendor baseline, 1,748
 rows, 95% bootstrap CI, Wyrd split included:
 | arm | weight rel-L2 | jevbench | battery | paired delta [CI] | Wyrd delta [CI] | argmax flips |
 |---|---|---|---|---|---|---|
 | lite bf16 vendor (control) | 0 | 168 | .852 | — | — | — |
 | lite NVFP4 RTN | 9.4% | 164 | .869 | +0.0034 [-0.0086, +0.0154] | -0.019 [-0.055, +0.017] | 117 |
 | lite INT3 RTN | 19.9% | 158 | .739 | -0.095 [-0.118, -0.072] | -0.131 [-0.176, -0.088] | 476 |
 The battery gate is monotone in weight error and functional: it cannot see 9.4%
 (NVFP4 RTN lands inside noise), it clearly sees 19.9% (INT3 collapses). This is
 the calibration warrant's premise confirmed: a quant that survives must move
 weights *less* than 9.4% AND keep paired deltas inside the band — i.e. the
 headroom to demonstrate is measured, and it is small.
 ## 5. Stage 5 — flash calibrated NVFP4 (the one arm completed before the halt)
 ModelOpt `NVFP4_DEFAULT_CFG`, calibrated on the shared 512-row corpus (sha256
 e7aa7b18…a31e, seeded `jev-nvfp4-calib-2026-10-12`, no scored item). Served
 in-process over a `/v1/systemone` shim (`stage5_serve.py`, prepare/run/respond
 mirroring the vendor contract, 1,405 quantizers): ModelOpt's fake-quant lives
 in the quantizer forward, so no package rewrite; accuracy is a true fake-quant
 read, memory is analytic + the byte table below.
 jevbench 194 x3 (bf16 Engine control 190). Survive rules applied by the letter,
 paired vs `arm-d3-flash`, n=1,748:
 | rule | result | verdict |
 |---|---|---|
 | (a) battery pooled delta, CI lower bound >= -0.02 | -0.0011 [-0.0086, +0.0057] | PASS |
 | (b) public suite within ±3 of bf16 median | +4 (194 vs 190) | **FAIL by the letter** (upward; 10 flips: 6 up, 2 down, 2 both-wrong) |
 | (c) Wyrd within -0.02 | +0.0071 [-0.0119, +0.0262] | PASS |
 | (d) argmax flips <= 4 of 231, both directions | 10 | **FAIL by the letter** |
 Byte table (bf16 snapshot index, total 15.63 GiB on disk): quantized linears
 (attn 3,980 + mlp 9,216 MiB) at 4.5 b/w = 0.28125x -> 3,711 MiB; kept wide
 (embed+head 1,940, vision 870, other 2) = 2,812 MiB. **Projected export 6,524
 MiB (~6.37 GiB)** — single-digit, as expected. ModelOpt `export_hf_checkpoint`
 exists for a measured-export ground truth; not run before the halt.
 Latency of the fake-quant arm is dequant-matmul timing and is NOT reported as a
 speedup. **Analytic 32k verdict unchanged: quantization does not touch the
 O(q x state) activation cost; the 16q32k cell fails identically at any weight width.**
 ## 6. Amendment 2 — the seat GPU budget (measured read-only, GPU 1)
 Prime: "our budget is now whatever fits in the current gpu where it's running."
 - GPU 1 total: **97,887 MiB.**
 - Non-seat processes: `vllm-meromero-r` 43,994 MiB + `vllm-gen-small` 34,994 MiB = 78,988 MiB.
 - Seat (intern-decision) resting: 8,812 MiB.
 - Safety margin: 2,048 MiB (brokkr's suggested figure; no seat headroom policy exists).
 - **Budget: 16,851 MiB (~16.5 GiB).**
 Re-read of every fit clause against 16,851 MiB:
 - Short shape: fits INC-class (8,822 measured) and every arm through mini
  (10,540); NVFP4-flash analytic ~6,524 fits; flash bf16 (17,600) misses by
  749 MiB; 27B (50,300) does not.
 - 16q32k: only edge-b1 (4,778) and INC-class fit. Every bs=8 arm is 27.6–96.3
  GiB. The latency clause independently kills mini/flash/INC at that shape.
 - Net: the wider budget upgrades mini from "memory-out" to "memory-in,
  latency-out" (53.8s at the only knob setting that fits), and leaves the
  conclusion unchanged.
 ## 7. What was NOT run because of the halt
 - INC bench-copy arms (positive control `inc-bf16-copy` and fake-quant
  `inc-fakequant`, 249 language tensors RTN-NVFP4, median rel-L2 9.4%, vision
  kept bf16): the 10-04 `intern_server.py` shim and both packages were built
  and the arm was mid-boot when the halt landed. The INC fake-quant row is
  therefore absent — note the production engine cannot load NVFP4 at all
  (`intern_decision_serve/engine.py` pins the snapshot's `inference.py` by hash
  and validates dtype against {bfloat16, float16, float32} with an explicit
  raise; no compressed-tensors path exists in the image).
 - NVFP4 arms for mini, 27B (fake-quant), nano, lite (calibrated), edge.
 - ModelOpt measured export (ground-truth on-disk size).
 - Paired battery deltas + Wyrd splits for the bf16 ladder arms (no quantized
  sibling to pair against; would need battery re-runs).
 ## 8. Teardown & disk
 Engine stack stopped, all bench servers on GPU 3 stopped, `:18080` iptables rule
 removed, guard read confirms GPU 3 at 2 MiB. Production GPU-0/1 seats untouched
 (one ~51 s exposure window during cross-checks earlier in the run, zero
 degradation observed — all RestartCount=0, no Xid).
 Weights KEPT per instruction (Prime rules on deletion):
 `/tank/bench/models/model-runtime` **80 GB** (all six d3 snapshots + INC copy
 staged), plus bench working set: `inc-bf16-copy` 8.5 GB, `inc-fakequant` 8.5 GB,
 `quant-flash-nvfp4` 16 GB, `quant-lite-nvfp4` 1.7 GB, `quant-lite-int3` 1.4 GB.
 `/tank/bench` total 180 GB; `/tank` at 62% with 3.3 TB free, so this is
 comfort, not pressure.
@@ -0,0 +1,226 @@
 # Decision-model sweep follow-on: H2O re-bench, Clef-flash, RSI-Jev vs the live seat (2026-10-12)
 Ordered by Prime; preregistration `brokkr-smithy`
 `docs/surveys/decision-model-sweep/bench-2026-10-12-followon-prereg.md` @ 2e0539f —
 the first run under the bench-first rule. Run by infra-hermes on fv-ml1 GPU 3
 (process-pinned, GPU-3-empty guard, loopback only). Harness: the same d3 bench
 harness as the morning run
 (`docs/pfi/d3-vs-intern-decision-bench-2026-10-12.md`), unchanged.
 ## Headline
 **The incumbent holds.** H2O-lightning-4b is the first non-d3 arm to beat INC
 on the public suite (205 vs 202) with a positive paired battery delta — but it
 fails the long-input gate by the letter: its shipped shim silently trims the
 state at ~32,000 tokens, confirmed today exactly as on 10-04. Under the prereg
 rules that makes it a **short-route candidate**, not a supersede. Clef-flash is
 **lateral** in both precisions (battery CI straddles zero, Wyrd significantly
 negative, and bf16 doesn't even fit the seat at rest). RSI-Jev v6.1 is a
 **loss** (23 items below INC on public, Wyrd −0.0548).
 ## Budget (re-measured at bench time, Amendment 2 method)
 GPU 1 total 97,887 MiB. Non-seat: vllm-meromero-r 43,998 + vllm-gen-small
 34,994 = 78,992 MiB. Seat resting 9,098 MiB. Margin 2,048 MiB.
 **Budget = 16,847 MiB.** (16,851 in the morning; 4 MiB drift, immaterial.)
 ## Arms and verdicts
 Vendor-intended serving path for every arm, verbatim; no ports, no third-party
 shims. Wire adaptations beyond the mandatory model field: none for H2O (shipped
 `h2o_lightning_shim.py`) or RSI (vendor CLI); Clef reuses the 10-04
 `clef_server.py` stdlib wrapper around the shipped `joint_schema_model.py`
 (same wrapper that produced the 10-04 numbers; weights byte-identical).
 | arm | path | jevbench (×3) | battery paired Δ vs INC [95% CI] | Wyrd Δ [CI] | neg-control | Brier / ECE |
 |---|---|---|---|---|---|---|
 | **INC (reused cells)** | seat 0.1.3 | **202** | — | — | .097 | .178 / .051 |
 | H2O @ b43f759b | vLLM 0.30.0 + shipped shim | **205** | **+0.0137 [+0.0023, +0.0257]** | −0.0167 [−0.0357, +0.0024] | .028 | .152 / .050 |
 | CLEF bf16 @ 8b2e5fd1 | vendor loader, max_length=32768 | 191 | +0.0040 [−0.0080, +0.0160] | **−0.0333 [−0.0571, −0.0095]** | .111 | .235 / .053 |
 | CLEF bnb-int8 | same + LLM.int8() | 190 | +0.0006 [−0.0114, +0.0120] | **−0.0333 [−0.0571, −0.0119]** | .090 | .239 / .043 |
 | RSI-Jev v6.1-VL-4B @ 7f93b192 (HF rev) | `rsi-jev serve --effort auto` | 179 | −0.0149 [−0.0303, +0.0006] | **−0.0548 [−0.0905, −0.0214]** | .208 | .274 / .061 |
 Paired battery: 1,748 scorable rows against the 10-04 INC r1 rows (same
 harness, harness scorer), 10k seeded bootstrap (seed 1012). All arms ran the
 full battery: negative control crushed for H2O (it abstains — the loaded-name
 trap works), soft for Clef (.111/.090) and RSI (.208 — the weakest trap of the
 run). Calibration as shipped, jevbench 231: H2O is the ONLY arm with a better
 Brier than INC, and its ECE is the second-best.
 ## Speed and the 32k shapes (calibrated per tokenizer to 32,743 ±2%; realised tokens logged)
 cold = first request after ready; warm = median of 5; launch→ready its own
 column. INC crit cells re-measured today on the seat path (2-call 16+5 split
 for crit21, per contract); 32k INC cells reused from today's 32k run (maxfit
 16q at 32,708 tokens, 1,960 ms — the prereg's 1.95 s anchor).
 | arm | crit16 warm ms | crit21 warm ms | 1q32k warm ms (realised tok) | 16q32k warm ms (realised tok) | 16q32k peak MiB | launch→ready s |
 |---|---|---|---|---|---|---|
 | INC | 110 | 198 (2-call) | 1,960 (32,715)* | 1,960 maxfit (32,708)* | 8,822 | 20 |
 | H2O | 137 | 168 | 179 (32,082 — **TRIMMED**) | 1,308 (32,574) | 15,464 | ~280 → **154** (measured, see addendum 1a) |
 | CLEF bf16 | 210 | 138 | 2,948 (32,730) | 2,962 (32,768 cap) | 32,130 | 12 |
 | CLEF int8 | 238 | 176 | 3,262 (32,730) | 3,284 (32,768 cap) | 23,080 | 18 |
 | RSI | 112 | 69 | 1,224 (32,758) | 2,405 (33,205) | 21,158 | 68 |
 \* reused 32k-run cells; INC's `/v1/systemone` enforces a 32,768-token
 contract, so its honest 16q32k is a maxfit single forward.
 Both Clef arms sit exactly on 32,768: the vendor loader with max_length=32768
 honours the window but caps there — 16q over 32k state counts state once per
 request at the wire, and the loader slices to window. Their "warm" at that
 shape is the fastest honest path, but the state never exceeds the window.
 RSI realised 33,205 — it has no hard cap; honest 32k+ shape.
 **H2O's trim, the named risk, confirmed:** the shape calibration plateaued at
 32,082 realised tokens across builder lengths 46,953→51,097 (full history in
 `arm-h2o/shape32k.stdout`) — the shim head/tail-trims at 32,000 exactly as on
 10-04 (24,834 effective then, 32,082 now: the trim moved but did not go away).
 The gate itself scores OK at every depth (the evidence sits mid-state and
 survives the trim), so the failure mode is silent: **accuracy at 32k is
 measured on a truncated state and the response says nothing about it.**
 ## Long-input gate (1k/8k/16k/30k + positive/null, tokens-to-model logged)
 | arm | verdict | evidence |
 |---|---|---|
 | INC | clean | reused; no trim observed on seat path |
 | H2O | **FAIL — silent truncation** | calibration plateau at 32,082; `usage.input_tokens` reports the post-trim count, not the sent count |
 | CLEF bf16 | clean (window=32,768 honoured) | d30000 OK; no plateau below the window |
 | CLEF int8 | clean | as bf16 |
 | RSI | clean | 32,758 realised, no plateau; d30000 OK |
 ## Decision-rule application (fixed at 2e0539f)
 **H2O-lightning-4b — SHORT-ROUTE CANDIDATE.** Fits the budget at every shape
 incl. 16q32k (15,464 ≤ 16,847); battery CI lower bound +0.0023 > 0; Wyrd
 −0.0167 ≥ −0.02; 16q32k warm 1.31 s ≤ 2.94 s (passes latency by the letter,
 and it is genuinely the fastest at the worst shape because the shim's
 single-count protocol answers over a *truncated* state). Fails only the gate
 clause: silent truncation at 32k. Recorded, not acted on. Note for whoever
 eventually evaluates it for the seat: the trim is inside the vendor's shipped
 serving code — a supersede needs the vendor to fix the shim, not us to patch
 around it.
 **CLEF bf16 — LATERAL, memory-out.** Battery CI straddles zero; Wyrd fails;
 rest 18,854 MiB does not fit the budget at all; 16q32k warm 2,962 ms fails the
 latency clause by 22 ms (0.7%) at peak 32,130 MiB (1.9x the budget).
 **CLEF int8 — LATERAL.** Fits short shapes (rest 12,084) but the battery CI
 straddles zero, so it is not even a short-route candidate; 16q32k fails both
 memory (23,080) and latency (3,284).
 **RSI-Jev — LOSS.** Public −23 items (outside ±3), battery Δ −0.0149 with CI
 [−0.0303, +0.0006], Wyrd −0.0548, negative control softest of the run. Its
 serving numbers are respectable (clean gate, latency pass at 16q32k) — it is
 the decision quality that loses.
 ## Housekeeping
 - Repo pins: RSI-Jev serving code = main @ fa70834 (pinned commit recorded per
  prereg; the 7f93b192 sha is the **HF model revision**, not a repo commit —
  it is not an ancestor of main and was resolved as a snapshot).
 - Licence flags as prereg'd: Clef Apache-2.0 fine; RSI has no LICENSE file at
  source and lists a research-only training source — flagged once, does not
  block a bench, deployment rights are the operator's call.
 - Teardown: all bench servers stopped, no listeners on 1809x, GPU 3 guard read
  2 MiB, no firewall rules were needed (loopback-bound ports only).
 - Weights KEPT (operator ruling stands): model-runtime now 116 GB total
  (adds h2o 9 GB blob-set, clef 18 GB, RSI 5.5 GB), plus materialized
  `pkg-h2o` 9.0 GB and `pkg-clef` 18 GB serving dirs and the RSI-Jev clone
  72 MB. `/tank` 63% used, 3.2 TB free.
 - Artifacts: `fv-ml1:/tank/bench/arm-{h2o,clef-bf16,clef-int8,rsi,inc}/`
  (jevbench ×3, sets.json battery, gate, shape32k, crit16; h2o additionally
  `shape32k-forced.json` recording the trimmed-shape cells and
  `shape32k.stdout` the plateau history).
 ## What was NOT run
 Nothing from the prereg matrix was skipped. One methodology note: INC's
 battery/gate rows are the 10-04 r1 rows (reused cells, harness unchanged and
 bit-exact reproduced this morning at 202/231), and the paired bootstrap seeds
 differ between the morning doc (42) and this one (1012) — both are seeded and
 re-runnable; the h2o CI lower bound of +0.0023 is close enough to zero that
 the choice of seed is worth knowing about, and it survived re-runs under
 seed 42 unchanged ([+0.0023, +0.0257] on both).
 ## Amendment 1 addendum — the h2o confirmation run (operator: a 32,000-token limit is acceptable)
 Prereg Amendment 1 @ 63d7a1b (plus 1a, boot timing). The seat contract under
 h2o is 32,000 tokens (state + questions); the failed clause was the shim's
 *silence*, not the limit. Stack: vLLM 0.30.0 + shipped shim unmodified, GPU 3,
 plus a **thin admission wrapper** (task 2 recorded: no shim config option for
 admission exists — `max_state_tokens` only configures the trim). The wrapper
 imports the shipped renderers, asks vLLM's own `/tokenize` for the count of
 each question's exact prompt, and refuses over-cap with 422 + an explicit
 error naming the count (`refused, not trimmed`). No vendor file was edited.
