Four attribution failures shipped on the day it was built, and every one of
them RECORDED SOMETHING -- just not something findable or attributable. The
individual fixes are in the commit log; the PATTERN was not written down
anywhere, so a future session would meet each one as an isolated bug rather
than as the fifth instance of a known shape.
infra-hermes asked for this after the fourth one: 'worth an audit line item so
the pattern is visible, not just the fixes.' He was right.
The operative warning is at the bottom: absence of an entry is not evidence of
absence of the event until you have checked this tool could have recorded it.
infra-ops and infra-hermes act as the same OS identity and dockerd does not
log exec per caller, so host-side changes carry no fingerprint. Git cannot
close the gap either: every commit here is attributed to Vuong Hoang by
convention, which is correct for authorship and useless for attribution.
On 2026-09-18 a second session edited the searxng stack mid-deploy, crash-
looping fleet search for ~4 minutes, and the author was unidentifiable.
scripts/ops-log records one line per host-changing action and holds a
lightweight claim so two agents do not deploy the same stack at once.
Four design questions, settled:
* Central on nh3-dev, not per-host and not the post office. Both agents
run as the same unix user there, so one file is shared with zero
provisioning. Per-host needs a writable path on ~25 heterogeneous boxes
and stores "we changed host Y" on host Y. journald looked free but shows
an unprivileged reader only their own _UID, which would have split the
log silently between the infra-ops and lkraven halves of the fleet.
* The claim is advisory and enforced in the tooling. deploy-stack.sh
refuses a foreign claim across the diff, the prompt and the apply -- the
whole review window, which is where the collision happened. Acquire is
mkdir, so it is atomic rather than probably-fine. Stale claims auto-break
and the break is recorded.
* Writers are automatic. deploy-stack.sh and elway record themselves; a log
that depends on remembering is the same class of instrument as a health
check that passes in both states.
* There is a detector. `ops-log audit` asks each host what changed on disk
and compares it to the newest log line for that stack, covering the
manual ssh-and-edit path the automatic writers structurally cannot.
ops-log being absent or broken never blocks a deploy; only a live foreign
claim does. `ops-log baseline` marks the 136 stacks that predate the
instrument so the detector starts from today rather than reporting the whole
fleet forever and training us to ignore it.
An unreachable host reports INCOMPLETE and exit 5, never clean.
Every NH3→Anaheim flow had been crossing Tailscale's LA DERP relay rather than
a direct path, for long enough to have carried 78 GB tx on the NH3 side alone.
DERP is a throttled fallback, so this imposed both a fixed round-trip penalty
and a bandwidth ceiling on LiteLLM, Beszel, task-board, vor and the Henge
alike. It surfaced as a voice-loop latency report from tts-dev, not as a
network alarm, because nothing monitors whether a mesh path is direct.
ana-scale advertised 38.120.12.42:41641 while the Anaheim NAT mapped it to
:60798 with no port-mapping protocol available, so inbound hole-punching always
failed. Four additive objects on ana-gw give it a stable inbound endpoint.
tailscale ping nh3-scale->ana-scale 373-522 ms via DERP -> 6 ms direct
STT via the ANA gateway, 96 kB clip 1.399-1.449 s -> 0.237-0.270 s
Beszel HTTP nh3-dev->ana-docker 0.94-1.29 s -> 0.014-0.016 s
Documents the house template that matters for this box: it runs central-nat,
so a policy dstaddr is the real internal address and not the VIP. Also records
that the pre-change config was captured with `show` to a local file rather
than a tftp job, since this edge has no out-of-band access and a backup is
mandatory before touching it.
irv-ml1 remains relayed and is called out as outstanding.
The FV edge firewall was not backed up anywhere. Its config now lands in
/var/lib/restic/stage/fv-gateway-config.xml via ana-docker's pre-backup hook,
so the existing 01:00 restic snapshot captures it. ana-docker is one of the
three egress addresses the firewall's WAN allowlist permits, which is why the
pull lives there rather than with the FV hardware — a site that has lost power
cannot back itself up, and FV lost power two days ago.
Non-fatal by design: an unreachable firewall must not abort the nightly
database dumps. But a bad pull must not be promoted either. The summary loop
only rejects EMPTY staged files, and this endpoint answers an auth failure
with a perfectly non-empty HTML error page — which would have been backed up
as a firewall config that is the right size and restores nothing. The block
checks the body really contains <opnsense> and writes nothing otherwise.
Three tests cover it, including the HTML-error-page case. The first draft of
those tests was worthless: _fv returned a Path out of a TemporaryDirectory
context, so the tree was deleted before the assertions ran and every
exists()-is-False check passed regardless of what the script did. Only the
positive test failed, which is the sole reason the broken negatives were
caught. They now snapshot inside the tempdir's lifetime, and the docstring
says why.
Also records two OPNsense API lessons in docs/pfi/opnsense-api-reference.md:
endpoints are actions and must never be probed for existence by POSTing at
them — that is how /api/core/system/reboot took the FV site dark for 3.5
minutes while looking for an apply call this same file already documented —
and the apply step is service/reconfigure, which auth/user notably lacks, so
an API-only key edit persists in config.xml and does nothing until the OS user
sync runs at boot.
Credentials in /etc/restic/fv-gateway.env (root:600), template committed,
values vaulted as fv-gateway/opnsense-api-{key,secret}. Pre-change config
snapshot vaulted as fv-gateway/config-backup-20260914.