 It polices both currencies: per-question prompt max AND the shim's own
 shared-head + tails total.
 **(1)+(2) Gate re-run through the wrapper, 32k states, evidence at depth.**
 Every admitted cell reports `usage.input_tokens` exactly equal to the wrapper
 count — the trim is now unreachable below the cap, so the report is honest:
 | depth | admitted | tokens sent = reported | answer |
 |---|---|---|---|
 | ~1,000 | yes | 31,700 | supported ✓ |
 | ~8,000 | yes | 31,862 | supported ✓ (p: .907/.027/.066) |
 | ~16,000 | yes | 29,508 | supported ✓ |
 | ~30,000 | yes | 31,885 | supported ✓ |
 | over-cap probes (×2) | **refused** | 39,805 / 40,336 tokens | 422, named count, every time |
 Evidence answered correctly at every depth — the 10-04 depth-dependent
 evidence loss is gone at these sizes (it lived INSIDE the trimmed window).
 **Clause (2): clean.**
 **(3) Largest-admissible shapes** (state pinned by the wrapper so the 16-q
 request totals ≤ 32,000; realised 31,377 max-prompt / 31,867 total):
 1q cold 213 ms / warm 194 ms; **16q cold 2,123 ms / warm 2,033 ms** (second
 pass: 2,141/2,041). Cold = first request after backend ready on a restarted
 container (compile cache mounted; calibration probes excluded).
 **Clause (3): PASS** — warm 16q is 2.03 s against the 2.93 s bar, and it is
 the fastest honest 32k cell of any arm this bench measured.
 **(4) Concurrency:** 4 concurrent 16q × 32k on `--max-num-seqs 16`: all four
 answered correctly, wall 7.08 s, **peak 15,566 MiB** against the GPU 1 budget
 re-measured at run time (16,847 MiB; non-seat residents unchanged: 43,998 +
 34,994). Headroom 1,281 MiB. **Clause (4): PASS** — the clause brokkr flagged
 as most likely to bite did not bite; KV pressure at 4× concurrency grows
 sub-linearly because all four requests share the prefix.
 **(1a) Boot matrix** (reported, not gated): cold launch→ready median **153.6 s**
 (min 149.4 / max 185.7, n=3), warm (persistent compile cache) **153.5 s** (n=3),
 enforce-eager 161.9 s (n=1). Phases: weights 4.3 s, engine init 133.6 s (CUDA
 graph capture ~77 s, torch.compile ~13 s), shim label-verification 3.1–3.7 s.
 Two corrections to the main doc's boot row: the earlier ~280 s was measured
 without a persistent cache mount and was never reproducible (use ~154 s);
 warm cache saves nothing here because boot is graph-capture-bound, and
 eager mode — which skips capture entirely — is *slower* (161.9 s), so there
 is no faster-boot knob to grab. INC boots in 35 s; h2o's swap would cost
 ~2 min of boot per restart.
 **(5) Wyrd item by item** (168 scorable rows; 8 disagree, h2o gains 3, loses 5,
 net −2 — matching the −0.0167 point estimate; for Wyrd's owner):
 | item | h2o | INC | gold | verdict |
 |---|---|---|---|---|
 | fletcher/f-go-tavern exit2 | stay | tavern | tavern | h2o wrong |
 | fletcher/f-leave-town exit | meet_wren | stay | stay | h2o wrong |
 | fletcher/f-memory exit | meet_wren | stay | stay | h2o wrong |
 | fletcher/f-wren-comes exit2 | stay | meet_wren | meet_wren | h2o wrong |
 | toll/t-granted place | moved | holds | moved | h2o right |
 | toll/t-questions exit | departure | stay | stay | h2o wrong |
 | vigil/v-to-pub exit2 | stay | outside | stay | h2o right |
 | vigil/v-to-pub place | moved | holds | moved | h2o right |
 Pattern worth noting: h2o's losses are all fletcher-scene *exit* decisions
 (six of them in one scene family); its gains are the toll/vigil *place*
 updates. No relabels (zero rows correct-but-renamed).
 **Amendment 1 verdict: all three pass clauses hold.** (2) clean, (3) 2.03 s,
 (4) 15,566 ≤ 16,847. Per the amendment's fixed rule, h2o-lightning-4b now
 passes every Supersede clause — with the two honest caveats that its contract
 is a REFUSING 32,000-token seat (not INC's silent-any-size behaviour), and
 boot is 4.4x INC's. The seat swap is a decision for the operator and
 infra-ops; the bench does not swap anything.
 Teardown after the confirmation: wrapper, shim and container stopped; GPU 3
 guard re-read 2 MiB before release; weights still kept (model-runtime 116 GB
 + 27 GB materialized serving dirs).
@@ -1,6 +1,6 @@
 # Persistent memory — eshpfi-management
-_Last updated: 2026-10-11 ~1400 PT (U13 cutover executed + acceptance; fv-ml1 hard hang and evidence capture staged; esh-nas-pve stale snapshots destroyed; vh albok-everything; eshpfi pushed.)_
+_Last updated: 2026-10-12 ~1710 PT (decision seat: h2o 1311–1403, rolled back, judge re-fitted, h2o LIVE again 1704; earlier: U13 cutover + acceptance, fv-ml1 evidence capture, vh albok-everything.)_
 > **Always check for `/tmp/infra-ops-handoff.md`** — if it exists and its
 > `Written:` stamp is under **8 hours** old, read it (it carries the in-flight
@@ -140,7 +140,7 @@ _As of 2026-10-09 ~1955 PT. The newest subsection is first; older subsections ca
 | 0 | parakeet-nemo | audio | rest 2,108 MiB, cap 3,840 | — | — |
 | 1 | gen-small (+summarizer/classifier) | 262,144 | 5.5 GiB | 461,215 | 1.76× (default output cap 32,768) |
 | 1 | meromero-rp (char-rp) | 262,144 | 22 GiB | 346,436 | 1.32× |
-| 1 | intern-decision | 32,768/call | cap 14.4 GiB, peak 15,220 | — | — |
+| 1 | intern-decision = **h2o-lightning-4b** since 10-12 1704 (INC = rollback target) | 40,960 (front admits 32,000/call) | 4.48 GiB (4,810,363,371 B) | 135,558 | 3.31× (seat peak 16,164 under 4×16q×32k, budget 16,847) |
 | 2 | flash-next (gen, gen-large, +6) | 262,144 | — | 344,155 | 1.31× |
 Every vLLM seat on a shared card (GPU 0, GPU 1) has byte-pinned KV (`--kv-cache-memory`), so util only gates the boot check (Free at start ≥ util × total). flash-next is alone on GPU 2 and stays util-sized (0.96). ⚠ Before restarting any seat on a shared card, gate on the arithmetic (`scripts/oneoff/fv-ml1-gpucheck.sh`), never trial-boot. Before restarting a user-facing service, the idle check must STOP the restart, not just print (auto-memory `feedback_idle_check_must_gate_the_restart`).
@@ -166,10 +166,12 @@ Every vLLM seat on a shared card (GPU 0, GPU 1) has byte-pinned KV (`--kv-cache-
 - **Legacy archive:** DESTROY it whole by **2026-10-30**, or at retirement-done, or on a subject-erasure request, whichever comes first. The runbook is in the detail file.
 - **TODO:** re-sweep both api logs after real traffic. After the b193 push, remove the retired config keys (list: detail file, "Post-b193 config retirements").
-### intern-decision (replaced SemIf on 2026-09-30)
+### The decision seat `intern-decision` :8033 (h2o-lightning-4b since 2026-10-12 1704; INC = rollback)
- **LIVE 0.1.3** at `intern-decision.fv.internal:8033`: semif-compatible `/decide` plus Jev `/v1/systemone`, 32k tokens, a Triton cache volume. Run `scripts/intern-decision-warmup` after an IMAGE change. → `persistent-memory.d/2026-09-30-semif-replaced-by-intern-decision.md`
+- **LIVE: h2o-lightning-4b @ b43f759b** (`compose.h2o.yaml`, `switch-to-h2o.sh`) behind `services/h2o-decision-front`: 32k refuse-not-trim admission, bearer, `/decide/shared` translator. Inline images only (`LIGHTNING_IMAGE_URLS=off` + vLLM `--allowed-media-domains media.invalid`, both verified). Nemi's judge uses the **v2 note question** (albok-nemi 0.1.18) at **concurrency 4**.
- **Open:** label ~50 real Wyrd/Cicada turns before trusting it in production; its card makes no contamination claim.
+- **Rollback:** `sudo bash /opt/docker/compose/intern-decision/rollback-to-inc.sh` (INC `intern-decision-serve:0.1.3`, ~35 s), **and albok-dev sets judge concurrency to 1**. INC scores on one thread (INV-2), and 4 in flight queued into 339 timeouts on 10-12.
 - ⚠ Seat headroom under 4-way 32k load is 683 MiB (peak 16,164 against 16,847). Re-measure before anything else lands on GPU 1.
 - INC history: → `persistent-memory.d/2026-09-30-semif-replaced-by-intern-decision.md`.
 ### nh3-pve + nh3-ml1 (2026-09-25/26)
@@ -178,6 +180,9 @@ Every vLLM seat on a shared card (GPU 0, GPU 1) has byte-pinned KV (`--kv-cache-
 ## Recent decisions
 - `[2026-10-12]` **Decision seat: h2o-lightning-4b served 1311–1403, then ROLLED BACK to INC** on albok-dev's word, with brokkr concurring (Prime's plan made the swap conditional on the judge's sign-off). Every eval-set check had passed (JevBench 205/231, 422 refusal, 16q@32k ~2.1 s, seat peak 16,164/16,847), but albok's full walk over 10,841 same texts (both verdicts, same prompt and policy) found h2o flags 242 adult at 0.5 against INC's 88, with 154 extra at 0.80–0.90 that no threshold separates; for personal, 5 vs 11 with only 2 in common. **Lesson: a passing benchmark plus a small borderline set (9 entries, 2 → 5) could not resolve a 2× shift in a production classifier; only the same-text fleet comparison did.** albok's re-stamp walk on INC (1406–1408) put the hidden set back exactly to its pre-cutover counts (fleet/memory 86, agent-feedback 21, nonfiction 0). Prime's inline-images-only order (via brokkr) is carried in the staged h2o compose. h2o returns only after a judge re-fit. Threads `01M4RSFC9J2R0J07EEBD4N04W3` (brokkr), `01M4RYSMP3DV9HZHJ0KNBH2S96` (albok). **Then:** the round 2 eval seat on GPU 3 (1428–1619, Prime 1420) found the cause. The judge's v1 note question asked two things at once: INC answered it inconsistently, h2o literally. The shim floor is ruled out (identical verdicts on and off). Prime ruled "text itself", and albok-nemi 0.1.18 (judge v2) re-fitted the question. Both seats agree under v2 (work notes INC 0/22, h2o 1/22 at 0.52). **Prime ruled the cutover at 1640 ("do it", via albok-dev): h2o LIVE again since 1704** (`switch-to-h2o.sh` 170223, healthy 170454, probes clean). albok-dev's v2 fleet walk on h2o followed, with expected counts 0/0/6/2. Threads `01M4S026GWS354WZGXZ5E34F31`, `01M4S91WPPPN0B17JBNX98K5CY`.
 - `[2026-10-12]` **U13 regression fixed 0943 (svos-dev's URGENT, Prime saw every app pane 403):** the browser frames a pane at the registry `base_url`, and U13 had set app-terminal's to the seat-host origin, which admits only nh3-svos. Fix: `base_url` back to `https://svos-terminal.phasefinal.com` (config_gen `acae45206bbc9fcb`), plus POST `/seats/{observe,check,launch,retire-launch}` allowed from the box on that origin. svos-dev's one-line fix alone would have 404'd the kernel's seat-host calls; the origin never carried them before. svos-dev owes a separate browser-facing origin field; then the seat-host goes back to being kernel-facing.
 - `[2026-10-12]` **Shared CI runner (ana-docker act_runner, capacity 2, `timeout: 30m`) is slow because pfi-pve's `ospool` is HDD mirrors with no SLOG:** 8 KiB `oflag=dsync` writes measured ~145 ms on ana-docker vs ~2.7 ms on nh3-docker (n=3 each; pfi-pve host on ospool 18–32 ms; corviduo-dev ~75 ms). Jobs past 30 min are killed with a truncated log and no footer (albok every run since 10-04; vastblue 10935). Worse since the 10-04 PVE 9 reboot, cause not pinned (drive write cache was the same before and after: sda WCE on, sdb–sdd off). vastblue fixed its side with PostgreSQL on tmpfs (run 11033 green, pytest 14m17s vs 7–8 min per 72-test batch); albok is adding pytest-timeout + durations in 0.1.19. **Prime's call, pending:** A opt-in fast runner on nh3-docker (recommended), B SLOG for ospool on the BOSS (check endurance), C leave it. ana-docker vCPU is x86-64-v2-AES (no AVX2), a separate lean-against.
 - `[2026-10-11]` **worldtree-dev is reading mail again (2219), answering four of our threads at once**: #416 residual REOPENED for a second writer pass (a filer-side deterministic expiry guard, U4 rev 1.17, after #422/#423; acceptance = 0 leaks over 10 daily batches with the filing/user median in band). Keep the daily report's "known #416 residual" classification until then. The writer/gate never call embed/rerank and the gate fails CLOSED, so the 10-04 leak was the writer's judgment. No confirmatory batch for 10-09 (10-10/10-11 = 0.83). Nothing of theirs sets a small max_tokens on gen-frontier. This makes Prime's "who sees gate failures" question moot. Thread `01M4BDETDGMNVDTFXN3T3Y63DP`.
 - `[2026-10-11]` **U13 cutover executed on Prime's go (0721) + his four rulings** — E1 0733, E2 steps 1–6 0734–0816, URGENT acceptance test 1350; fv-ml1 evidence capture, stale snapshots, Mac keys, no key rotation. → `persistent-memory.d/2026-10-11-u13-cutover-executed.md`
 - `[2026-10-11]` **fv-ml1 hard hang ~0005, self-reset, no panic record; pre-patch snapshot HELD** — and infra-ops' soft-off that landed mid-boot. → `persistent-memory.d/2026-10-11-fv-ml1-hard-hang.md`
@@ -0,0 +1,21 @@
 # h2o decision front
 The front of the h2o-lightning-4b decision seat on fv-ml1 (seat swap ordered by Prime 2026-10-12, through brokkr).
 `front.py` sits in front of the vendor's **unmodified** `h2o_lightning_shim.py`, which sits in front of stock vLLM
 0.30.0. It adds what the seat contract needs and the vendor stack lacks:
 - **Admission** (infra-hermes's `tools/h2o-admission/admission_proxy.py` @ 30f36f2, logic unchanged): count every
  decision request with h2o's own tokenizer over the shim's own renders, and refuse over 32,000 tokens with 422 + the
  count. The shim would otherwise trim silently.
 - **The seat's bearer token** on everything but `GET /health`, as intern-decision-serve has.
 - **`/decide` and `/decide/shared` continuity** for albok-nemi's sensitivity judge: intern-decision-serve's own
  mapping (one Jev `choice` question per decision, calls of at most 16, fields `q`/`q1..qN`), re-keyed into the semif
  result shape (`results[i].native.probabilities[option id]`). `orderings` other than `none` → 422.
 - **`/health`** with `model: {name, source, revision}`, which the judge stamps into its cache key.
 `run.sh` runs the shim (loopback) and the front in one container (`python:3.12-slim`, host network). Stacks copy both
 files into `conf/front/`: `stacks/h2o-eval` (TEMPORARY, GPU 3, :8034) today.
 Verified 2026-10-12 1300 on the eval seat: 401 without the token; `/decide/shared` in the judge's shape; a planted
 personal record → `personal` p_yes .911, a neutral note → .056 (n=1 each, smoke test only); 40,088 tokens → 422 with
 the count; `/v1/systemone` passthrough with `_admission` counts; reachable from nh3-dev.
@@ -0,0 +1,231 @@
 #!/usr/bin/env python3
 """The front of the h2o-lightning-4b decision seat (fv-ml1). Stdlib only.
 It sits in front of the UNMODIFIED shipped h2o_lightning_shim.py (which sits in front of stock vLLM 0.30.0) and adds
 the three things the seat contract needs and the vendor stack does not have:
 1. ADMISSION (from infra-hermes's tools/h2o-admission/admission_proxy.py @ 30f36f2, logic unchanged): the shim
   silently head-and-tail-trims state above ~32,000 tokens, so every decision request is counted first with h2o's own
   tokenizer (vLLM /tokenize over the shim's own per-question renders) and refused with 422 + the count when either the
   largest prompt or the shim-accounted total (shared head + tails) exceeds FRONT_CAP. Refused, never trimmed.
 2. THE SEAT'S AUTH: a bearer token on everything but GET /health, as intern-decision-serve has (OPEN_PATHS /health).
 3. CONTINUITY for /decide and /decide/shared (semif's surface, albok-nemi's sensitivity judge): each decision becomes
   one Jev `choice` question exactly as intern-decision-serve builds it ({type: choice, instructions: question,
   criteria: {option id: description}}), asked through /v1/systemone in calls of at most 16 questions, and re-keyed
   into the semif result shape the judge reads (results[i].native.probabilities[option id]). orderings other than
   "none" are refused (422): the judge does its own forward/reverse. /health carries {model: {name, revision}}, which
   the judge stamps into its cache key.
 Env: FRONT_TOKEN (required), FRONT_BIND, FRONT_PORT, FRONT_SHIM (shim URL), FRONT_VLLM (vLLM URL), FRONT_CAP (32000),
 FRONT_SHIM_FILE (vendor shim path, imported for its renderers), FRONT_SERVE_CONFIG, FRONT_MODEL_NAME,
 FRONT_MODEL_SOURCE, FRONT_MODEL_REVISION, FRONT_MAX_DECISIONS (64).
 """
 import hmac, importlib.util, json, os, sys, time, urllib.error, urllib.request
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 TOKEN = os.environ["FRONT_TOKEN"]
 BIND = os.environ.get("FRONT_BIND", "127.0.0.1")
 PORT = int(os.environ.get("FRONT_PORT", "8034"))
 UP = os.environ.get("FRONT_SHIM", "http://127.0.0.1:18191")
 VLLM = os.environ.get("FRONT_VLLM", "http://127.0.0.1:18199")
 CAP = int(os.environ.get("FRONT_CAP", "32000"))
 SHIM_FILE = os.environ.get("FRONT_SHIM_FILE", "/model/h2o_lightning_shim.py")
 SERVE_CONFIG = os.environ.get("FRONT_SERVE_CONFIG", "/model/serve_config.json")
 MODEL = {"name": os.environ.get("FRONT_MODEL_NAME", "h2o-lightning-4b"),
         "source": os.environ.get("FRONT_MODEL_SOURCE", "h2oai/h2o-lightning-4b"),
         "revision": os.environ.get("FRONT_MODEL_REVISION", "b43f759b6385d21182f87c3829931ec3d4dfa98a")}
 MAX_DECISIONS = int(os.environ.get("FRONT_MAX_DECISIONS", "64"))
 MAX_QUESTIONS_PER_CALL = 16
 MIN_OPTIONS, MAX_OPTIONS = 2, 16
 spec = importlib.util.spec_from_file_location("h2o_shim", SHIM_FILE)
 shim_mod = importlib.util.module_from_spec(spec)
 sys.modules["h2o_shim"] = shim_mod
 spec.loader.exec_module(shim_mod)
 CFG = shim_mod.Contract(json.load(open(SERVE_CONFIG)))
 CFG.max_state_tokens = 10**9   # count the UNTRUNCATED render; the shim's own trim is what admission polices
 class Refused(Exception):
    def __init__(self, code, obj):
        super().__init__(obj)
        self.code, self.obj = code, obj
 def sent_tokens(body):
    """(max tokens of one question's prompt, shim-accounted total = shared head + tails), counted by vLLM's own
    tokenizer over the shim's own per-question renders (admission_proxy.py @ 30f36f2, unchanged)."""
    state = body.get("state") or ""
    if isinstance(state, (dict, list)):
        state = json.dumps(state)
    state_txt = shim_mod.render_state(state, CFG.state_compact)
    all_toks = []
    for qid, q in body.get("questions", {}).items():
        prompt_text = CFG.question_prompt(state_txt, q, qid)[0]
        req = urllib.request.Request(VLLM + "/tokenize",
            data=json.dumps({"model": CFG.model, "prompt": prompt_text, "add_special_tokens": False}).encode(),
            headers={"Content-Type": "application/json"})
        with urllib.request.urlopen(req, timeout=30) as r:
            all_toks.append(json.loads(r.read())["tokens"])
    mx = max(len(t) for t in all_toks)
    shared = shim_mod.common_prefix_len(all_toks)
    return mx, shared + sum(len(t) - shared for t in all_toks)
 def admit(body):
    try:
        mx, total = sent_tokens(body)
    except Refused:
        raise
    except Exception as e:  # noqa: BLE001
        raise Refused(502, {"error": f"admission: tokenizer count failed: {type(e).__name__}: {e}"})
    if mx > CAP or total > CAP:
        raise Refused(422, {"error": f"admission: {mx} max-prompt / {total} total tokens against the {CAP}-token seat "
                                     f"limit; refused, not trimmed", "max_prompt_tokens": mx,
                            "shim_accounted_total": total, "cap": CAP})
    return mx, total
 def post_up(path, obj):
    req = urllib.request.Request(UP + path, data=json.dumps(obj).encode(), headers={"Content-Type": "application/json"},
                                 method="POST")
    try:
        with urllib.request.urlopen(req, timeout=300) as r:
            return r.status, json.loads(r.read())
    except urllib.error.HTTPError as e:
        raw = e.read() or b"{}"
        try:
            return e.code, json.loads(raw)
        except ValueError:
            return e.code, {"error": raw[:300].decode(errors="replace")}
 def systemone(body):
    mx, total = admit(body)
    code, obj = post_up("/v1/systemone", body)
    if isinstance(obj, dict):
        obj.setdefault("_admission", {}).update({"max_prompt_tokens": mx, "shim_accounted_total": total, "cap": CAP})
    return code, obj
 def check_decision(d, i):
    if not isinstance(d, dict) or not isinstance(d.get("id"), str) or not isinstance(d.get("question"), str):
        raise Refused(422, {"error": f"decision {i}: needs string id and question"})
    opts = d.get("options")
    if not isinstance(opts, list) or not MIN_OPTIONS <= len(opts) <= MAX_OPTIONS:
        raise Refused(422, {"error": f"decision {d['id']!r}: options must contain {MIN_OPTIONS}-{MAX_OPTIONS} entries"})
    if any(not isinstance(o, dict) or not isinstance(o.get("id"), str) or not isinstance(o.get("description"), str)
           for o in opts):
        raise Refused(422, {"error": f"decision {d['id']!r}: every option needs string id and description"})
    if len({o["id"] for o in opts}) != len(opts):
        raise Refused(422, {"error": f"decision {d['id']!r}: option ids must be unique"})
    if d.get("orderings", "none") != "none":
        raise Refused(422, {"error": f"decision {d['id']!r}: orderings {d.get('orderings')!r} is not served by this "
                                     "seat (h2o); send orderings 'none' and combine on the client"})
 def decide(state, decisions):
    """semif /decide/shared on top of /v1/systemone: intern-decision-serve's mapping (one Jev choice question per
    decision, calls of at most 16, fields q / q1..qN), re-keyed into its result shape."""
    if not isinstance(decisions, list) or not 1 <= len(decisions) <= MAX_DECISIONS:
        raise Refused(422, {"error": f"decisions must be a list of 1..{MAX_DECISIONS}"})
    for i, d in enumerate(decisions):
        check_decision(d, i)
    started, results, sizes, tokens = time.perf_counter(), [], [], []
    for k in range(0, len(decisions), MAX_QUESTIONS_PER_CALL):
        chunk = decisions[k:k + MAX_QUESTIONS_PER_CALL]
        fields = ["q"] if len(chunk) == 1 else [f"q{j + 1}" for j in range(len(chunk))]
        questions = {f: {"type": "choice", "instructions": d["question"],
                         "criteria": {o["id"]: o["description"] for o in d["options"]}} for f, d in zip(fields, chunk)}
        code, resp = systemone({"state": state, "questions": questions})
        if code != 200:
            raise Refused(code, resp if isinstance(resp, dict) else {"error": str(resp)})
        sizes.append(len(chunk))
        tokens.append((resp.get("usage") or {}).get("input_tokens"))
        for f, d in zip(fields, chunk):
            answer = (resp.get("answers") or {}).get(f)
            ids = [o["id"] for o in d["options"]]
            probs = (answer or {}).get("probabilities") or {}
            if answer is None or any(i not in probs for i in ids):
                raise Refused(500, {"error": f"scoring_failed: the model's answer for field {f!r} lacks option ids"})
            results.append({"id": d["id"], "option_ids": ids, "probabilities": [probs[i] for i in ids],
                            "top": answer.get("choice"), "confidence": answer.get("confidence"),
                            "calibration": resp.get("calibration"), "native": answer,
                            "input_tokens": tokens[-1], "model": MODEL,
                            "call": {"index": len(sizes) - 1, "field": f, "questions": len(chunk)}})
    return {"results": results, "timing": {"total_seconds": time.perf_counter() - started, "calls": len(sizes),
                                           "questions_per_call": sizes, "input_tokens": tokens}}
 class H(BaseHTTPRequestHandler):
    protocol_version = "HTTP/1.1"
    def log_message(self, fmt, *args):
        sys.stderr.write("%s - %s\n" % (self.address_string(), fmt % args))
    def _send(self, code, obj):
        b = json.dumps(obj).encode()
        self.send_response(code)
        self.send_header("Content-Type", "application/json")
        self.send_header("Content-Length", str(len(b)))
        self.end_headers()
        self.wfile.write(b)
    def _authorized(self):
        scheme, _, tok = (self.headers.get("Authorization") or "").partition(" ")
        return scheme.lower() == "bearer" and hmac.compare_digest(tok.strip().encode(), TOKEN.encode())
    def do_GET(self):
        path = self.path.split("?")[0]
        if path == "/health":
            try:
                with urllib.request.urlopen(UP + "/health", timeout=10) as r:
                    shim = json.loads(r.read())
            except Exception as e:  # noqa: BLE001
                return self._send(503, {"status": "unavailable", "model": MODEL, "error": f"shim: {type(e).__name__}"})
            ok = bool(shim.get("ok"))
            return self._send(200 if ok else 503, {
                "status": "ok" if ok else "starting", "model": MODEL, "cap_tokens": CAP, "admission": "refuse, never trim",
                "endpoints": ["/decide", "/decide/shared", "/v1/systemone"], "max_decisions": MAX_DECISIONS,
                "max_questions_per_call": MAX_QUESTIONS_PER_CALL, "orderings": ["none"], "backend": shim})
        if not self._authorized():
            return self._send(401, {"error": "unauthorized"})
        if path == "/v1/models":
            try:
                with urllib.request.urlopen(UP + "/v1/models", timeout=10) as r:
                    return self._send(r.status, json.loads(r.read()))
            except urllib.error.HTTPError as e:
                return self._send(e.code, {"error": "upstream"})
        return self._send(404, {"error": "not found"})
    def do_POST(self):
        path = self.path.split("?")[0]
        if not self._authorized():
            return self._send(401, {"error": "unauthorized"})
        n = int(self.headers.get("Content-Length", 0) or 0)
        if n > 4 * 1024 * 1024:
            return self._send(413, {"error": "body over 4 MiB"})
        try:
            body = json.loads(self.rfile.read(n) or b"null")
        except ValueError:
            return self._send(422, {"error": "body is not JSON"})
        if not isinstance(body, dict):
            return self._send(422, {"error": "body must be a JSON object"})
        try:
            if path == "/v1/systemone":
                return self._send(*systemone(body))
            if path == "/api/alpha/decisions":
                admit(body)
                return self._send(*post_up(path, body))
            if path == "/decide/shared":
                return self._send(200, decide(body.get("state"), body.get("decisions")))
            if path == "/decide":
                d = {k: body.get(k) for k in ("id", "question", "options", "orderings") if k in body}
                out = decide(body.get("state"), [d])
                return self._send(200, {**out["results"][0], "total_seconds": out["timing"]["total_seconds"]})
        except Refused as r:
            return self._send(r.code, r.obj)
        return self._send(404, {"error": "not found"})
 ThreadingHTTPServer((BIND, PORT), H).serve_forever()
@@ -0,0 +1,13 @@
 #!/bin/bash
 # One container: the unmodified vendor shim on loopback, then the front. If either dies, the container exits and
 # Docker restarts it (restart: unless-stopped). Both inherit the container's whole environment, so the shim's own
 # settings (SHIM_*, e.g. SHIM_NOUL_FLOOR, and LIGHTNING_*, e.g. LIGHTNING_IMAGE_URLS=off) set in compose reach it.
 set -eu
 python3 /model/h2o_lightning_shim.py --config /model/serve_config.json --vllm "$SHIM_VLLM" \
  --host 127.0.0.1 --port "$SHIM_PORT" &
 SHIM=$!
 python3 /front/front.py &
 FRONT=$!
 wait -n "$SHIM" "$FRONT" 2>/dev/null || true
 kill "$SHIM" "$FRONT" 2>/dev/null || true
 exit 1
@@ -0,0 +1,35 @@
 # h2o-eval (TEMPORARY) — fv-ml1 GPU 3, :8034
 > **DOWN since 2026-10-12 1619 PT** (round 2 done: the floor was ruled out, with identical verdicts on and off, and albok-nemi 0.1.18
 > re-fitted the judge's note question). The host keeps `/opt/docker/compose/h2o-eval` and the weights.
 >
 > **Round 2 (was UP 2026-10-12 1428–1619 PT)** (Prime 1420, via albok-dev): "I'd like to understand why h2o failed, so
 > stand it up on gpu3 and refit it and let's get an understanding." albok-dev runs the investigation and the judge
 > re-fit (thread `01M4S026GWS354WZGXZ5E34F31`). **No load tests or other traffic while their runs are in.** Round 1
 > (1300–1319) was the pre-cutover validation. Take it down with `docker compose down` when albok-dev is done.
 >
 > | port | front | floor | judge stamp (`/health` model) |
 > |---|---|---|---|
 > | `:8034` | `h2o-eval-front` | ON, 0.801, as the live cutover ran | `h2o-lightning-4b@b43f759b…`, identical to the cutover's |
 > | `:8035` | `h2o-eval-front-nofloor` | OFF (`SHIM_NOUL_FLOOR=0`) | `h2o-lightning-4b+noul-floor-0@b43f759b…` |
 >
 > Both fronts sit over ONE vLLM (14,420 MiB, KV 135,070 tokens, util-sized because GPU 3 is otherwise empty). Each
 > `/health` also reports `backend.noul_floor {value, source}`.
 >
 > Checked at 1428:
 > - `/decide/shared` 200 on both, with the cutover's smoke values reproduced (adult 0.052, personal 0.876).
 > - An https image URL is refused on both: 422 "not accepted here".
 > - **The vLLM backstop answered for the first time.** An https image URL posted straight to vLLM got 400 "The URL must
 >   be from one of the allowed domains: ['media.invalid']", with no fetch.
 h2o-lightning-4b @ b43f759b behind `services/h2o-decision-front` (copied into `conf/front/`), for albok-dev's
 re-validation of Nemi's sensitivity judge before the GPU 1 cutover (Prime 2026-10-12 via brokkr). **Remove right after
 the cutover:** GPU 3 is for transient work only (Scriberr and MIA use it).
 - Weights: `/tank/aimodels/h2o-lightning-4b-b43f759b` (copied from the bench kit; model.safetensors sha256
  `2d51561f221e1549d91bd1c38b3512a1435dd32950d454a9f040f1d4d11497b6`, tokenizer.json `5f9e4d49…`, shim git oid
  `fee7807e…`, serve_config `5e6fef1c…`: all equal to Hugging Face at b43f759b).
 - vLLM exactly as benched: `--max-model-len 40960 --max-num-seqs 16 --gpu-memory-utilization 0.1506`, vision limits as
  the vendor's serve.sh; healthy ~3 min after `up` (graph capture), ~14.4 GB on GPU 3.
 - `.env` on the host: `FRONT_TOKEN` = intern-decision's token (= vault `albok/judge-token`).
 - Remove: `cd /opt/docker/compose/h2o-eval && docker compose down` (the weights stay for the GPU 1 seat).
@@ -0,0 +1,90 @@
 # TRANSIENT h2o-lightning-4b eval seat on fv-ml1 GPU 3 (GPU 3 is for transient work only).
 # Round 1, 2026-10-12 1300-1319: albok-dev validated Nemi's judge on h2o before the GPU 1 cutover (Prime via brokkr).
 # Round 2, from 2026-10-12 ~1430 (Prime 1420, via albok-dev): "I'd like to understand why h2o failed, so stand it up
 #   on gpu3 and refit it and let's get an understanding." albok-dev investigates the 154 texts h2o alone called adult
 #   and re-fits the judge. Two fronts over ONE vLLM:
 #     :8034 front          the shim's yes/no floor ON (0.801), exactly as the live cutover ran; /health model stamp
 #                          h2o-lightning-4b@b43f759b..., identical to the cutover's
 #     :8035 front-nofloor  SHIM_NOUL_FLOOR=0 (floor OFF, brokkr's lever); stamp h2o-lightning-4b+noul-floor-0@b43f759b...