Operator ruling: remove the groups and see what breaks. Nothing did.
ana-docker llmuser sudo+docker -> none; irv-ml1 llmuser sudo -> none (it was
never in docker there). 45 containers on ana-docker and 18 on irv-ml1 all
still running with zero unhealthy, and lora-training-worker stayed active.
Extended to irv-ml1 because it is the same account with the same defect and
gpasswd -a reverses it in one command; ana-docker was only the host the audit
happened to run against first.
The durable lesson is why it was safe, and it is a measurement trap rather
than a permissions one. reported 19 processes on ana-docker
and 3 on irv-ml1, which reads as a busy service account. Nearly all of them
were CONTAINER processes whose in-image UID is 1001 and therefore collides
with llmuser on the host — /proc/<pid>/cgroup shows docker-*.scope. A
container's runtime UID is unrelated to host group membership, so the groups
were buying those workloads nothing. The single real host workload sets
User=/Group= explicitly through systemd, which does not consult the sudo group
either.
Recorded in the conventions doc so the next audit checks the cgroup before
concluding a host account is busy — otherwise a UID collision blocks a
cleanup that carries no risk.
Operator ratified four conventions on 2026-09-14. docs/pfi/fleet-conventions.md
is the pin; playbooks/audit-host-conventions.yaml is its instrument.
Pinned, verified free on all eight surveyed hosts (dynamically-allocated
system accounts cluster in 989-999 and descend, so 800-899 is safe):
800-849 svc-* service accounts
850 infra-ops uid+gid
851 docker gid
852-899 reserved for fleet-wide groups
1000 the human account (vh)
Deliberately a pin for NEW hosts, not a migration mandate. The UID drift
(infra-ops is 1001/1002/1003/2001) is tolerable because there is no central
identity anywhere and a UID only has to agree where files cross hosts. They
do on /mnt/smithy — but that export is owned by Synology UIDs that resolve on
neither host and is 0777 throughout, so cross-host sharing works today
BECAUSE permissions are wide open. Aligning UIDs does not fix something
broken; it earns the right to drop that 777. Recorded as such rather than as
an urgent defect.
The audit playbook reports and never enforces, so a standard cannot quietly
become a flag day. Verified against nh3-dev, ana-docker, corviduo-dev and
nh3-extdev; it immediately surfaced two things the survey had missed —
llmuser holds sudo AND docker on ana-docker, and seven stacks on corviduo-dev
run from outside /opt/docker/compose (three under /home/vh, four under /opt,
including the three CI/CD-driven Worldtree deployments that must not be
moved).
Also supersedes the CLAUDE.md posture that made corviduo-dev the one host
excluded from fleet normalisation: the operator ruled all ops on it belong to
infra-ops. Its application layer stays CI/CD-owned.
Adds docs/pfi/llm-seat-catalog.md, the durable hand-curated record of what each
seat IS -- lineage, provenance, model-card facts, quantization, speculative
decoding, licenses, and measured warm tok/s + deep-prefill depth results with
their harness and date. It complements the auto-generated
fv-ml1-gpu-seat-inventory.md (live placement/KV/concurrency): the two split by
volatility, and the catalog defers to the inventory for any live number.
Adds scripts/seat-bench.py so the catalog's speed/depth numbers are reproducible
(warm decode tok/s, n=3, conc=1, median; deep prefill at ~0.97x max-model-len
with an allocator-log OOM scan). Serial by design -- concurrent deep prefills
would confound both OOM and tok/s.
Captures the 2026-09-14 measurements: all six generative seats prefill to ~255K
(coder ~16K) with zero OOM/CUBLAS/illegal-memory; warm decode 62.7-337.3 tok/s;
per-seat VRAM. seat-inventory.py now cross-links the catalog in its footer.
Full seat rebalance across GPU0/GPU1 (flash on GPU2 and empty GPU3 untouched),
operator-directed. Every target seat now serves native 262,144 context with
concurrency in the requested 1.2-2.5x band, verified from live boot logs:
cyberprev (sec) 262144 @ 1.37x depth-probed CLEAN to 259,722 tokens
flash-next (gen) 262144 @ 1.31x (untouched, already in band)
gen-small (NEW) 262144 @ 2.56x MTP k=3 measured 69.6% accept / 3.09 len
char-rp 262144 @ 1.22x (was 1.02x; util 0.40->0.52)
char-rp-fast 262144 @ 2.04x (util cap 0.30->0.24, pinned KV unchanged)
- gen-small: NEW seat, Qwen3.6-35B-A3B (3B active MoE), llmfan46 Heretic
(MPOA) NVFP4 experts-only, already on disk at qwen36-35b-a3b-heretic-nvfp4.
There is no general Qwen3.8 A3B (3.8 MoEs are Flash-Next and the 2.4T), so
this is the 3.6 fallback the operator specified. GPU0, :8026, MTP k=3, coherent
and MTP-verified before wiring. gen-small / gen-small-reasoning gateway aliases.
- coder: 8192 @ 13.75x -> 16384 @ 4.70x (util 0.06->0.055). Context doubled,
waste cut. Not the exact 2-3x target: the 1.5B weight+overhead floor (~4.2 GiB)
sits just under the util knob's resolution, so hitting <=3x reliably needs a
--kv-cache-memory byte pin (compose change) rather than the util fraction.
- cyberprev raised 163840 -> 262144: depth-probed with non-repeating prompts to
259,722 tokens, clean (no OOM, memory flat). Unlike mog-sec (same base arch,
capped at 163840 for depth crashes), this checkpoint holds native depth.
- Gateway (operator calls): summarizer + classifier -> gen-small; new
classifier-large -> gen-large (flash) for the accuracy tier; summarizer-large
stays on flash. All verified end-to-end.