 #   Each /health also reports backend.noul_floor {value, source}. NO load tests or other traffic while albok-dev's
 #   runs are in (they name start/end on thread 01M4S026GWS354WZGXZ5E34F31). INC keeps :8033 on GPU 1, untouched.
 # Serving exactly as benched (docs/pfi/decision-sweep-followon-bench-2026-10-12.md, Amendment 1): stock vLLM 0.30.0
 # + the vendor's unmodified shim + the admission front (services/h2o-decision-front, copied into conf/front).
 # Weights: /tank/aimodels/h2o-lightning-4b-b43f759b (hashes match HF at b43f759b; see the stack README).
 # Inline images only (Prime 2026-10-12): LIGHTNING_IMAGE_URLS=off on both fronts + vLLM --allowed-media-domains.
 name: h2o-eval
 services:
  vllm:
    image: vllm/vllm-openai:v0.30.0
    container_name: h2o-eval-vllm
    restart: unless-stopped
    ports:
      - "127.0.0.1:18199:8000"
    environment:
      VLLM_USE_FLASHINFER_SAMPLER: "0"
    volumes:
      - /tank/aimodels/h2o-lightning-4b-b43f759b:/model:ro
      - h2o-eval-vllm-cache:/root/.cache/vllm
    command: ["--model", "/model", "--served-model-name", "h2oai/h2o-lightning-4b",
              "--max-model-len", "40960", "--max-num-seqs", "16", "--gpu-memory-utilization", "0.1506",
              "--limit-mm-per-prompt", "{\"image\": 4, \"video\": 0}",
              "--mm-processor-kwargs", "{\"max_pixels\": 1605632}",
              "--allowed-media-domains", "media.invalid"]
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["3"]
              capabilities: [gpu]
  front:
    image: python:3.12-slim
    container_name: h2o-eval-front
    restart: unless-stopped
    network_mode: host
    depends_on: [vllm]
    env_file: .env
    environment:
      SHIM_VLLM: http://127.0.0.1:18199
      SHIM_PORT: "18191"
      SHIM_NOUL_FLOOR: "0.801"
      LIGHTNING_IMAGE_URLS: "off"   # Prime 2026-10-12: inline images only, the seat never fetches an image URL
      FRONT_SHIM: http://127.0.0.1:18191
      FRONT_VLLM: http://127.0.0.1:18199
      FRONT_BIND: 0.0.0.0
      FRONT_PORT: "8034"
      FRONT_CAP: "32000"
      FRONT_SHIM_FILE: /model/h2o_lightning_shim.py
      FRONT_SERVE_CONFIG: /model/serve_config.json
      FRONT_MODEL_REVISION: b43f759b6385d21182f87c3829931ec3d4dfa98a
    volumes:
      - /tank/aimodels/h2o-lightning-4b-b43f759b:/model:ro
      - /opt/docker/conf/h2o-eval/front:/front:ro
    command: ["bash", "/front/run.sh"]
  front-nofloor:
    image: python:3.12-slim
    container_name: h2o-eval-front-nofloor
    restart: unless-stopped
    network_mode: host
    depends_on: [vllm]
    env_file: .env
    environment:
      SHIM_VLLM: http://127.0.0.1:18199
      SHIM_PORT: "18192"
      SHIM_NOUL_FLOOR: "0"           # the floor OFF (the shim: 0 turns it off)
      LIGHTNING_IMAGE_URLS: "off"   # Prime 2026-10-12: inline images only, the seat never fetches an image URL
      FRONT_SHIM: http://127.0.0.1:18192
      FRONT_VLLM: http://127.0.0.1:18199
      FRONT_BIND: 0.0.0.0
      FRONT_PORT: "8035"
      FRONT_CAP: "32000"
      FRONT_SHIM_FILE: /model/h2o_lightning_shim.py
      FRONT_SERVE_CONFIG: /model/serve_config.json
      FRONT_MODEL_NAME: h2o-lightning-4b+noul-floor-0   # a distinct judge stamp
      FRONT_MODEL_REVISION: b43f759b6385d21182f87c3829931ec3d4dfa98a
    volumes:
      - /tank/aimodels/h2o-lightning-4b-b43f759b:/model:ro
      - /opt/docker/conf/h2o-eval/front:/front:ro
    command: ["bash", "/front/run.sh"]
 volumes:
  h2o-eval-vllm-cache:
@@ -0,0 +1,231 @@
 #!/usr/bin/env python3
 """The front of the h2o-lightning-4b decision seat (fv-ml1). Stdlib only.
 It sits in front of the UNMODIFIED shipped h2o_lightning_shim.py (which sits in front of stock vLLM 0.30.0) and adds
 the three things the seat contract needs and the vendor stack does not have:
 1. ADMISSION (from infra-hermes's tools/h2o-admission/admission_proxy.py @ 30f36f2, logic unchanged): the shim
   silently head-and-tail-trims state above ~32,000 tokens, so every decision request is counted first with h2o's own
   tokenizer (vLLM /tokenize over the shim's own per-question renders) and refused with 422 + the count when either the
   largest prompt or the shim-accounted total (shared head + tails) exceeds FRONT_CAP. Refused, never trimmed.
 2. THE SEAT'S AUTH: a bearer token on everything but GET /health, as intern-decision-serve has (OPEN_PATHS /health).
 3. CONTINUITY for /decide and /decide/shared (semif's surface, albok-nemi's sensitivity judge): each decision becomes
   one Jev `choice` question exactly as intern-decision-serve builds it ({type: choice, instructions: question,
   criteria: {option id: description}}), asked through /v1/systemone in calls of at most 16 questions, and re-keyed
   into the semif result shape the judge reads (results[i].native.probabilities[option id]). orderings other than
   "none" are refused (422): the judge does its own forward/reverse. /health carries {model: {name, revision}}, which
   the judge stamps into its cache key.
 Env: FRONT_TOKEN (required), FRONT_BIND, FRONT_PORT, FRONT_SHIM (shim URL), FRONT_VLLM (vLLM URL), FRONT_CAP (32000),
 FRONT_SHIM_FILE (vendor shim path, imported for its renderers), FRONT_SERVE_CONFIG, FRONT_MODEL_NAME,
 FRONT_MODEL_SOURCE, FRONT_MODEL_REVISION, FRONT_MAX_DECISIONS (64).
 """
 import hmac, importlib.util, json, os, sys, time, urllib.error, urllib.request
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 TOKEN = os.environ["FRONT_TOKEN"]
 BIND = os.environ.get("FRONT_BIND", "127.0.0.1")
 PORT = int(os.environ.get("FRONT_PORT", "8034"))
 UP = os.environ.get("FRONT_SHIM", "http://127.0.0.1:18191")
 VLLM = os.environ.get("FRONT_VLLM", "http://127.0.0.1:18199")
 CAP = int(os.environ.get("FRONT_CAP", "32000"))
 SHIM_FILE = os.environ.get("FRONT_SHIM_FILE", "/model/h2o_lightning_shim.py")
 SERVE_CONFIG = os.environ.get("FRONT_SERVE_CONFIG", "/model/serve_config.json")
 MODEL = {"name": os.environ.get("FRONT_MODEL_NAME", "h2o-lightning-4b"),
         "source": os.environ.get("FRONT_MODEL_SOURCE", "h2oai/h2o-lightning-4b"),
         "revision": os.environ.get("FRONT_MODEL_REVISION", "b43f759b6385d21182f87c3829931ec3d4dfa98a")}
 MAX_DECISIONS = int(os.environ.get("FRONT_MAX_DECISIONS", "64"))
 MAX_QUESTIONS_PER_CALL = 16
 MIN_OPTIONS, MAX_OPTIONS = 2, 16
 spec = importlib.util.spec_from_file_location("h2o_shim", SHIM_FILE)
 shim_mod = importlib.util.module_from_spec(spec)
 sys.modules["h2o_shim"] = shim_mod
 spec.loader.exec_module(shim_mod)
 CFG = shim_mod.Contract(json.load(open(SERVE_CONFIG)))
 CFG.max_state_tokens = 10**9   # count the UNTRUNCATED render; the shim's own trim is what admission polices
 class Refused(Exception):
    def __init__(self, code, obj):
        super().__init__(obj)
        self.code, self.obj = code, obj
 def sent_tokens(body):
    """(max tokens of one question's prompt, shim-accounted total = shared head + tails), counted by vLLM's own
    tokenizer over the shim's own per-question renders (admission_proxy.py @ 30f36f2, unchanged)."""
    state = body.get("state") or ""
    if isinstance(state, (dict, list)):
        state = json.dumps(state)
    state_txt = shim_mod.render_state(state, CFG.state_compact)
    all_toks = []
    for qid, q in body.get("questions", {}).items():
        prompt_text = CFG.question_prompt(state_txt, q, qid)[0]
        req = urllib.request.Request(VLLM + "/tokenize",
            data=json.dumps({"model": CFG.model, "prompt": prompt_text, "add_special_tokens": False}).encode(),
            headers={"Content-Type": "application/json"})
        with urllib.request.urlopen(req, timeout=30) as r:
            all_toks.append(json.loads(r.read())["tokens"])
    mx = max(len(t) for t in all_toks)
    shared = shim_mod.common_prefix_len(all_toks)
    return mx, shared + sum(len(t) - shared for t in all_toks)
 def admit(body):
    try:
        mx, total = sent_tokens(body)
    except Refused:
        raise
    except Exception as e:  # noqa: BLE001
        raise Refused(502, {"error": f"admission: tokenizer count failed: {type(e).__name__}: {e}"})
    if mx > CAP or total > CAP:
        raise Refused(422, {"error": f"admission: {mx} max-prompt / {total} total tokens against the {CAP}-token seat "
                                     f"limit; refused, not trimmed", "max_prompt_tokens": mx,
                            "shim_accounted_total": total, "cap": CAP})
    return mx, total
 def post_up(path, obj):
    req = urllib.request.Request(UP + path, data=json.dumps(obj).encode(), headers={"Content-Type": "application/json"},
                                 method="POST")
    try:
        with urllib.request.urlopen(req, timeout=300) as r:
            return r.status, json.loads(r.read())
    except urllib.error.HTTPError as e:
        raw = e.read() or b"{}"
        try:
            return e.code, json.loads(raw)
        except ValueError:
            return e.code, {"error": raw[:300].decode(errors="replace")}
 def systemone(body):
    mx, total = admit(body)
    code, obj = post_up("/v1/systemone", body)
    if isinstance(obj, dict):
        obj.setdefault("_admission", {}).update({"max_prompt_tokens": mx, "shim_accounted_total": total, "cap": CAP})
    return code, obj
 def check_decision(d, i):
    if not isinstance(d, dict) or not isinstance(d.get("id"), str) or not isinstance(d.get("question"), str):
        raise Refused(422, {"error": f"decision {i}: needs string id and question"})
    opts = d.get("options")
    if not isinstance(opts, list) or not MIN_OPTIONS <= len(opts) <= MAX_OPTIONS:
        raise Refused(422, {"error": f"decision {d['id']!r}: options must contain {MIN_OPTIONS}-{MAX_OPTIONS} entries"})
    if any(not isinstance(o, dict) or not isinstance(o.get("id"), str) or not isinstance(o.get("description"), str)
           for o in opts):
        raise Refused(422, {"error": f"decision {d['id']!r}: every option needs string id and description"})
    if len({o["id"] for o in opts}) != len(opts):
        raise Refused(422, {"error": f"decision {d['id']!r}: option ids must be unique"})
    if d.get("orderings", "none") != "none":
        raise Refused(422, {"error": f"decision {d['id']!r}: orderings {d.get('orderings')!r} is not served by this "
                                     "seat (h2o); send orderings 'none' and combine on the client"})
 def decide(state, decisions):
    """semif /decide/shared on top of /v1/systemone: intern-decision-serve's mapping (one Jev choice question per
    decision, calls of at most 16, fields q / q1..qN), re-keyed into its result shape."""
    if not isinstance(decisions, list) or not 1 <= len(decisions) <= MAX_DECISIONS:
        raise Refused(422, {"error": f"decisions must be a list of 1..{MAX_DECISIONS}"})
    for i, d in enumerate(decisions):
        check_decision(d, i)
    started, results, sizes, tokens = time.perf_counter(), [], [], []
    for k in range(0, len(decisions), MAX_QUESTIONS_PER_CALL):
        chunk = decisions[k:k + MAX_QUESTIONS_PER_CALL]
        fields = ["q"] if len(chunk) == 1 else [f"q{j + 1}" for j in range(len(chunk))]
        questions = {f: {"type": "choice", "instructions": d["question"],
                         "criteria": {o["id"]: o["description"] for o in d["options"]}} for f, d in zip(fields, chunk)}
        code, resp = systemone({"state": state, "questions": questions})
        if code != 200:
            raise Refused(code, resp if isinstance(resp, dict) else {"error": str(resp)})
        sizes.append(len(chunk))
        tokens.append((resp.get("usage") or {}).get("input_tokens"))
        for f, d in zip(fields, chunk):
            answer = (resp.get("answers") or {}).get(f)
            ids = [o["id"] for o in d["options"]]
            probs = (answer or {}).get("probabilities") or {}
            if answer is None or any(i not in probs for i in ids):
                raise Refused(500, {"error": f"scoring_failed: the model's answer for field {f!r} lacks option ids"})
            results.append({"id": d["id"], "option_ids": ids, "probabilities": [probs[i] for i in ids],
                            "top": answer.get("choice"), "confidence": answer.get("confidence"),
                            "calibration": resp.get("calibration"), "native": answer,
                            "input_tokens": tokens[-1], "model": MODEL,
                            "call": {"index": len(sizes) - 1, "field": f, "questions": len(chunk)}})
    return {"results": results, "timing": {"total_seconds": time.perf_counter() - started, "calls": len(sizes),
                                           "questions_per_call": sizes, "input_tokens": tokens}}
 class H(BaseHTTPRequestHandler):
    protocol_version = "HTTP/1.1"
    def log_message(self, fmt, *args):
        sys.stderr.write("%s - %s\n" % (self.address_string(), fmt % args))
    def _send(self, code, obj):
        b = json.dumps(obj).encode()
        self.send_response(code)
        self.send_header("Content-Type", "application/json")
        self.send_header("Content-Length", str(len(b)))
        self.end_headers()
        self.wfile.write(b)
    def _authorized(self):
        scheme, _, tok = (self.headers.get("Authorization") or "").partition(" ")
        return scheme.lower() == "bearer" and hmac.compare_digest(tok.strip().encode(), TOKEN.encode())
    def do_GET(self):
        path = self.path.split("?")[0]
        if path == "/health":
            try:
                with urllib.request.urlopen(UP + "/health", timeout=10) as r:
                    shim = json.loads(r.read())
            except Exception as e:  # noqa: BLE001
                return self._send(503, {"status": "unavailable", "model": MODEL, "error": f"shim: {type(e).__name__}"})
            ok = bool(shim.get("ok"))
            return self._send(200 if ok else 503, {
                "status": "ok" if ok else "starting", "model": MODEL, "cap_tokens": CAP, "admission": "refuse, never trim",
                "endpoints": ["/decide", "/decide/shared", "/v1/systemone"], "max_decisions": MAX_DECISIONS,
                "max_questions_per_call": MAX_QUESTIONS_PER_CALL, "orderings": ["none"], "backend": shim})
        if not self._authorized():
            return self._send(401, {"error": "unauthorized"})
        if path == "/v1/models":
            try:
                with urllib.request.urlopen(UP + "/v1/models", timeout=10) as r:
                    return self._send(r.status, json.loads(r.read()))
            except urllib.error.HTTPError as e:
                return self._send(e.code, {"error": "upstream"})
        return self._send(404, {"error": "not found"})
    def do_POST(self):
        path = self.path.split("?")[0]
        if not self._authorized():
            return self._send(401, {"error": "unauthorized"})
        n = int(self.headers.get("Content-Length", 0) or 0)
        if n > 4 * 1024 * 1024:
            return self._send(413, {"error": "body over 4 MiB"})
        try:
            body = json.loads(self.rfile.read(n) or b"null")
        except ValueError:
            return self._send(422, {"error": "body is not JSON"})
        if not isinstance(body, dict):
            return self._send(422, {"error": "body must be a JSON object"})
        try:
            if path == "/v1/systemone":
                return self._send(*systemone(body))
            if path == "/api/alpha/decisions":
                admit(body)
                return self._send(*post_up(path, body))
            if path == "/decide/shared":
                return self._send(200, decide(body.get("state"), body.get("decisions")))
            if path == "/decide":
                d = {k: body.get(k) for k in ("id", "question", "options", "orderings") if k in body}
                out = decide(body.get("state"), [d])
                return self._send(200, {**out["results"][0], "total_seconds": out["timing"]["total_seconds"]})
        except Refused as r:
            return self._send(r.code, r.obj)
        return self._send(404, {"error": "not found"})
 ThreadingHTTPServer((BIND, PORT), H).serve_forever()
@@ -0,0 +1,13 @@
 #!/bin/bash
 # One container: the unmodified vendor shim on loopback, then the front. If either dies, the container exits and
 # Docker restarts it (restart: unless-stopped). Both inherit the container's whole environment, so the shim's own
 # settings (SHIM_*, e.g. SHIM_NOUL_FLOOR, and LIGHTNING_*, e.g. LIGHTNING_IMAGE_URLS=off) set in compose reach it.
 set -eu
 python3 /model/h2o_lightning_shim.py --config /model/serve_config.json --vllm "$SHIM_VLLM" \
  --host 127.0.0.1 --port "$SHIM_PORT" &
 SHIM=$!
 python3 /front/front.py &
 FRONT=$!
 wait -n "$SHIM" "$FRONT" 2>/dev/null || true
 kill "$SHIM" "$FRONT" 2>/dev/null || true
 exit 1
@@ -1,6 +1,128 @@
-# intern-decision
+# intern-decision — the decision seat (fv-ml1 GPU 1, :8033)
-> **LIVE since 2026-09-30 0941 PT** on fv-ml1 GPU 1, with cap 9.0 GiB and `MAX_TOKENS` 7,168.