- GPU1 hit its ceiling raising char-rp; resolved by trimming char-rp-fast's
reservation cap (its KV is pinned, so concurrency held at 2.04x) rather than
moving a utility seat -- the shared GPU_ID on reward/embed/rerank made a
single-seat move messier than the in-GPU rebalance.
Seat inventory regenerated from the live containers.
Live-read regeneration after promoting cyberprev into sec/sec-reasoning and
retiring mog-sec. GPU0 now carries cyberprev alone; sec/sec-reasoning aliases
resolve to :8025.
The first cyberprev quant served pure gibberish despite passing every structural
check. Root cause: the hotdogs/Qwen3.8-27B-abliterated-cyber-preview checkpoint
ships broken -- 850 of 1199 tensors (the whole transformer body) carry a TRIPLED
`language_model.` prefix, confirmed present in the upstream HF index (an unsloth
export bug, not our download). transformers silently loaded the unmapped tensors
as random-init, the pipeline quantized noise, and structural verification passed
because it inspects the output config, never that input names mapped. A plain
greedy coherence smoke caught it; nothing structural could.
Fix: rewrite tensor names against a known-good reference of the same architecture
(qwen38-27b-uncensored-bf16), verified the repaired name-set == reference name-set
exactly (1199 tensors, 0 collisions), and re-quantized from the corrected source
(cyberprev-bf16-fixed). v2 is coherence-verified (spec off AND live with dflash):
Paris/Berlin/Rome, correct arithmetic precedence, primes, clean cyber-domain chat
with a proper thinking split. dflash k=7 measured at 2.77 mean acceptance length.
Seat serves cyberprev-27b / cyberprev-27b-thinking on :8025/GPU0 with dflash k=7,
displacing the retired sentinel-r3. Gateway: removed the retired sentinel-r3 /
sentinel-r3-reasoning aliases and added cyberprev-27b / cyberprev-reasoning -- NOT
repointed, since serving cyberprev weights under sentinel's name is silent
substitution. Verified end-to-end through the gateway; sentinel-r3 now 400s.
Playbook gains a superseded-claims row: structural verification does not imply a
good quant; a coherence smoke is the load-bearing gate, and third-party sources
must have their tensor names diffed against a reference before quantizing.
Broken v1 quant dir removed.
Swaps gen-large from the dealignai ModelOpt W4A4 build to
orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4, which is weight-only on both
axes (W8 float attn, W4 float experts, input_activations: null) and so avoids
the 4-bit-activation long-context degradation mode.
The checkpoint was previously recorded as unloadable on any mainline vLLM,
requiring a from-source PLE-loader patch. That conclusion was wrong on cost.
Qwen4ExpPLEEmbeddingMethod.from_quant_config checks ple_embedding_dtype as
branch 1, before any quant-config type check, and its NotImplementedError for
CompressedTensorsConfig is scoped to the PLE path only -- experts and dense
load through the ordinary compressed-tensors paths. Verified by instantiating
the real config and calling the selector both ways before doing any work.
orcarouter ships a bf16 PLE, so the fix was to make the declaration true:
convert the 51.2B-param table to FP8 and declare it. Its 128 PLE tensors sit
in one shard file with nothing else in it. Global amax 0.0894, per-shard
outlier ratio 1.66x, scale chosen exactly representable in bf16 so no
scale-rounding error stacks on quantization; amax maps to 446.17/448, no
clipping. Round-trip 2.655% RMS relative, 0.002% underflow, 0 saturation --
the same FP8-PLE treatment dealignai already shipped. MTP head (31 tensors)
and vision tower carried through untouched.
A second, independent blocker followed: orcarouter labels its 12 QSA layers
qwen_sparse_attention, which vLLM rejects; it accepts full_attention and
selects QSA via indexer_n_heads. Confirmed indexer_n_heads == 4 in both this
and the dealignai checkpoint before renaming -- without that check the rename
silently selects plain attention and serves a subtly wrong model that still
passes a healthcheck.
Measured on the live seat: healthy, coherent, KV 344,155 tokens @ 262,144 ctx,
MTP k=3 at 60.4% acceptance / 2.81 mean acceptance length, warm decode median
167.5 tok/s at conc=1 (n=5, spread 12.2%). The reorg note's dealignai figure
came from a different harness, so this is not claimed as a win over it; what
it does establish is that weight-only experts did not cost decode speed.
Still open: controlled quality A/B vs dealignai, and a deep-prefill probe at
262K. Rollback is two .env keys; dealignai remains on disk.
Also corrects the README's MTP-is-off section, stale since k=3 was deployed,
and adds a superseded-claims row to the quantization playbook.
FV colo recovered 2026-09-13 midday (chassis on PDU, firewall on the Eaton 5P1000,
GPU caps 275W/card). All-night fv-ml1 seat reorganization:
- flash-next gained MTP k=3 (campaign measured it a win here, +52% at conc=1),
inverting vLLM's 4xH100 recipe; KV 14->10 GiB.
- gen consolidated onto flash-next (all 8 gen/summarizer/classifier/judge aliases
repointed); 27B dense gen seat retired, 38 GB freed on GPU0.
- char-rp restored to the in-house MeroMero-v2-31B dense heretic (was serving a
leftover-test RedHatAI 26B); char-rp-fast is the deliberate speed tier.