+> **LIVE: h2o-lightning-4b since 2026-10-12 1704 PT** (Prime 1640, via albok-dev). `compose.yaml` = `compose.h2o.yaml`
 > (`switch-to-h2o.sh` copies one over the other). Nemi's judge runs the **re-fitted v2 note question**
 > (albok-nemi 0.1.18, contract 4.4 rev 1.15) at concurrency 4.
 >
 > **What happened on 2026-10-12:**
 > 1. **1311–1403, h2o's first stint, rolled back.** Under the judge's v1 question, h2o flagged 242 texts adult
 >    against INC's 88, over 10,841 identical texts.
 > 2. **1428–1619, the investigation on GPU 3.** The v1 question asked two things at once: INC answered it
 >    inconsistently and h2o answered it as written. The shim's floor was ruled out, with identical verdicts on and
 >    off. Prime ruled "text itself".
 > 3. **Under v2 the two seats agree:** work notes INC 0/22, h2o 1/22 at 0.52; nulls 0/18 on both. fleet/memory adult
 >    went 86 → 0 on INC.
 > 4. **Rollback:** `rollback-to-inc.sh`, and set the judge's concurrency back to 1 (see "Switching").
 >
 > The section headed "The live seat: Intern-Decision-4B" further down describes INC, the rollback target.
 ### The h2o seat (`compose.h2o.yaml`, live since 1704)
 | | |
 |---|---|
 | **URL** | `http://10.251.50.54:8033` = `http://intern-decision.fv.internal:8033`. `/health` is open; everything else needs `Authorization: Bearer $(secret get intern-decision/api-token)` (the same value as `albok/judge-token`) |
 | **Model** | [`h2oai/h2o-lightning-4b`](https://huggingface.co/h2oai/h2o-lightning-4b) @ `b43f759b6385d21182f87c3829931ec3d4dfa98a`, from `/tank/aimodels/h2o-lightning-4b-b43f759b` (read-only). `model.safetensors` sha256 `2d51561f…d4d11497b6`; tokenizer, shim, serve_config, config and chat template all equal Hugging Face at that revision |
 | **Serving** | `vllm/vllm-openai:v0.30.0` (container `intern-decision-vllm`, loopback `:18299`), args exactly as benched: `--max-model-len 40960 --max-num-seqs 16`, the vendor's vision limits, `VLLM_USE_FLASHINFER_SAMPLER=0` |
 | **KV** | **pinned in bytes**: `--kv-cache-memory 4810363371` (4.48 GiB = 135,558 tokens, 3.31× concurrency at 40,960). GPU 1 is shared with meromero-rp and gen-small, so a util-sized KV would drift on a co-boot. `--gpu-memory-utilization 0.1506` now only gates vLLM's boot check |
 | **Front** | container `intern-decision` (python:3.12-slim, host network, `:8033`), running `conf/front/run.sh`: the vendor's **unmodified** `h2o_lightning_shim.py` on loopback `:18291`, behind `services/h2o-decision-front/front.py` |
 | **State** | none beyond the `intern-decision_vllm-cache` volume (compile cache) |
 | **Rollback** | `compose.inc.yaml` is the INC compose, same `.env`. See **Rollback to INC** below |
 ### What the front adds (`services/h2o-decision-front`)
 - **Admission, refuse-not-trim, at 32,000 tokens.** It counts with h2o's own tokenizer through vLLM `/tokenize`, over
  the shim's own render of each question, and checks both the largest prompt and the shim's own total (shared prefix
  plus tails). Over the cap gets `422 "admission: N max-prompt / M total tokens against the 32000-token seat limit;
  refused, not trimmed"`. The logic is infra-hermes's `tools/h2o-admission/admission_proxy.py` (30f36f2), unchanged.
 - **The bearer token** on every path except `GET /health`.
 - **`/decide` and `/decide/shared`**, translated onto `/v1/systemone`. This keeps albok-nemi's judge, the seat's only
  production caller, working unchanged. One call holds at most 16 questions and is split greedily beyond that, as
  INC did. `orderings` other than `none` → 422. Each result carries `probabilities`, `top`, `confidence`,
  `calibration`, `native` (the shim's answer object, unchanged), `input_tokens`, `model` and `call`. INC's
  `prompt_sha256`, `readout` and timing fields are gone.
 - **`/health`** reports `model {name, source, revision}`, the cap, the endpoints and vLLM's `max_model_len`, and
  returns 503 when the shim is not healthy.
 - `/v1/systemone` and `/api/alpha/decisions` pass through after admission, and `/v1/models` passes through as is.
 ### Verification at the 1311 cutover (2026-10-12, 1311–1318 PT, on fv-ml1 loopback)
 | check | result | expected |
 |---|---|---|
 | `/health` | `ok`, model `h2o-lightning-4b` @ `b43f759b…`, cap 32,000, `max_model_len` 40,960 | the pin |
 | over the cap: one question over a ~40k-token state | **422**, "40078 max-prompt / 40078 total … refused, not trimmed" | 422 |
 | 16 questions at the cap (31,528 max-prompt / 31,999 shim total), warm, N = 3 | **200**; 2.105 / 2.061 / 1.999 s | ~2 s (bench) |
 | `/decide/shared` smoke: two yes/no decisions over a line holding a phone number and a street address | 200 in 0.52 s; p(yes) explicit 0.052, personal info 0.876 | the right way round |
 | GPU 1, h2o's vLLM process, at rest / after the 16q@32k calls | 14,420 / 15,590 MiB | budget 16,847, bench peak 15,566 |
 The 16q@32k and `/decide/shared` rows show the path works; the speed and memory numbers are infra-hermes's, below.
 **infra-hermes, 1316–1318, same seat**: the bench's own harness (jevbench.cli v1.2.16 with the typesafe adapter, plus
 the Amendment 1 gate probe driver). Bearer passed by `--key-env`. Artifacts are on fv-ml1 in `/tank/bench/cutover/jevbench/`
 and `/tank/bench/cutover-probe.json`.
 | check | result | reference |
 |---|---|---|
 | **JevBench public, one pass** | **205/231** | 205 ×3 in the Amendment 1 bench |
 | depth cells: evidence at depth 1k / 8k / 16k / 30k | `supported` at every depth (p 0.89–0.94), admitted at 31,505 / 31,212 / 12,208 / 31,744 tokens | the 16k cell was under-filled by the probe's calibration, not by the seat. It passed at 29,508 in Amendment 1, and the 30k depth passes here |
 | over the cap | 422 at 32,020 / 32,046 / 33,463 / 39,845 tokens, each naming the count | |
 | `usage.input_tokens` against the admitted count | equal in every admitted cell | the admission honesty property |
 | largest admissible, 1 question / 16 questions (31,417 / 31,907 tokens) | first 251 / 2,395 ms; warm median 226 / **2,159 ms** | bench front 2,033 ms; the bar is 2.93 s. The +126 ms is the front's translation and auth |
 | **4 concurrent 16q × 32k calls** | all 4 correct, wall 7,178 ms. GPU 1 peak at 50 ms sampling: **95,178 MiB** = meromero 43,998 + gen-small 34,994 + **seat 16,164** | seat budget 16,847, so **683 MiB headroom**; the bench peak was 15,566 |
 ⚠ **683 MiB is the seat's margin under 4-way 32k load.** That is tighter than the bench's peak, because of the byte-pinned KV
 size and concurrent growth. It passes, but it is the number to re-measure before anything else lands on GPU 1.
 ### Before any (re)create
 GPU 1 holds meromero-rp (~44.0 GB) and gen-small (~35.0 GB). Read the free memory first, with
 `scripts/oneoff/fv-ml1-gpucheck.sh 1 0.1506 intern-decision-vllm`, and stop if it fails. The vLLM container takes about
 3 minutes to reach healthy (graph capture). The front waits on it, and the healthcheck has a 300 s start period.
 ### Images: inline only (Prime 2026-10-12, via brokkr)
 The h2o seat must never fetch an image URL. `compose.h2o.yaml` sets `LIGHTNING_IMAGE_URLS=off` on the front, and the
 vendor shim then refuses every image URL with 422 "image URLs are not accepted here". Inline `data:` URIs and base64
 keep working. vLLM also runs with `--allowed-media-domains media.invalid` as a backstop. That is a reserved name that
 never resolves, and vLLM 0.30.0 checks every http(s) media URL against it before fetching, while `data:` URLs bypass
 the check.
 The shim switch was verified live at 1402, on the h2o seat with the front recreated alone:
 - the same `https://media.invalid/red.png` request went from "could not be fetched" (a fetch was attempted) to "not
  accepted here" (refused before any fetch);
 - inline red still answered red at 0.984;
 - before the change, inline red/blue/green answered at 0.98+ and no-image came out near-uniform.
 The vLLM backstop was first exercised on the GPU 3 eval seat at 1428 (same image and flag): an https image URL
 posted straight to vLLM got 400 "The URL must be from one of the allowed domains: ['media.invalid']", with no fetch.
 Repeat that check on this seat's loopback port (`:18299`) on the next switch. INC needs neither setting, because its served checkpoint has no vision tower.
 ### Switching between the two
 The 1704 switch: `switch-to-h2o.sh` ran from 170223 (rc 0) and h2o was healthy at 170454. The 1705 checks: `/health` at the
 pin, with floor 0.801; `/decide/shared` 200; a 422 over the cap; inline red → red at 0.984; an https image URL refused
 by the shim; the vLLM backstop refusing example.com (400); GPU 1 resting at 14,430 MiB.
 Both run ON fv-ml1 and share one `.env`.
 - **To h2o:** `sudo bash /opt/docker/compose/intern-decision/switch-to-h2o.sh`. It stops INC and checks that GPU 1 has
  16,900 MiB free (h2o's measured budget). If not, it restarts INC and exits 2. Otherwise it copies
  `compose.h2o.yaml` over `compose.yaml` and waits up to 300 s for `/health`. Only do this once albok has re-fitted
  the judge.
 - **To INC:** `sudo bash /opt/docker/compose/intern-decision/rollback-to-inc.sh`. It stops h2o, copies `compose.inc.yaml`
  over `compose.yaml` and waits up to 180 s. Image `intern-decision-serve:0.1.3` and its
  `intern-decision_triton-cache` volume survive the `down`. Used 2026-10-12, 140328 → healthy 140356.
 After either switch, make `stacks/intern-decision/compose.yaml` match the host again (it is h2o since 1704).
 ⚠ **Tell albok-dev on either switch: Nemi's judge concurrency depends on the seat.** INC scores on one inference
 thread (INV-2, ~16 calls/s at most), so the judge runs with concurrency 1 against it (albok's `nemi.yaml`, since
 2026-10-12). With 4 in flight on INC, the 1533–1619 walk queued behind itself and lost 339 calls to 30 s timeouts. On
 h2o, vLLM batches, so it goes back to 4.
 ---
 # The live seat: Intern-Decision-4B (2026-09-30 → 2026-10-12 1308, and 1403–1702; now the rollback target)
 > **LIVE from 2026-09-30 0941 PT to 2026-10-12 1308 PT, and again since 2026-10-12 1403 PT** on fv-ml1 GPU 1, with cap 9.0 GiB and `MAX_TOKENS` 7,168.
 > Acceptance on the live service is below. `semif` is stopped and kept as the rollback.
 **Intern-Decision-4B typed decisions** on **fv-ml1 GPU 1**, the utility card beside `vllm-coder`,
@@ -293,7 +415,7 @@ Read `docker logs intern-decision`. Record the deploy with `scripts/ops-log`.
 Re-run the bench's sets through the service (`services/intern-decision-serve/acceptance/`), and
 re-measure the VRAM before changing the cap.
-## Rollback
+## Rollback (INC → semif)
 `semif` is **stopped, not removed**, and is kept as the rollback. Only on Prime's word:
@@ -0,0 +1,75 @@
 # STAGED, NOT LIVE (since the 2026-10-12 1403 rollback): the h2o-lightning-4b decision seat, which switch-to-h2o.sh
 # copies over compose.yaml. It served 1311-1403 on 2026-10-12 (Prime via brokkr) and was rolled back when Nemi's judge
 # flagged 242 texts adult against INC's 88 at fleet scale. Address intern-decision.fv.internal:8033 (brokkr rule 8).
 # h2o-lightning-4b @ b43f759b. Stock vLLM 0.30.0 + the vendor's UNMODIFIED shim + services/h2o-decision-front (copied to
 # conf/front): 32,000-token REFUSE-not-trim admission, the seat's bearer token (the same INTERN_DECISION_API_TOKEN),
 # /decide + /decide/shared translated onto /v1/systemone, /health model {name, revision}.
 # Rollback to Intern-Decision-4B: compose.inc.yaml (the previous compose, same .env): see the stack README.
 # GPU 1 is shared (meromero-rp, gen-small): KV is pinned in BYTES (4.48 GiB = 135,070 tokens, the benched size);
 # util only gates the boot check. Gate on scripts/oneoff/fv-ml1-gpucheck.sh before any (re)create.
 # Images: INLINE ONLY (Prime 2026-10-12). LIGHTNING_IMAGE_URLS=off makes the shim refuse every image URL; vLLM's
 # --allowed-media-domains media.invalid (a reserved name that never resolves) is the backstop if the shim ever changes.
 name: intern-decision
 services:
  vllm:
    image: vllm/vllm-openai:v0.30.0
    container_name: intern-decision-vllm
    restart: unless-stopped
    ports:
      - "127.0.0.1:18299:8000"
    environment:
      VLLM_USE_FLASHINFER_SAMPLER: "0"
    volumes:
      - /tank/aimodels/h2o-lightning-4b-b43f759b:/model:ro
      - vllm-cache:/root/.cache/vllm
    command: ["--model", "/model", "--served-model-name", "h2oai/h2o-lightning-4b",
              "--max-model-len", "40960", "--max-num-seqs", "16",
              "--gpu-memory-utilization", "0.1506", "--kv-cache-memory", "4810363371",
              "--limit-mm-per-prompt", "{\"image\": 4, \"video\": 0}",
              "--mm-processor-kwargs", "{\"max_pixels\": 1605632}",
              "--allowed-media-domains", "media.invalid"]
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["${GPU_ID:-1}"]
              capabilities: [gpu]
  front:
    image: python:3.12-slim
    container_name: intern-decision
    restart: unless-stopped
    network_mode: host
    depends_on: [vllm]
    environment:
      FRONT_TOKEN: ${INTERN_DECISION_API_TOKEN:?set INTERN_DECISION_API_TOKEN}
      SHIM_VLLM: http://127.0.0.1:18299
      SHIM_PORT: "18291"
      SHIM_NOUL_FLOOR: "0.801"
      LIGHTNING_IMAGE_URLS: "off"   # Prime 2026-10-12: inline images only, the seat never fetches an image URL
      FRONT_SHIM: http://127.0.0.1:18291
      FRONT_VLLM: http://127.0.0.1:18299
      FRONT_BIND: 0.0.0.0
      FRONT_PORT: ${PORT:-8033}
      FRONT_CAP: "32000"
      FRONT_SHIM_FILE: /model/h2o_lightning_shim.py
      FRONT_SERVE_CONFIG: /model/serve_config.json
      FRONT_MODEL_REVISION: b43f759b6385d21182f87c3829931ec3d4dfa98a
    volumes:
      - /tank/aimodels/h2o-lightning-4b-b43f759b:/model:ro
      - /opt/docker/conf/intern-decision/front:/front:ro
    command: ["bash", "/front/run.sh"]
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8033/health', timeout=5)"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 300s
    labels:
      - homepage.group=AI - Eval & Retrieval
      - homepage.name=Decision seat (h2o-lightning-4b)
      - homepage.icon=mdi-scale-balance
      - homepage.description=Typed decisions, 32k refusing (h2o-lightning-4b on vLLM, fv-ml1 GPU1; /decide/shared kept)
      - homepage.href=http://${HOST_IP:-10.251.50.54}:${PORT:-8033}/health
 volumes:
  vllm-cache:
@@ -0,0 +1,74 @@
 # intern-decision: Intern-Decision-4B (internlm, Apache-2.0) behind intern-decision-serve, on
 # fv-ml1 GPU 1 (the utility card, beside vllm-coder and the erp/meromero seats; scriberr moved to
 # GPU 3 on 2026-09-30 1322, Prime, to free this card's headroom for 32k-token calls).
 # Replaces semif (Prime, 2026-09-30: "replace semif with intern-decision now").
 #
 # One forward pass per call, scored by the checkpoint's OWN inference.py (sha256-pinned); the
 # service keeps semif-serve's HTTP surface (/decide, /decide/shared, /health). Service code +
 # contract: services/intern-decision-serve/ (intern-decision-serve.contract.md). Image built on
 # fv-ml1 from that dir.
 #
 # ⚠ VRAM_CAP_GIB is a HARD cap on torch's allocator (per-process memory fraction), set so the
 #   container's WHOLE nvidia-smi footprint, CUDA context included, fits GPU 1's free memory beside
 #   the static vLLM seats (infra-ops budget, 2026-09-30); MAX_TOKENS keeps every accepted call under the cap. A request that needs more
 #   gets 503 out_of_memory and the service stays up. See the README before changing it.
 #
 # .env (tunables): IMAGE, PORT, GPU_ID, VRAM_CAP_GIB, MAX_TOKENS, HOST_IP, INTERN_DECISION_API_TOKEN
 #   (vault intern-decision/api-token, >= 32 chars).
 name: intern-decision
 services:
  intern-decision:
    image: ${IMAGE:?set IMAGE}
    container_name: intern-decision
    restart: unless-stopped
    ports:
      - "${PORT:-8033}:8000"
    environment:
      INTERN_DECISION_API_TOKEN: ${INTERN_DECISION_API_TOKEN:?set INTERN_DECISION_API_TOKEN}
      INTERN_DECISION_DEVICE: cuda
      INTERN_DECISION_VRAM_CAP_GIB: ${VRAM_CAP_GIB:?set VRAM_CAP_GIB}
      # Coupled to VRAM_CAP_GIB: the largest call measured to fit under the cap (README "VRAM").
      INTERN_DECISION_MAX_TOKENS: ${MAX_TOKENS:?set MAX_TOKENS}
      INTERN_DECISION_MAX_DECISIONS: ${MAX_DECISIONS:-64}
      # POSTs in progress (queued + scoring) before new ones get 429 busy.
      INTERN_DECISION_MAX_QUEUE: ${MAX_QUEUE:-32}
    volumes:
      # Pinned weights AND the checkpoint's inference.py, read offline. Never downloads.
      - /tank/aimodels/huggingface:/hf:ro
      # Triton/fla autotune cache (README "Cold-shape latency"): the ~6.5 s first-call-per-bucket
      # autotune results survive RECREATES (deploys, image upgrades, .env edits), not just restarts.
      # The mount point exists in the image owned by 10001, so the fresh volume is intern-writable.
      - triton-cache:/tmp/triton-cache
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["${GPU_ID:-1}"]
              capabilities: [gpu]
    healthcheck:
      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=5)"]
      interval: 30s
      timeout: 10s
      retries: 3
      # Startup loads ~9 GB of weights and scores the warm-up three times before it serves.
      start_period: 300s
    networks:
      - tnet
    labels:
      - homepage.group=AI - Eval & Retrieval
      - homepage.name=Intern-Decision — typed decisions
      - homepage.icon=mdi-scale-balance
      - homepage.description=Typed decisions from one forward pass (Intern-Decision-4B, fv-ml1 GPU1)
      - homepage.href=http://${HOST_IP:-10.251.50.54}:${PORT:-8033}/health
 networks:
  tnet:
    name: traefik-net
    external: true
 volumes:
  # -> intern-decision_triton-cache on the host
  triton-cache:
@@ -1,46 +1,33 @@
-# intern-decision: Intern-Decision-4B (internlm, Apache-2.0) behind intern-decision-serve, on
+# STAGED, NOT LIVE (since the 2026-10-12 1403 rollback): the h2o-lightning-4b decision seat, which switch-to-h2o.sh
-# fv-ml1 GPU 1 (the utility card, beside vllm-coder and the erp/meromero seats; scriberr moved to
+# copies over compose.yaml. It served 1311-1403 on 2026-10-12 (Prime via brokkr) and was rolled back when Nemi's judge
-# GPU 3 on 2026-09-30 1322, Prime, to free this card's headroom for 32k-token calls).
+# flagged 242 texts adult against INC's 88 at fleet scale. Address intern-decision.fv.internal:8033 (brokkr rule 8).
-# Replaces semif (Prime, 2026-09-30: "replace semif with intern-decision now").
+# h2o-lightning-4b @ b43f759b. Stock vLLM 0.30.0 + the vendor's UNMODIFIED shim + services/h2o-decision-front (copied to
-#
+# conf/front): 32,000-token REFUSE-not-trim admission, the seat's bearer token (the same INTERN_DECISION_API_TOKEN),
-# One forward pass per call, scored by the checkpoint's OWN inference.py (sha256-pinned); the
+# /decide + /decide/shared translated onto /v1/systemone, /health model {name, revision}.
-# service keeps semif-serve's HTTP surface (/decide, /decide/shared, /health). Service code +
+# Rollback to Intern-Decision-4B: compose.inc.yaml (the previous compose, same .env): see the stack README.
-# contract: services/intern-decision-serve/ (intern-decision-serve.contract.md). Image built on
+# GPU 1 is shared (meromero-rp, gen-small): KV is pinned in BYTES (4.48 GiB = 135,070 tokens, the benched size);
-# fv-ml1 from that dir.
+# util only gates the boot check. Gate on scripts/oneoff/fv-ml1-gpucheck.sh before any (re)create.
-#
+# Images: INLINE ONLY (Prime 2026-10-12). LIGHTNING_IMAGE_URLS=off makes the shim refuse every image URL; vLLM's
-# ⚠ VRAM_CAP_GIB is a HARD cap on torch's allocator (per-process memory fraction), set so the
+# --allowed-media-domains media.invalid (a reserved name that never resolves) is the backstop if the shim ever changes.
 #   container's WHOLE nvidia-smi footprint, CUDA context included, fits GPU 1's free memory beside
 #   the static vLLM seats (infra-ops budget, 2026-09-30); MAX_TOKENS keeps every accepted call under the cap. A request that needs more
 #   gets 503 out_of_memory and the service stays up. See the README before changing it.
 #
 # .env (tunables): IMAGE, PORT, GPU_ID, VRAM_CAP_GIB, MAX_TOKENS, HOST_IP, INTERN_DECISION_API_TOKEN
 #   (vault intern-decision/api-token, >= 32 chars).
 name: intern-decision
 services:
-  intern-decision:
+  vllm:
-    image: ${IMAGE:?set IMAGE}
+    image: vllm/vllm-openai:v0.30.0
-    container_name: intern-decision
+    container_name: intern-decision-vllm
    restart: unless-stopped
    ports:
-      - "${PORT:-8033}:8000"
+      - "127.0.0.1:18299:8000"
    environment:
-      INTERN_DECISION_API_TOKEN: ${INTERN_DECISION_API_TOKEN:?set INTERN_DECISION_API_TOKEN}
+      VLLM_USE_FLASHINFER_SAMPLER: "0"
      INTERN_DECISION_DEVICE: cuda
      INTERN_DECISION_VRAM_CAP_GIB: ${VRAM_CAP_GIB:?set VRAM_CAP_GIB}
      # Coupled to VRAM_CAP_GIB: the largest call measured to fit under the cap (README "VRAM").
      INTERN_DECISION_MAX_TOKENS: ${MAX_TOKENS:?set MAX_TOKENS}
      INTERN_DECISION_MAX_DECISIONS: ${MAX_DECISIONS:-64}
      # POSTs in progress (queued + scoring) before new ones get 429 busy.
      INTERN_DECISION_MAX_QUEUE: ${MAX_QUEUE:-32}
    volumes:
-      # Pinned weights AND the checkpoint's inference.py, read offline. Never downloads.
+      - /tank/aimodels/h2o-lightning-4b-b43f759b:/model:ro
-      - /tank/aimodels/huggingface:/hf:ro
+      - vllm-cache:/root/.cache/vllm
-      # Triton/fla autotune cache (README "Cold-shape latency"): the ~6.5 s first-call-per-bucket
+    command: ["--model", "/model", "--served-model-name", "h2oai/h2o-lightning-4b",
-      # autotune results survive RECREATES (deploys, image upgrades, .env edits), not just restarts.
+              "--max-model-len", "40960", "--max-num-seqs", "16",
-      # The mount point exists in the image owned by 10001, so the fresh volume is intern-writable.
+              "--gpu-memory-utilization", "0.1506", "--kv-cache-memory", "4810363371",
-      - triton-cache:/tmp/triton-cache
+              "--limit-mm-per-prompt", "{\"image\": 4, \"video\": 0}",
              "--mm-processor-kwargs", "{\"max_pixels\": 1605632}",
              "--allowed-media-domains", "media.invalid"]
    deploy:
      resources:
        reservations:
@@ -48,27 +35,41 @@ services:
            - driver: nvidia
              device_ids: ["${GPU_ID:-1}"]
              capabilities: [gpu]
  front:
    image: python:3.12-slim
    container_name: intern-decision
    restart: unless-stopped
    network_mode: host
    depends_on: [vllm]
    environment:
      FRONT_TOKEN: ${INTERN_DECISION_API_TOKEN:?set INTERN_DECISION_API_TOKEN}
      SHIM_VLLM: http://127.0.0.1:18299
      SHIM_PORT: "18291"
      SHIM_NOUL_FLOOR: "0.801"
      LIGHTNING_IMAGE_URLS: "off"   # Prime 2026-10-12: inline images only, the seat never fetches an image URL
      FRONT_SHIM: http://127.0.0.1:18291
      FRONT_VLLM: http://127.0.0.1:18299
      FRONT_BIND: 0.0.0.0
      FRONT_PORT: ${PORT:-8033}
      FRONT_CAP: "32000"
      FRONT_SHIM_FILE: /model/h2o_lightning_shim.py
      FRONT_SERVE_CONFIG: /model/serve_config.json
      FRONT_MODEL_REVISION: b43f759b6385d21182f87c3829931ec3d4dfa98a
    volumes:
      - /tank/aimodels/h2o-lightning-4b-b43f759b:/model:ro
      - /opt/docker/conf/intern-decision/front:/front:ro
    command: ["bash", "/front/run.sh"]
    healthcheck:
-      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=5)"]
+      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8033/health', timeout=5)"]
      interval: 30s
      timeout: 10s
      retries: 3
      # Startup loads ~9 GB of weights and scores the warm-up three times before it serves.
      start_period: 300s
    networks:
      - tnet
    labels:
      - homepage.group=AI - Eval & Retrieval
-      - homepage.name=Intern-Decision — typed decisions
+      - homepage.name=Decision seat (h2o-lightning-4b)
      - homepage.icon=mdi-scale-balance
-      - homepage.description=Typed decisions from one forward pass (Intern-Decision-4B, fv-ml1 GPU1)
+      - homepage.description=Typed decisions, 32k refusing (h2o-lightning-4b on vLLM, fv-ml1 GPU1; /decide/shared kept)
      - homepage.href=http://${HOST_IP:-10.251.50.54}:${PORT:-8033}/health
 networks:
  tnet:
    name: traefik-net
    external: true
 volumes:
-  # -> intern-decision_triton-cache on the host
+  vllm-cache:
  triton-cache:
@@ -0,0 +1,231 @@
 #!/usr/bin/env python3
 """The front of the h2o-lightning-4b decision seat (fv-ml1). Stdlib only.
 It sits in front of the UNMODIFIED shipped h2o_lightning_shim.py (which sits in front of stock vLLM 0.30.0) and adds
 the three things the seat contract needs and the vendor stack does not have:
 1. ADMISSION (from infra-hermes's tools/h2o-admission/admission_proxy.py @ 30f36f2, logic unchanged): the shim
   silently head-and-tail-trims state above ~32,000 tokens, so every decision request is counted first with h2o's own
   tokenizer (vLLM /tokenize over the shim's own per-question renders) and refused with 422 + the count when either the
   largest prompt or the shim-accounted total (shared head + tails) exceeds FRONT_CAP. Refused, never trimmed.
 2. THE SEAT'S AUTH: a bearer token on everything but GET /health, as intern-decision-serve has (OPEN_PATHS /health).
 3. CONTINUITY for /decide and /decide/shared (semif's surface, albok-nemi's sensitivity judge): each decision becomes
   one Jev `choice` question exactly as intern-decision-serve builds it ({type: choice, instructions: question,
   criteria: {option id: description}}), asked through /v1/systemone in calls of at most 16 questions, and re-keyed
   into the semif result shape the judge reads (results[i].native.probabilities[option id]). orderings other than
   "none" are refused (422): the judge does its own forward/reverse. /health carries {model: {name, revision}}, which
   the judge stamps into its cache key.
 Env: FRONT_TOKEN (required), FRONT_BIND, FRONT_PORT, FRONT_SHIM (shim URL), FRONT_VLLM (vLLM URL), FRONT_CAP (32000),
 FRONT_SHIM_FILE (vendor shim path, imported for its renderers), FRONT_SERVE_CONFIG, FRONT_MODEL_NAME,
 FRONT_MODEL_SOURCE, FRONT_MODEL_REVISION, FRONT_MAX_DECISIONS (64).
 """
 import hmac, importlib.util, json, os, sys, time, urllib.error, urllib.request
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 TOKEN = os.environ["FRONT_TOKEN"]
 BIND = os.environ.get("FRONT_BIND", "127.0.0.1")
 PORT = int(os.environ.get("FRONT_PORT", "8034"))
 UP = os.environ.get("FRONT_SHIM", "http://127.0.0.1:18191")
 VLLM = os.environ.get("FRONT_VLLM", "http://127.0.0.1:18199")
 CAP = int(os.environ.get("FRONT_CAP", "32000"))
 SHIM_FILE = os.environ.get("FRONT_SHIM_FILE", "/model/h2o_lightning_shim.py")
 SERVE_CONFIG = os.environ.get("FRONT_SERVE_CONFIG", "/model/serve_config.json")
 MODEL = {"name": os.environ.get("FRONT_MODEL_NAME", "h2o-lightning-4b"),
         "source": os.environ.get("FRONT_MODEL_SOURCE", "h2oai/h2o-lightning-4b"),
         "revision": os.environ.get("FRONT_MODEL_REVISION", "b43f759b6385d21182f87c3829931ec3d4dfa98a")}
 MAX_DECISIONS = int(os.environ.get("FRONT_MAX_DECISIONS", "64"))
 MAX_QUESTIONS_PER_CALL = 16
 MIN_OPTIONS, MAX_OPTIONS = 2, 16
 spec = importlib.util.spec_from_file_location("h2o_shim", SHIM_FILE)
 shim_mod = importlib.util.module_from_spec(spec)
 sys.modules["h2o_shim"] = shim_mod
 spec.loader.exec_module(shim_mod)
 CFG = shim_mod.Contract(json.load(open(SERVE_CONFIG)))
 CFG.max_state_tokens = 10**9   # count the UNTRUNCATED render; the shim's own trim is what admission polices
 class Refused(Exception):
    def __init__(self, code, obj):
        super().__init__(obj)
        self.code, self.obj = code, obj
 def sent_tokens(body):
    """(max tokens of one question's prompt, shim-accounted total = shared head + tails), counted by vLLM's own
    tokenizer over the shim's own per-question renders (admission_proxy.py @ 30f36f2, unchanged)."""