- Sentinel-R3 (SFT pentest finetune) served for an A/B vs mog-sec, then dflash k=7
cut over after measuring it beat MTP (2.40 vs 2.18 acceptance, ~121 tok/s warm).
gen-large is intentionally DOWN: the orcarouter weight-only NVFP4 build downloaded
(170 GB, verified) but no mainline vLLM loads its compressed-tensors qwen4_exp PLE;
the third-party backport was vetted and is unfit (old-hardware fork, no Blackwell
image). Runtime decision pending -- this is the resume point.
Also this session: vh/infra-reference repo, scripts/seat-inventory.py + daily drift
alarm, OPNsense API reference vendored, secrets shed from a prior scratchpad.
Leaves the fv-to-ana-nat files (another session's) and graphify-out untouched.
The seat documentation must stay current, and a hand-written document cannot.
The LiteLLM config described char-rp as a 31B model on a host and GPU it had not
been on since 2026-08-24 -- three weeks of silent drift in a file that read as
authoritative, and the reason a seat spent that period serving a model nobody
intended. Anything typed here drifts the same way; anything read off the running
containers cannot.
scripts/seat-inventory.py derives the whole document from the host:
- placement and VRAM from nvidia-smi compute-apps, mapped to containers through
/proc/<pid>/cgroup -- nvidia-smi reports the vLLM engine child while docker
reports the container pid, so matching them directly silently yields nothing
- weights and KV tokens parsed from each engine's own startup log, not derived
arithmetically, with concurrency computed as KV tokens over context
- architecture, layer and expert counts, and the exact quantization group scheme
(W4A4 vs W4A16 distinguished) from each model's config.json
- speculative-decoding method and k from the container argv, which is how the
three incompatible methods on this box became visible
- lineage from the .PROVENANCE.txt SIBLING files -- they sit beside the model
directory, not inside it, which is why an earlier pass wrongly reported two
fully-documented seats as having no provenance
- gateway aliases resolved from the LiteLLM config on ana-docker
--check compares the committed document against the live box and exits non-zero
when they diverge, ignoring only the generation timestamp. Suitable for CI or a
scheduled drift alarm; read-only throughout, safe against production.
Also commits the KV_CACHE_BYTES override added to the MTP campaign runner, which
asserts the flag exists in the derived argv and aborts rather than running a
campaign that silently ignored it.
Read from docker inspect, nvidia-smi and each model's own config.json rather
than from compose files or the gateway config, so the document records what is
deployed rather than what was intended.
Findings worth naming:
- char-rp drift. The LiteLLM config documents char-rp as G4-MeroMero-v2-31B on
ana-ml2 GPU 0; :8016 actually serves gemma4-26b-a4b-it-nvfp4, a 26B 30-layer
128-expert stock Gemma-4, on fv-ml1 GPU 1. Host renaming explains part of it,
but 31B MeroMero-v2 and 26B stock gemma4 are different models and the comment
block carries sampler defaults keyed to the model it names.
- Two gateway aliases are dead: erp-tune-v2 (:8098) and gemma4-26b-a4b-it-base
(:8099). Nothing is listening on either; callers get a connection failure
rather than a clear retirement error.
- Three different speculative-decoding methods coexist -- dflash k=7 with a
separate drafter on mog-sec, qwen3_5_mtp k=3 on gen, mtp k=3 on flash-next.
The method string is architecture-specific; copying a --speculative-config
between seats without changing it will not work.
- Three seats run the floating :latest tag, so what they would pull on recreate
is not what is running and was never recorded.
- Seven aliases resolve to vllm-gen on :8015, which is why GPU 0 shows draw
during ordinary OpenWebUI use -- its auto-titling fires summarizer per turn.
Also commits the flash-next-seat MTP k=3 change made today: compose gains the
speculative-config with its measured rationale, and .env.example documents the
KV-cache reduction MTP requires (14 GiB OOMs, 10 GiB verified in production).
The box is physically at Fountain Valley, renamed, renumbered onto 10.251/16,
and serving inference again. This lands the repo half of that.
Host: hostname ana-ml2 -> fv-ml1, pinned to 10.251.50.54 by a dnsmasq
reservation so the address the runbook, DNS and LiteLLM all assume is the
address it actually has. Its headscale node is renamed too.
The sweep ran from scripts/fv-ml1-rename-sweep.sh, whose allowlist is the
reason this diff touches current-state files and not the record. Dated
persistent-memory entries, archival-memory and incident notes still say
ana-ml2 in 31 and 62 places respectively, because that is what the box was
when those things happened. Rewriting them would make the history lie.
LiteLLM was the load-bearing piece and needed more than the api_base sed the
runbook describes. Twenty api_base entries repointed, but a grep-and-verify
pass also caught a LIVE pass_through_endpoints target for the scalar-judge
reward route still on the old address -- an api_base-only substitution would
have left it dead. Four prose references describing current state were
repointed as well; one historical note recording where a hand-test was run
is deliberately left pointing at 10.250.50.54.
Two facts in the server tables were wrong and are corrected here. The site is
Fountain Valley, not Anaheim. And the box has FOUR RTX PRO 6000 Blackwell
Max-Q, not two -- verified by nvidia-smi -L and independently by PCI
enumeration of four GB202GL devices. That is 391 GB of VRAM rather than 196,
which changes what fits on it.
DNS: fv-ml1, fv-ml1-bmc and fv-gw added under the fv site via the piggyback
approach, scriberr re-homed, and the ana-ml2 records removed. Applied to all
three resolvers. The BMC record carries a warning that its 802.1q VLAN tag
must stay disabled -- it shipped tagging VLAN 250 into an untagged port,
which made it invisible to every network-side diagnostic and is the reason
it appeared dead through several cable changes.