    state = body.get("state") or ""
    if isinstance(state, (dict, list)):
        state = json.dumps(state)
    state_txt = shim_mod.render_state(state, CFG.state_compact)
    all_toks = []
    for qid, q in body.get("questions", {}).items():
        prompt_text = CFG.question_prompt(state_txt, q, qid)[0]
        req = urllib.request.Request(VLLM + "/tokenize",
            data=json.dumps({"model": CFG.model, "prompt": prompt_text, "add_special_tokens": False}).encode(),
            headers={"Content-Type": "application/json"})
        with urllib.request.urlopen(req, timeout=30) as r:
            all_toks.append(json.loads(r.read())["tokens"])
    mx = max(len(t) for t in all_toks)
    shared = shim_mod.common_prefix_len(all_toks)
    return mx, shared + sum(len(t) - shared for t in all_toks)
 def admit(body):
    try:
        mx, total = sent_tokens(body)
    except Refused:
        raise
    except Exception as e:  # noqa: BLE001
        raise Refused(502, {"error": f"admission: tokenizer count failed: {type(e).__name__}: {e}"})
    if mx > CAP or total > CAP:
        raise Refused(422, {"error": f"admission: {mx} max-prompt / {total} total tokens against the {CAP}-token seat "
                                     f"limit; refused, not trimmed", "max_prompt_tokens": mx,
                            "shim_accounted_total": total, "cap": CAP})
    return mx, total
 def post_up(path, obj):
    req = urllib.request.Request(UP + path, data=json.dumps(obj).encode(), headers={"Content-Type": "application/json"},
                                 method="POST")
    try:
        with urllib.request.urlopen(req, timeout=300) as r:
            return r.status, json.loads(r.read())
    except urllib.error.HTTPError as e:
        raw = e.read() or b"{}"
        try:
            return e.code, json.loads(raw)
        except ValueError:
            return e.code, {"error": raw[:300].decode(errors="replace")}
 def systemone(body):
    mx, total = admit(body)
    code, obj = post_up("/v1/systemone", body)
    if isinstance(obj, dict):
        obj.setdefault("_admission", {}).update({"max_prompt_tokens": mx, "shim_accounted_total": total, "cap": CAP})
    return code, obj
 def check_decision(d, i):
    if not isinstance(d, dict) or not isinstance(d.get("id"), str) or not isinstance(d.get("question"), str):
        raise Refused(422, {"error": f"decision {i}: needs string id and question"})
    opts = d.get("options")
    if not isinstance(opts, list) or not MIN_OPTIONS <= len(opts) <= MAX_OPTIONS:
        raise Refused(422, {"error": f"decision {d['id']!r}: options must contain {MIN_OPTIONS}-{MAX_OPTIONS} entries"})
    if any(not isinstance(o, dict) or not isinstance(o.get("id"), str) or not isinstance(o.get("description"), str)
           for o in opts):
        raise Refused(422, {"error": f"decision {d['id']!r}: every option needs string id and description"})
    if len({o["id"] for o in opts}) != len(opts):
        raise Refused(422, {"error": f"decision {d['id']!r}: option ids must be unique"})
    if d.get("orderings", "none") != "none":
        raise Refused(422, {"error": f"decision {d['id']!r}: orderings {d.get('orderings')!r} is not served by this "
                                     "seat (h2o); send orderings 'none' and combine on the client"})
 def decide(state, decisions):
    """semif /decide/shared on top of /v1/systemone: intern-decision-serve's mapping (one Jev choice question per
    decision, calls of at most 16, fields q / q1..qN), re-keyed into its result shape."""
    if not isinstance(decisions, list) or not 1 <= len(decisions) <= MAX_DECISIONS:
        raise Refused(422, {"error": f"decisions must be a list of 1..{MAX_DECISIONS}"})
    for i, d in enumerate(decisions):
        check_decision(d, i)
    started, results, sizes, tokens = time.perf_counter(), [], [], []
    for k in range(0, len(decisions), MAX_QUESTIONS_PER_CALL):
        chunk = decisions[k:k + MAX_QUESTIONS_PER_CALL]
        fields = ["q"] if len(chunk) == 1 else [f"q{j + 1}" for j in range(len(chunk))]
        questions = {f: {"type": "choice", "instructions": d["question"],
                         "criteria": {o["id"]: o["description"] for o in d["options"]}} for f, d in zip(fields, chunk)}
        code, resp = systemone({"state": state, "questions": questions})
        if code != 200:
            raise Refused(code, resp if isinstance(resp, dict) else {"error": str(resp)})
        sizes.append(len(chunk))
        tokens.append((resp.get("usage") or {}).get("input_tokens"))
        for f, d in zip(fields, chunk):
            answer = (resp.get("answers") or {}).get(f)
            ids = [o["id"] for o in d["options"]]
            probs = (answer or {}).get("probabilities") or {}
            if answer is None or any(i not in probs for i in ids):
                raise Refused(500, {"error": f"scoring_failed: the model's answer for field {f!r} lacks option ids"})
            results.append({"id": d["id"], "option_ids": ids, "probabilities": [probs[i] for i in ids],
                            "top": answer.get("choice"), "confidence": answer.get("confidence"),
                            "calibration": resp.get("calibration"), "native": answer,
                            "input_tokens": tokens[-1], "model": MODEL,
                            "call": {"index": len(sizes) - 1, "field": f, "questions": len(chunk)}})
    return {"results": results, "timing": {"total_seconds": time.perf_counter() - started, "calls": len(sizes),
                                           "questions_per_call": sizes, "input_tokens": tokens}}
 class H(BaseHTTPRequestHandler):
    protocol_version = "HTTP/1.1"
    def log_message(self, fmt, *args):
        sys.stderr.write("%s - %s\n" % (self.address_string(), fmt % args))
    def _send(self, code, obj):
        b = json.dumps(obj).encode()
        self.send_response(code)
        self.send_header("Content-Type", "application/json")
        self.send_header("Content-Length", str(len(b)))
        self.end_headers()
        self.wfile.write(b)
    def _authorized(self):
        scheme, _, tok = (self.headers.get("Authorization") or "").partition(" ")
        return scheme.lower() == "bearer" and hmac.compare_digest(tok.strip().encode(), TOKEN.encode())
    def do_GET(self):
        path = self.path.split("?")[0]
        if path == "/health":
            try:
                with urllib.request.urlopen(UP + "/health", timeout=10) as r:
                    shim = json.loads(r.read())
            except Exception as e:  # noqa: BLE001
                return self._send(503, {"status": "unavailable", "model": MODEL, "error": f"shim: {type(e).__name__}"})
            ok = bool(shim.get("ok"))
            return self._send(200 if ok else 503, {
                "status": "ok" if ok else "starting", "model": MODEL, "cap_tokens": CAP, "admission": "refuse, never trim",
                "endpoints": ["/decide", "/decide/shared", "/v1/systemone"], "max_decisions": MAX_DECISIONS,
                "max_questions_per_call": MAX_QUESTIONS_PER_CALL, "orderings": ["none"], "backend": shim})
        if not self._authorized():
            return self._send(401, {"error": "unauthorized"})
        if path == "/v1/models":
            try:
                with urllib.request.urlopen(UP + "/v1/models", timeout=10) as r:
                    return self._send(r.status, json.loads(r.read()))
            except urllib.error.HTTPError as e:
                return self._send(e.code, {"error": "upstream"})
        return self._send(404, {"error": "not found"})
    def do_POST(self):
        path = self.path.split("?")[0]
        if not self._authorized():
            return self._send(401, {"error": "unauthorized"})
        n = int(self.headers.get("Content-Length", 0) or 0)
        if n > 4 * 1024 * 1024:
            return self._send(413, {"error": "body over 4 MiB"})
        try:
            body = json.loads(self.rfile.read(n) or b"null")
        except ValueError:
            return self._send(422, {"error": "body is not JSON"})
        if not isinstance(body, dict):
            return self._send(422, {"error": "body must be a JSON object"})
        try:
            if path == "/v1/systemone":
                return self._send(*systemone(body))
            if path == "/api/alpha/decisions":
                admit(body)
                return self._send(*post_up(path, body))
            if path == "/decide/shared":
                return self._send(200, decide(body.get("state"), body.get("decisions")))
            if path == "/decide":
                d = {k: body.get(k) for k in ("id", "question", "options", "orderings") if k in body}
                out = decide(body.get("state"), [d])
                return self._send(200, {**out["results"][0], "total_seconds": out["timing"]["total_seconds"]})
        except Refused as r:
            return self._send(r.code, r.obj)
        return self._send(404, {"error": "not found"})
 ThreadingHTTPServer((BIND, PORT), H).serve_forever()
@@ -0,0 +1,13 @@
 #!/bin/bash
 # One container: the unmodified vendor shim on loopback, then the front. If either dies, the container exits and
 # Docker restarts it (restart: unless-stopped). Both inherit the container's whole environment, so the shim's own
 # settings (SHIM_*, e.g. SHIM_NOUL_FLOOR, and LIGHTNING_*, e.g. LIGHTNING_IMAGE_URLS=off) set in compose reach it.
 set -eu
 python3 /model/h2o_lightning_shim.py --config /model/serve_config.json --vllm "$SHIM_VLLM" \
  --host 127.0.0.1 --port "$SHIM_PORT" &
 SHIM=$!
 python3 /front/front.py &
 FRONT=$!
 wait -n "$SHIM" "$FRONT" 2>/dev/null || true
 kill "$SHIM" "$FRONT" 2>/dev/null || true
 exit 1
@@ -0,0 +1,12 @@
 #!/bin/bash
 # One-command rollback of the decision seat to Intern-Decision-4B (the pre-2026-10-12 seat), run ON fv-ml1:
 #   sudo bash /opt/docker/compose/intern-decision/rollback-to-inc.sh
 # It stops the h2o stack, puts the INC compose back (same .env) and starts it (~35 s to healthy).
 # Used 2026-10-12 1403. The h2o compose is kept as compose.h2o.yaml (canonical in eshpfi); switch-to-h2o.sh goes back.
 set -euo pipefail
 cd /opt/docker/compose/intern-decision
 docker compose down
 cp compose.inc.yaml compose.yaml
 docker compose up -d
 for i in $(seq 1 60); do curl -sf http://127.0.0.1:8033/health >/dev/null && { echo "INC healthy"; exit 0; }; sleep 3; done
 echo "INC not healthy after 180 s"; exit 1
@@ -0,0 +1,18 @@
 #!/bin/bash
 # Switch the decision seat to h2o-lightning-4b (the staged compose.h2o.yaml), run ON fv-ml1:
 #   sudo bash /opt/docker/compose/intern-decision/switch-to-h2o.sh
 # Only once albok-dev has re-fitted Nemi's judge for h2o (2026-10-12 rollback; see the stack README).
 # It stops INC, checks GPU 1 has room for h2o's measured peak, and either starts h2o or puts INC back.
 # h2o takes ~3 min to healthy (graph capture). Undo with rollback-to-inc.sh.
 set -euo pipefail
 cd /opt/docker/compose/intern-decision
 NEED_MIB=16900   # h2o seat budget 16,847 MiB (peak 16,164 under 4 x 16q x 32k on 2026-10-12)
 docker compose down
 FREE=$(nvidia-smi -i 1 --query-gpu=memory.free --format=csv,noheader,nounits | tr -d ' ')
 if [ "$FREE" -lt "$NEED_MIB" ]; then
  echo "GPU 1 free ${FREE} MiB < ${NEED_MIB}: not starting h2o; restarting INC"; docker compose up -d; exit 2
 fi
 cp compose.h2o.yaml compose.yaml
 docker compose up -d
 for i in $(seq 1 100); do curl -sf http://127.0.0.1:8033/health >/dev/null && { echo "h2o healthy"; exit 0; }; sleep 3; done
 echo "h2o not healthy after 300 s: run rollback-to-inc.sh"; exit 1
@@ -0,0 +1,24 @@
 # h2o admission proxy + gate probe set
 Production-relevant because the seat swap to h2o-lightning-4b (ordered
 2026-10-12) inherits a hard constraint: the shipped vendor shim SILENTLY
 trims state at 32,000 tokens. The seat contract is a REFUSING 32,000-token
 seat, and this proxy is what makes the refusal real. It must sit in front of
 the shim on any deployment.
 - `admission_proxy.py` — thin wrapper in front of the unmodified shipped
  `h2o_lightning_shim.py`. Imports the vendor's own renderers, counts each
  question's exact prompt with vLLM's `/tokenize`, and refuses over-cap with
  422 + the named count (both currencies: per-question prompt max and the
  shim's shared-head+tails total). Env: ADMIT_CAP (default 32000), ADMIT_UP
  (shim URL), ADMIT_VLLM, ADMIT_PORT, ADMIT_SHIM (path to the vendor file).
  Measured gate overhead: the shape numbers pass with it in-line.
 - `gate_probe_set.py` — the Amendment 1 confirmation driver (calibrate to the
  cap, depth-gate cells, cold/warm shapes, 4x concurrency, over-cap refusal).
  Written for the bench layout on fv-ml1 (`/tank/bench`); paths are hardcoded,
  adapt before re-running anywhere else.
 Provenance: fv-ml1:/tank/bench/h2o_admission.py @ the confirmation run,
 results in docs/pfi/decision-sweep-followon-bench-2026-10-12.md (Amendment 1
 addendum). Measured boot for capacity planning: ~154 s launch-to-ready,
 graph-capture bound (no warm-cache or eager-mode saving).
@@ -0,0 +1,92 @@
 #!/usr/bin/env python3
 """Thin admission wrapper for the h2o confirmation run (Amendment 1, brokkr 63d7a1b).
 Sits IN FRONT of the unmodified shipped h2o_lightning_shim.py and refuses over-cap
 input loudly, mirroring INC's refusal. Counts with h2o's OWN tokenizer: renders the
 state exactly as the shim does (imported verbatim from the shipped file) and asks
 vLLM's /tokenize for the count. Refuses > CAP tokens with an explicit 422; forwards
 everything else untouched. No trim. The shim is not modified.
 """
 import importlib.util, json, os, sys, time, urllib.request
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 CAP = int(os.environ.get("ADMIT_CAP", "32000"))
 UP = os.environ.get("ADMIT_UP", "http://127.0.0.1:18091")   # the shim
 VLLM = os.environ.get("ADMIT_VLLM", "http://127.0.0.1:18099")
 PORT = int(os.environ.get("ADMIT_PORT", "18092"))
 SHIM_PATH = os.environ.get("ADMIT_SHIM", "/tank/bench/pkg-h2o/h2o_lightning_shim.py")
 spec = importlib.util.spec_from_file_location("h2o_shim", SHIM_PATH)
 shim_mod = importlib.util.module_from_spec(spec)
 sys.modules["h2o_shim"] = shim_mod
 spec.loader.exec_module(shim_mod)
 CFG = shim_mod.Contract(json.load(open("/tank/bench/pkg-h2o/serve_config.json")))
 CFG.max_state_tokens = 10**9   # wrapper must count the UNTRUNCATED render; the shim's own trim is what we police
 def sent_tokens(body):
    """Tokens per prompt as vLLM will actually receive them, counted by vLLM's own tokenizer via
    the shim's own renderers (chat path: tokenize over messages, matching the image-free prompt
    the shim sends per question). Returns (max_per_prompt, shim_accounted_total).
    The shim accounts usage as shared_prefix + sum(question tails); we mirror that exactly."""
    import concurrent.futures
    state = body.get("state") or ""
    if isinstance(state, dict):
        state = json.dumps(state)
    state_txt = shim_mod.render_state(state, CFG.state_compact)   # UNTRUNCATED: what admission guards
    all_toks = []
    for qid, q in body.get("questions", {}).items():
        prompt_text = CFG.question_prompt(state_txt, q, qid)[0]   # EXACTLY what the shim sends per question
        req = urllib.request.Request(VLLM + "/tokenize",
            data=json.dumps({"model": CFG.model, "prompt": prompt_text, "add_special_tokens": False}).encode(),
            headers={"Content-Type": "application/json"})
        with urllib.request.urlopen(req, timeout=30) as r:
            all_toks.append(json.loads(r.read())["tokens"])
    mx = max(len(t) for t in all_toks)
    shared = shim_mod.common_prefix_len(all_toks)
    total = shared + sum(len(t) - shared for t in all_toks)   # the shim's own accounting: head + tails
    return mx, total
 class H(BaseHTTPRequestHandler):
    protocol_version = "HTTP/1.1"
    def log_message(self, *a): pass
    def _send(self, code, obj):
        b = json.dumps(obj).encode()
        self.send_response(code); self.send_header("Content-Type", "application/json")
        self.send_header("Content-Length", str(len(b))); self.end_headers(); self.wfile.write(b)
    def do_GET(self):
        try:
            req = urllib.request.Request(UP + self.path)
            with urllib.request.urlopen(req, timeout=30) as r:
                self._send(r.status, json.loads(r.read()))
        except urllib.error.HTTPError as e:
            self._send(e.code, json.loads(e.read() or b"{}"))
    def do_POST(self):
        n = int(self.headers.get("Content-Length", 0))
        raw = self.rfile.read(n)
        mx = total = None
        if self.path.split("?")[0] in ("/v1/systemone", "/api/alpha/decisions"):
            try:
                body = json.loads(raw)
            except ValueError:
                return self._send(422, {"error": "admission: body is not JSON"})
            try:
                mx, total = sent_tokens(body)
            except Exception as e:
                return self._send(502, {"error": f"admission: tokenizer count failed: {e}"})
            if mx > CAP or total > CAP:
                return self._send(422, {"error": f"admission: {mx} max-prompt / {total} total tokens against the "
                                                f"{CAP}-token seat limit; refused, not trimmed",
                                        "max_prompt_tokens": mx, "shim_accounted_total": total, "cap": CAP})
        req = urllib.request.Request(UP + self.path, data=raw,
            headers={"Content-Type": "application/json"}, method="POST")
        try:
            with urllib.request.urlopen(req, timeout=300) as r:
                obj = json.loads(r.read())
                obj.setdefault("_admission", {})["max_prompt_tokens"] = mx if "mx" in dir() else None
                obj["_admission"]["shim_accounted_total"] = total if "total" in dir() else None
                self._send(r.status, obj)
        except urllib.error.HTTPError as e:
            self._send(e.code, json.loads(e.read() or b"{}"))
 ThreadingHTTPServer(("127.0.0.1", PORT), H).serve_forever()
@@ -0,0 +1,150 @@
 #!/usr/bin/env python3
 # Amendment 1 confirmation driver for h2o (brokkr 63d7a1b). Stack assumed up:
 #   vLLM :18099, shipped shim :18091, admission wrapper :18092.
 # Shape sizes are pinned by the WRAPPER's own count (max per-prompt prompt tokens, pre-trim):
 # the shim's usage.input_tokens reports post-trim and is not trustworthy at the cap.
 # The gate's filler budget was written for INC's ~3.6 chars/token; h2o's tokenizer runs ~5.1,
 # so every builder target is iteratively rescaled against the wrapper count.
 import json, os, sys, time, statistics, subprocess, threading, random, urllib.request, concurrent.futures
 sys.path.insert(0, "/tank/bench")
 from gate_longctx_s1 import build_state
 CAP = 32000
 MODEL = "h2oai/h2o-lightning-4b"
 URL = "http://127.0.0.1:18092"   # through the admission wrapper
 OUT = "/tank/bench/arm-h2o"
 TAG = os.environ.get("CONFIRM_TAG","A")
 rng = random.Random("h2o-confirm-2026-10-12")
 def post(body, timeout=300):
    req = urllib.request.Request(URL+"/v1/systemone", data=json.dumps(body).encode(),
                                 headers={"Content-Type":"application/json"})
    with urllib.request.urlopen(req, timeout=timeout) as r:
        return json.loads(r.read())
 def topper(r, qid):
    a = (r.get("answers") or {}).get(qid) or {}
    p = a.get("probabilities") or {}
    return max(p, key=lambda k: float(p[k])) if p else None
 def questions(nq):
    return {f"q{i}": {"type":"choice","instructions":f"Does the state satisfy criterion {i}?",
                      "criteria":{"yes":f"topic-{i} mentioned","no":"not mentioned"}} for i in range(nq)}
 def adm(body):
    r = post(body)
    return r, (r.get("_admission") or {})
 EVID=("DECISIVE: the underwriter's signed memo of 2026-03-14 states the claim was established "
      "by documents filed before the deadline; the renewal decision must rest on this memo.")
 GQ={"type":"choice","instructions":"Based only on the state above, does the evidence establish the claim, "
    "contradict it, or is it insufficient?",
    "criteria":{"supported":"The evidence establishes the claim","contradicted":"The evidence establishes the opposite",
                "insufficient":"The evidence does not establish either"}}
 def count_gate(st):
    # wrapper counts then forwards; over-cap it REFUSES with the count in the body — either way we learn the count
    try:
        r,a = adm({"model":MODEL,"state":st,"questions":{"q1":GQ}})
        return a.get("max_prompt_tokens")
    except urllib.error.HTTPError as e:
        if e.code == 422:
            try: return json.loads(e.read()).get("max_prompt_tokens")
            except Exception: return None
        raise
 peak=0; stop=False
 def poll():
    global peak, stop
    while not stop:
        try: peak=max(peak,int(subprocess.run(["nvidia-smi","-i","3","--query-gpu=memory.used",
             "--format=csv,noheader,nounits"],capture_output=True,text=True).stdout.split()[0]))
        except Exception: pass
        time.sleep(0.05)
 res={"tag":TAG}
 # ---- largest-admissible state: iterate until wrapper max-prompt is within 160 tok under CAP
 STATE_FILE = os.environ.get("CONFIRM_STATE","")
 if os.environ.get("CONFIRM_LOAD") == "1" and STATE_FILE and os.path.exists(STATE_FILE):
    state32 = open(STATE_FILE).read()
    hist=[{"note":"state loaded from file; first engine request is the first shape cell"}]
    target=None
 else:
  target=31900; hist=[]
  for it in range(8):
    st = build_state(target, None, 0, rng)
    try:
        r, a = adm({"model":MODEL,"state":st,"questions":questions(16)})   # pin the TOTAL at 16q, per clause 3
        mx = a.get("max_prompt_tokens"); tot = a.get("shim_accounted_total")
    except urllib.error.HTTPError as e:
        if e.code != 422: raise
        b = json.loads(e.read()); mx = b.get("max_prompt_tokens"); tot = b.get("shim_accounted_total")
    hist.append({"iter":it,"builder_tok_target":int(target),"wrapper_max_prompt":mx,"wrapper_total":tot})
    if tot is None: raise SystemExit(str(mx))
    if tot <= CAP and (CAP-tot) <= 160: break
    target = int(target * max(0.5, min(1.3, (CAP-80)/max(tot,1))))
 if target is not None: state32 = st
 res["calibration"]=hist
 def shape(nq, cold_expected=False):
    body={"model":MODEL,"state":state32,"questions":questions(nq)}
    t=time.perf_counter(); r=post(body); first=(time.perf_counter()-t)*1000
    xs=[first]
    for _ in range(4):
        t=time.perf_counter(); post(body); xs.append((time.perf_counter()-t)*1000)
    r2, a2 = adm(body)
    return {"nq":nq,"first_ms":round(first),"warm_median_ms":round(statistics.median(xs)),
            "cold_expected":cold_expected,
            "max_prompt_tokens":a2.get("max_prompt_tokens"),"shim_total":a2.get("shim_accounted_total"),
            "usage_input_tokens":(r2.get("usage") or {}).get("input_tokens")}
 STATE_FILE=os.environ.get("CONFIRM_STATE","")
 if STATE_FILE: open(STATE_FILE,"w").write(state32)
 cold_first=int(os.environ.get("COLD_FIRST","0"))   # which nq gets the truly-cold first call (0 = none)
 th=threading.Thread(target=poll); th.start()
 res["cells"]={}
 order=([cold_first] if cold_first in (1,16) else [1,16])
 order=order+[n for n in (1,16) if n not in order]
 for nq in order:
    res["cells"][f"{nq}q32k"]=shape(nq, cold_expected=(nq==cold_first))
 # ---- concurrency: 4x concurrent 16q at the same state
 body={"model":MODEL,"state":state32,"questions":questions(16)}
 t=time.perf_counter()
 with concurrent.futures.ThreadPoolExecutor(4) as ex:
    futs=[ex.submit(post,body) for _ in range(4)]
    oks=[]
    for f in futs:
        try: oks.append(topper(f.result(),"q0") is not None)
        except Exception: oks.append(False)
 res["concurrency_4x16q"]={"wall_ms":round((time.perf_counter()-t)*1000),"all_ok":all(oks),"per_call_ok":oks}
 # ---- gate cells: evidence at absolute h2o-token depths ~1k/8k/16k/30k, whole state pinned ~31,900
 gate=[]
 for depth_tok in (1000,8000,16000,30000):
    tgt=31900
    g=tgt*1.42; hist2=[]
    for it in range(8):
        d_scaled = int(depth_tok * (g/tgt))   # relative depth scales with the char budget
        stg = build_state(int(g/3.6), EVID, d_scaled, rng)
        mx = count_gate(stg)
        hist2.append({"chars_target":int(g),"count":mx})
        if mx is None: break
        if abs(mx-tgt)<=200: break
        g = int(g * tgt/mx)
    try:
        r,a = adm({"model":MODEL,"state":stg,"questions":{"q1":GQ}})
        cell={"depth_tok_target":depth_tok,"max_prompt_tokens":a.get("max_prompt_tokens"),
              "top":topper(r,"q1"),"usage_input_tokens":(r.get("usage") or {}).get("input_tokens"),
              "refused":False,"cal":hist2[-1]}
    except urllib.error.HTTPError as e:
        cell={"depth_tok_target":depth_tok,"refused":True,"detail":json.loads(e.read()),"cal":hist2[-1]}
    gate.append(cell)
    time.sleep(1)
 res["gate"]=gate
 stop=True; th.join()
 res["peak_MiB"]=peak
 # ---- over-cap REFUSAL (the whole point): take state32 and grow it past the cap
 st_over = state32 + ("Policy renewal review cycle X: coverage schedules, endorsements and rider terms are "
                     "reconciled against the master binder. " * 400)
 refused=False; detail=None; mx_over=count_gate(st_over)
 try:
    r_ok = post({"model":MODEL,"state":st_over,"questions":{"q1":GQ}})
    detail={"note":"not refused","max_prompt_tokens":(r_ok.get("_admission") or {}).get("max_prompt_tokens")}
 except urllib.error.HTTPError as e:
    refused=True; detail={"status":e.code,"body":json.loads(e.read())}
 res["over_cap_refusal"]={"refused":refused,"counted_tokens":mx_over,"detail":detail}
 json.dump(res,open(f"{OUT}/confirm-{TAG}.json","w"),indent=1)
 print("CONFIRM-DONE", json.dumps(res)[:2400])
Author	SHA1	Message	Date
vh	6a01da759d	memory: decision seat live on h2o again since 1704 under the re-fitted judge	2026-10-12 17:06:20 -07:00
vh	98254ca077	feat(intern-decision): decision seat back on h2o-lightning-4b under the re-fitted judge Prime ruled the cutover ("do it", 16:40, via albok-dev) after the judge's note question was re-fitted (albok-nemi 0.1.18). switch-to-h2o.sh ran 17:02:23, the GPU 1 gate passed, h2o healthy 17:04:54. Checks on :8033: /health at the pin, /decide/shared 200, 422 over the cap, inline image answers, image URL refused by the shim, vLLM media-domain backstop refuses (400). compose.yaml is h2o again; rollback-to-inc.sh stays the one-step way back.	2026-10-12 17:05:59 -07:00
vh	014a4fd0ec	docs(intern-decision): seat switch checklist — Nemi's judge concurrency is 1 on INC, 4 on h2o	2026-10-12 16:21:25 -07:00
vh	8d9fcd9ffb	memory: h2o judge investigation closed — v1 note question was the cause, re-fitted in albok-nemi 0.1.18; eval seat down	2026-10-12 16:20:49 -07:00
vh	2b40805b84	feat(h2o-eval): round 2 eval seat on GPU 3 with floor-on/floor-off fronts for the judge investigation Prime (via albok-dev): understand why h2o failed and re-fit the judge. One vLLM on GPU 3 behind two fronts: :8034 with the shim's yes/no floor at 0.801 exactly as the cutover ran (same judge stamp), :8035 with SHIM_NOUL_FLOOR=0 and a distinct stamp (h2o-lightning-4b+noul-floor-0). Both refuse image URLs; the vLLM --allowed-media-domains backstop is now verified (400, no fetch).	2026-10-12 14:29:15 -07:00
vh	501bda078d	memory: decision-seat rollback closed out — hidden set back to INC counts	2026-10-12 14:09:11 -07:00
vh	f979cca1ea	fix(intern-decision): roll the decision seat back to Intern-Decision-4B; h2o staged with inline-images-only albok-dev withdrew the judge sign-off on fleet-scale evidence and brokkr concurred: over 10,841 texts judged by both models, h2o flagged 242 adult at 0.5 against INC's 88, and no threshold reproduces INC's set. rollback-to-inc.sh ran 14:03:28, INC healthy 14:03:56, /decide/shared smoke 200 x3. compose.yaml is INC again; the h2o compose is kept as compose.h2o.yaml with switch-to-h2o.sh (GPU 1 free-memory gate, falls back to INC). It now carries Prime's inline-images-only setting: LIGHTNING_IMAGE_URLS=off on the front (verified live: an https image URL is refused before any fetch, inline images still answer) and vLLM --allowed-media-domains media.invalid as a backstop. rollback-to-inc.sh no longer leaves timestamped compose copies behind.	2026-10-12 14:05:44 -07:00
vh	c08992f95f	feat(intern-decision): decision seat swapped to h2o-lightning-4b on fv-ml1 GPU 1; eval seat down The :8033 seat now serves h2oai/h2o-lightning-4b @ b43f759b on vLLM 0.30.0 with byte-pinned KV (4.48 GiB, 135,558 tokens) behind services/h2o-decision-front: 32k refuse-not-trim admission, the seat's bearer token, and /decide + /decide/shared translated onto /v1/systemone for albok-nemi's judge. URL, port, token and stack name are unchanged. INC down 13:08:18, h2o healthy 13:11:27. Post-checks on the live seat: /health at the pinned revision; JevBench public 205/231 (= bench); 422 over the cap; 16q at the cap warm 2.0-2.16 s; seat peak 16,164 MiB under 4 concurrent 16q x 32k calls against a 16,847 budget. compose.inc.yaml + rollback-to-inc.sh restore Intern-Decision-4B in one step. The temporary GPU 3 eval seat (stacks/h2o-eval) is down.	2026-10-12 13:19:46 -07:00
vh	152137e41d	memory: decision-seat swap held for albok's judge; h2o eval seat on GPU 3	2026-10-12 13:02:04 -07:00
vh	2c3ef7a8af	feat(h2o): temporary h2o-lightning-4b eval seat on fv-ml1 GPU 3 + the seat front (admission, bearer, /decide/shared translator) Prime's call via brokkr: h2o-lightning-4b replaces the intern-decision seat once albok-dev re-validates Nemi's sensitivity judge on it. services/h2o-decision-front builds on infra-hermes's admission proxy (`30f36f2`): 32,000-token refuse-not-trim, the seat's bearer token, and intern-decision-serve's /decide/shared mapping onto /v1/systemone so the judge runs unchanged. stacks/h2o-eval serves it on GPU 3 :8034 exactly as benched (vLLM 0.30.0 + unmodified vendor shim; weights hashed against HF b43f759b). INC keeps serving production on GPU 1.	2026-10-12 13:01:53 -07:00
vh	30f36f2478	tools(h2o): admission proxy + gate probe set — the seat swap inherits a hard dep The vendor shim silently trims state at 32,000 tokens; the admission proxy is what turns the seat contract into a loud refusal, so it must front the shim in any deployment. Committing it with the Amendment 1 confirmation driver so executing the swap does not depend on a bench scratch dir surviving.	2026-10-12 12:29:57 -07:00
vh	14af41b945	docs(pfi): Amendment 1 addendum — h2o confirmation run passes all three clauses Admission via thin wrapper (vendor shim unmodified; no shim config knob exists). Gate re-run clean at every depth with usage honest; 16q32k warm 2,033 ms; 4x concurrent 16q32k peak 15,566 MiB vs 16,847 budget; boot matrix replaces the unmeasured 280 s figure (cold median 153.6 s, warm-cache saves nothing, eager slower). Wyrd itemised: 8 diffs, h2o net -2. h2o passes every Supersede clause under the accepted 32,000-token refusing contract; seat swap is the operator's call.	2026-10-12 12:13:31 -07:00
vh	4098a23dda	docs(pfi): decision-sweep follow-on bench — H2O re-bench, Clef-flash, RSI-Jev vs the live seat First bench-first run (prereg brokkr-smithy 2e0539f). H2O-lightning-4b is a short-route candidate: public 205, battery CI lower bound +0.0023, fits the GPU 1 budget at every shape — but its shipped shim silently trims state at ~32k and fails the long-input gate by the letter. Clef-flash lateral in both precisions (Wyrd -0.0333 in both; bf16 does not fit at rest). RSI-Jev v6.1 loses on decision quality (public -23, Wyrd -0.0548). INC holds the seat. Torn down; weights kept (model-runtime 116 GB).	2026-10-12 10:22:03 -07:00
vh	0fe081042d	memory: U13 seat-host 403 regression fixed	2026-10-12 09:44:12 -07:00
vh	82987828f8	fix(svos-front): kernel seat-host calls allowed on the terminal origin (app-terminal base_url reverted to it)	2026-10-12 09:44:06 -07:00
vh	ca9e385316	memory: CI runner slowness diagnosed (ospool fsync ~145 ms on ana-docker); A/B/C pending Prime	2026-10-12 09:41:23 -07:00
vh	a9bae04ba6	docs(pfi): D3 vs Intern-Decision-4B seat bench — final stashed findings (halted by operator call) Measured: bf16 ladder across six sizes, calibrated 32k shape cells with O(q*state) wire accounting, bs=1 memory/latency curve, Engine-vs-vendor fidelity rows, Stage 4 control quant trio (battery monotone in weight error), flash calibrated-NVFP4 with survive rules applied by the letter, and the Amendment 2 GPU 1 budget (16,851 MiB) re-read against every fit clause. Halted before the INC bench-copy arms and mini/27B/small-size NVFP4 arms; stated plainly. Weights kept pending operator ruling (80 GB).	2026-10-12 08:08:39 -07:00