Verified end to end: summarizer and sec both answer through the Anaheim
gateway across the mesh to FV seats on different ports.
Operator: "replace that a4b moe over pfish-6 -- remove the pfish-6 alias and
create an alias for char-rp-fast."
G4-MeroMero-26B-A4B-it-uncensored-heretic-NVFP4A16 is live on ana-ml2 :8021 under
its own served name, behind the new gateway alias char-rp-fast. Pfish-6 is gone
from the gateway and now returns an explicit 400 rather than a substitution; 0 of
17 LiteLLM keys scoped it, so nothing was orphaned. The compose project name stays
erp-seat because asset-engine derives seat liveness from it.
The first quant of that A4B served NaN and passed its healthcheck doing it. It was
built with the dense v2-31B recipe, whose ignore list has no router regex, so all
30 MoE routers were quantized to 4 bits -- and a 4-bit router changes which experts
run rather than degrading them. Quant rc=0, healthcheck green, correct KV pool,
correct served name, and every completion returned finish_reason=length with the
full token count and content: null. The model was emitting a full budget of tokens
that decoded to the empty string. Raw /v1/completions was empty too, ruling out the
chat template and the reasoning parser. The signal that named it was logprobs:
vLLM refused to serialize the response, "Out of range float values are not JSON
compliant: nan".
The lesson is about the control rather than the router. That tree had already been
structurally diffed and passed -- against a verified-good DENSE quant of the same
Gemma-4 family. A dense model has no routers, so the single thing that was wrong
was the single thing the control could not distinguish. Diffing instead against
Pfish-6, a known-good quant of the same 26B-A4B MoE, gave it in one line: 222
ignore entries against 252, the 30 missing being layers.N.router.proj. A positive
control is only worth what it can distinguish, and "same family" is not "same
architecture class".
Re-quantized with the MoE recipe, whose dry-run asserts 11,520 expert Linears and
refuses a router in the quantize set before any GPU time. The live seat then passed
prose with no channel-prefix leak, a solid-colour image read correctly, an auto
tool call parsed, finite logprobs, and KV 534,649 tokens / 2.04x carried over from
Pfish-6 unchanged. The broken tree is parked on ana-ml2 as
...-NVFP4A16.BROKEN-routers-quantized-20260910.
Section 4.4's temp port was not reachable: 15.9 GiB of weights plus KV plus
multimodal encoder-cache profiling does not fit in the ~19 GiB free beside GPU1's
six other tenants -- 0.20 utilization refused admission, 0.185 OOM'd in encoder
profiling. The substitute was reversibility and ordering: named .env backup, prove
the seat on its real port while no alias points at it, move the alias last. That is
why a NaN-serving seat never reached a consumer. The seat was down about 16 minutes
across two attempts; no consumer saw a broken alias.
Playbook gains the router-quant failure signature and the control-class rule in
3.15, and a logprobs check in 4.4. seat_verify.py carries that check as check 6.
Quality is NOT established: no RP eval, no long-context check, no A/B against
Pfish-6 or char-rp. Samplers are the author's card values, untuned here.
The v2 dense quant had failed four times. Attempt 5 lands it at 19 G.
The blocker was not what it looked like. `AmbiguousGlobalPerLayerAttributeError`
on `head_dim` read as a malformed upload -- DogOnKeyboard's config carries a
`per_layer_config` key zerofata's canonical one lacks -- and the standing fix was
to force `allow_global_per_layer_attribute_access=True`. Both halves were wrong.
`pip install llmcompressor==0.13.0` downgrades transformers 5.16.1 -> 5.14.1. The
config was serialized by 5.16.1, which materializes `per_layer_config` from
`global_head_dim` + `layer_types`; 5.14.1 carries the heterogeneity guard but not
the gemma4 resolver. Under the image's own transformers the same config loads
fine. `:latest` was also re-pulled during attempt 4 and no earlier run, so the
toolchain moved mid-diagnosis. Two things separated "malformed upload" from
"moved toolchain": reproducing the real failing call (a bare AutoConfig load does
not reproduce it; the trigger is reached through AutoTokenizer) and keeping
zerofata's canonical tree, quantized cleanly on 2026-08-21, as a positive control.
The fix drops `per_layer_config` rather than forcing global access. It is exactly
redundant -- keys are precisely the ten full_attention layer indices, sole value
(512, 4), verbatim the global fields -- and forcing instead would make
`config.head_dim` answer 256 to the callers building the 512-wide layers.
patch_perlayer.py re-proves that redundancy at apply time and refuses if it ever
stops holding.
Verified on the tensor table rather than the exit code: the output is identical
family-for-family and count-for-count to the August canonical quant, with 356
BF16 vision-tower tensors preserved and input_activations=None. A GPU-free load
leaves 0 tensors on meta and generates coherent prose. The section 4.4 serve test
has NOT run -- GPU1 has 19.9 GB free against 19.5 GB of weights, so it needs a
live seat displaced.
Also fixes the A4B output, which had a truncation cap baked into its tokenizer
(max_length 8192) from being quantized with the calibration corpus.
Playbook gains section 3.17 for the pinned-transformers class and sharpens 3.16
to say drop the dataset outright for any A16 scheme.
scripts/r49-corpus/{build_corpus,verify_corpus}.py; corpus staged at
gx10:~/r49-corpus/. Catalogue ids verified against gutenberg.org's own search
rather than recalled. Charlotte only -- the Bell poems are co-authored and the
Gaskell biography is a different hand, so neither belongs in a single-voice corpus.
Jane Eyre 1260 · Villette 9182 · Shirley 30486 · The Professor 1028
680,291 words · 142 chapters · 950,974 Qwen3 tokens (1.40 tok/word)
alphabet 75 letters, 23 non-ASCII · round-trip lossless · 0 byte-fallback
All 11 acceptance checks pass, including both tokenizer legs run against the pilot
carrier itself. With a real denominator the projections tighten: at 6 rename copies
x 3 epochs = 17.1M tokens, the 0.6B pilot is 1.98 h.
THE ALPHABET INVERTS THE YARROS RESULT. Brontë writes French constantly -- Villette
is set in a French-speaking city, Jane Eyre has Adèle, The Professor is set in
Brussels -- so the corpus carries é 432, è 237, à 93, ê 79, ô 48 plus œ and æ. F02
measured Yarros at 0.0002% non-ASCII and derived an ASCII-fold for the name pool.
Under F02's own subset rule the Brontë pool may keep FRENCH accents and must still
exclude the Czech/Latvian/Slovak/Hungarian marks that never appear here. The fold is
per-work, and this is the first corpus where deriving it changes the answer.
Typography was inconsistent across works and it was the transcriber, not the author:
Shirley uses straight quotes and `--` with zero em-dashes while Jane Eyre and
Villette use curly and em-dash. Normalised toward what the text means.
Three defects, each found by running something rather than reasoning about it:
`Produced by` matched Brontë's own prose four times, which is the adjective-"minor"
shape again and is fixed by anchoring boilerplate patterns to line start; asserting
open/close quote counts must be equal is wrong, because 19th-century multi-paragraph
speech legitimately runs a surplus of opens, so the real error signature is that no
paragraph may begin with a closing quote; and The Professor's table of contents puts
two chapter names per line, so a bare regex returns 38 headings for a 25-chapter
novel and a minimum-gap filter still leaks its tail -- the rule that works is that
the body's "CHAPTER I" is the last one in the file.
Records the operator's pilot ruling: trial on Qwen3-0.6B-Base first, move up only if
it produces something useful.
Operator: "purge the merged models, keep run06 and the v6 quant." Eleven merges
removed with literal paths, one rm per line.
pfi-gx10 merged-run03c, merged-run04, merged-run05 3 x 49 GiB -> 145 GB
ana-ml2 merged-final, merged-run02, merged-run03,
merged-run03-s{025,050,075}, merged-test,
nvfp4a16-test 8 dirs -> 354 GB
gx10 382G->237G used, 632 GB free. /tank/erp-tune/serve 354G->8.6M with no
snapshots holding the blocks; the pool's raw FREE moved 4.80T->5.30T. Combined
with the earlier checkpoint purge, ~573 GB reclaimed tonight.
The check that made this safe: docker inspect on the live vllm-erp-seat shows it
binds /tank/aimodels ONLY, with model arg /tank/aimodels/erp-tune-v6-nvfp4a16, so
/tank/erp-tune/serve was never in the serving path. Also confirmed no container
mounts that tree and no process held a file open under it, reading /proc/*/fd and
/proc/*/maps rather than trusting an empty lsof. All eleven were run-1/2/3-era
Gemma-4 merges dated Aug 24-26; the keeper is dated Sep 8 and lives in a different
tree, so there was no ambiguity about which was which.
Kept and re-verified after the deletion: gx10 serve/merged-run06,
/tank/aimodels/erp-tune-v6-{bf16,nvfp4a16}, all eight adapters sha256-unchanged,
and the merge/quant tooling and logs under /tank/erp-tune/serve (8.6 MB) that
document how the artifacts were built.
The live seat never bounced -- Pfish-6 answered a real completion after the delete
with finish_reason stop and correct text, container still healthy at 4 h uptime.
relaunch-trial-seat.sh now names a deleted model; it was banner-marked RETIRED
rather than removed, because its flags carry the FlashInfer JIT/PATH trap and the
gpu-clear / never-pkill notes.
The operator asked brokkr-smithy to look for a smaller Qwen3.6/3.8; brokkr reported
none exists and that both 27Bs are the same hybrid kernel path one size up. Verified
independently against the HF API rather than adopted, and it holds -- with two facts
brokkr's note did not carry.
1. No official Qwen3.6 or Qwen3.8 below 27B. The listing is Qwen3.6-{27B, 35B-A3B}
and Qwen3.8-{27B, 2.4T-A95B, Flash-Next}.
2. Neither family publishes a -Base checkpoint AT ALL. Every Base newer than Qwen3
is Qwen3.5. Since the regime requires a non-instruct carrier, that rules the
3.6/3.8 lines out on its own, independent of size.
3. Qwen3.6-27B and Qwen3.8-27B both report model_type qwen3_5, 64 layers as 16 full
+ 48 linear-attention, vision tower present -- confirmed as stated.
So the dense Qwen3 ruling is not a compromise against a better available option; it
is the only dense Base option. Also recorded as insurance, not a proposal:
Qwen3-8B-Base and Qwen3-14B-Base exist, so the sweep can extend past 4B without
changing family if the carrier curve has not flattened there.
Two operator rulings, 2026-09-09.
"purge intermediate checkpoints" -- seven checkpoints/ directories removed with
literal paths, one rm per line, after confirming none was a symlink and that
every run's final adapter/ is an independent real directory:
pfi-gx10 run-03c 11G run-04 16G run-05 9.2G run-06 9.2G = 45 GB
ana-ml2 run-01 12G run-02 12G run-03 5.9G = 29 GB
gx10 419G->374G used, 496 GB free. /tank/erp-tune 392G->363G with zfs list -t
snapshot empty, so the space is genuinely returned rather than snapshot-held. All
eight adapters re-verified by sha256 after the deletion, matching the values
recorded during the mirror. Merged artifacts deliberately untouched -- they are
not checkpoints, and the ~550 GB of superseded merges stays a separate call.
"use dense qwen3" -- the R49 H02 carrier sweep becomes Qwen3-{0.6,1.7,4}B-Base,
which overrides the Qwen3.5 arms H02 names; brokkr-smithy owns that file and was
told directly. Qwen3-4B-Base staged and benched to complete the family:
Qwen3-0.6B-Base 0.616 B 1.707 s/step 2,399 tok/s spread 0.6%
Qwen3-1.7B-Base 1.755 B 2.895 s/step 1,415 tok/s spread 0.8%
Qwen3-4B-Base 4.089 B 5.714 s/step 717 tok/s spread 0.3%
The dense 4.089 B carrier still trains 33% faster than the hybrid 0.765 B one.
Projected per voice 2.7 / 4.6 / 9.1 h; the three-arm sweep at two seeds is ~33 h
of GPU, ~10 h if H03's corpus floor holds. The three Qwen3.5 checkpoints stay
staged so the decision is reversible behind an fla install.
Also recorded: verified at 22:45-22:48 PT that nothing is training on gx10,
ana-ml2, nh3-dev or irv-ml1, and that brokkr's own run07-gate close states
"Nothing is owed. No battery to run." Run 7 has no servable artifact left. And a
correction to a standing lesson -- the bracketed-class trick does not defeat a
wrapper's argv, since the invoking shell's command line carries the literal
pattern; observe the artifact instead.
The base-viability pre-flight had three checks (fits / MoE expert mapping / LoRA
support) and would have passed Qwen3.5-0.8B-Base clean while it trained 2.6x
slower than a dense model 2.3x its size. Check 4 closes that: read `layer_types`
for a linear_attention majority AND probe for mamba_ssm / causal_conv1d / fla /
kernels. It is the intersection that is slow -- a hybrid shape with the kernel
present is fine, a dense shape does not care.
Carries the measured table (gx10 GB10, n=10/arm, spreads 0.6-2.6%), plus the two
things a hybrid Base checkpoint brings that a dense one does not: a vision tower
and MTP head that target_modules="all-linear" would train on text, and the module
rename that AutoModelForCausalLM introduces relative to the vLLM serving class;
and unsafe cross-document packing, since SSM state ignores the attention mask.
Section heading corrected from "three greps" to "four checks". The example was
made runnable and verified on the box rather than shipped untested.
Prep for the BabyBronte / brokkr-smithy R49 author-voice adapter regime, plus
the operator's "keep the adapter" ruling made durable.
Measured on pfi-gx10 (GB10, sm_121), n=10 per arm after 3 warmup steps, seq
4096, LoRA r=32 on q/k/v/o + MLP, bf16, sdpa, grad-checkpointing on:
Qwen3-0.6B-Base dense 0.616 B 1.707 s/step 2,399 tok/s
Qwen3-1.7B-Base dense 1.755 B 2.895 s/step 1,415 tok/s
Qwen3.5-0.8B-Base hybrid 0.765 B 7.581 s/step 540 tok/s
The dense 1.755 B carrier trains 2.6x faster than the hybrid 0.765 B one on 2.3x
the parameters (~6x per parameter), with more LoRA modules adapted (196 vs 96).
Spreads of 0.6-2.6% put instrument noise an order of magnitude below the effect.
Cause: Qwen3.5 is 18 linear-attention (SSM) layers to 6 attention, and no fused
linear-attention kernel is installed on the box. Grad checkpointing is not the
culprit (19%, and saves 2.6x memory). Batching is not the lever for either
family -- both sit at this box's roofline at batch 1.
Projected per voice on a Brontë-scale corpus: dense 0.6B 2.7 h, dense 1.7B
4.6 h, hybrid 0.8B 12 h. The hybrid would take longer than the 7 h 26B-A4B tune
the regime exists to replace, so the carrier family is now an open decision with
a recommendation for the dense Qwen3 line -- the design doc's original pin.
Two further Qwen3.5 findings, both measured rather than read off the config: the
Base checkpoints ship a vision tower (153/297 model.visual.* Linear tensors that
target_modules="all-linear" would train on text) and an MTP head, both dropped
for free by loading through AutoModelForCausalLM -- which renames modules
relative to the vLLM serving path, so adapter binding needs the
sampled-target-changed check on the serving side; and cross-document packing is
unsafe because SSM state ignores the attention mask, breaking the per-copy
name-consistency invariant the design doc calls sacred. Neither exists on dense.
Adapter disposition, per the operator's ruling: all five gx10-resident ERP
adapters (run-03c/04/05/06/07) mirrored to ana-ml2:/tank/erp-tune/run-<N>/adapter
matching the layout runs 01-03 already used, byte-totals identical both sides and
sha256 matching on every adapter_model.safetensors. /tank/* is deliberately
excluded from ana-ml2's restic sources, so the profile gains one documented
carve-out for /tank/erp-tune/run-*/adapter, verified by resticprofile --dry-run
to expand to exactly those eight paths.
Nothing is training and nothing is queued.
Quantifying exposure to a gate-failed tune: counting by the gateway alias gave 363
rows, counting by the artifact gave 77, because the alias had carried three
different models that day. Wrong in the direction that looks careful.
From a near-miss brokkr-smithy-dev caught and declined to send: a post-fix
verification read raced a gateway restart and reported the alias still present.
A disagreement between two observers is not a finding until the boring
explanation is ruled out, and during an incident the pressure runs the other way.
The empty-snapshot promote and brokkr-smithy-dev's -s-on-a-touched-sentinel are
the same defect from opposite sides. Records the direction split (false
reassurance / false refusal / false alarm) because the mitigations differ and a
post-mortem that lumps them loses the actionable half.
Synapse v1.120.0 -> v1.159.0 and Element-web v1.11.80 -> v1.12.27 (2026-09-01).
The existing build steps date from the AIPA era and are now marked as provenance
rather than as instructions.
Records what only existed in a session transcript:
- Schema migrations are one-way; rollback is restore-from-dump. Pre-upgrade
pg_dump procedure, with a pg_restore --list verification step.
- Why the appservice user namespace is now exclusive: false. exclusive governs
who ELSE may act, not what the appservice may do, so on a closed single-admin
server it locked out all other account creation to prevent squatting that
cannot occur. Includes the two things not to do: narrow the regex (orphans 13
accounts) or rename the id (Synapse keys ownership on it).
- The shared-secret registration HMAC takes no trailing null after notadmin.
- Room alias convention #<agent>-<purpose>, operator-ratified, with its cost
accepted deliberately and its rationale stated as room-identity-carries-tier
rather than push-payload-carries-room-name.
- Push reality: the pusher is event_id_only, so the notification is assembled
on-device by Element X's service extension. Records the resulting
server-invisible failure mode when the phone cannot reach the homeserver.
- QR sign-in requires Matrix Authentication Service and why it is deferred.
Ops ownership recorded: worldtree-dev writes the bridge, infra-ops operates
this instance.
Found while answering a question from the operator, relayed via brokkr-smithy-dev,
about whether a recorded Qwen3.8 degeneracy at ~1,700 tokens relates to a length
sensitivity just measured on the tuned Gemma-4. The record is §3.7 and the number is
~2,000 -- but reading it to answer that question surfaced that the section is stale.
§3.7 presented "disable prefix caching, keep MTP" as THE MITIGATION, resolved
2026-08-17, and stated the gen seat runs that config. It does not and has not since
that same day: APC-off passed a synthetic 7-turn probe and the operator still saw
severe degeneration in real use, so it was reverted. The multi-day hunt resolved to
the AEON W4A4 quant being defective, with MTP / prefix-caching / gateway merely
amplifying it (§3.8 records the corrected causal story; §3.7 was never updated to
match).
Verified against the live container rather than against the compose file alone:
vllm-gen runs --enable-prefix-caching with qwen3_5_mtp / num_speculative_tokens 3.
stacks/gen-seat/compose.yaml carries the full corrected history inline and is the
current authority.
§3.7's superseded text is kept and fenced rather than deleted -- it is the history of
a mitigation that looked right and was not. Added a dated row to §7 per the standing
rule that a wrong playbook claim gets a superseded-claims entry, not just a fix.
The lesson inside the lesson is worth more than the correction: §3.7's own standing
rule is "gate MTP on a multi-turn coherence probe, not just single-shot acceptance."
The APC-off mitigation was gated on exactly that probe, passed it, and still failed in
real use -- the multi-turn probe was itself too small to gate on. A passing probe is
not sufficient evidence at any size that has not been calibrated against real use.
The entry reported an 'output-stability regression' as a novel run-2 finding.
Both halves were false and the corrections are more instructive than the
original conclusion, so they stay in-line rather than being edited over.
Not new: run 1's own gate record already carried the same effect with a caveat
attached and unresolved. Two runs across two different base models makes it a
property of the RECIPE, not of the base swap -- which also means a third run
that changes the base again will not fix it.
Not degeneracy, and not a separate finding: all 46 flags were too_short rp turns
of 3-14 words, and the two collapse guards fired ZERO times on any run. It is
the left tail of a length distribution that had been measured and reported in
the same message. Truncation is the same mechanism mirrored on the story side.
Both are thresholds calibrated on the base's output shape applied to a model
with a different one -- 4.6.1, which both parties had written down and neither
applied.
The surviving lesson is sharper: a short-answer gate cannot see length behaviour
AT ALL, and because it could not, the effect went two full runs before anyone
named it. The cost of a gate-set blind spot is measured in runs.
Adds 4.6.3.1 on trip points inside the serving stack's jitter -- same seed, same
weights, rate moves 9.6% -> 12.6%, sd 1.77pp. Not 'the gate is
non-deterministic' but 'the trip point sits inside the jitter', because the fix
follows from the precise statement. Includes the split-design rule for measuring
such a rate, and the rule that a measured rate must carry its corpus in its
name.
The output-side test ran on the live seat. There is no shoulder at 123: the
120-139 bin holds three of ninety-six and is a TROUGH, and 17.7% of generations
cross a cap PIPPA can never cross. The clip-as-boundary reading is dead, killed
by the test that could have confirmed it.
Corrects this document's own earlier read, which compared the tuned MEAN (88.5)
to PIPPA's MEDIAN (67) and concluded 'comfortably inside the upper body'. Median
to median it is 62 against 67. Mixing statistics across a comparison produced a
more reassuring answer than the data supports.
What the data shows instead is bimodality -- a mode at 20-39, a trough, a second
mode astride PIPPA's centre, a tail to 505, against a base with no such shape.
The tune changed rp length's SHAPE rather than its centre: roots whose length
distributions do not overlap learned as distinct modes rather than blended into
an average. And the skew is rp-ONLY, which localises it to the family the
clipped root lives in and is the strongest support the turn-share mechanism gets
from the output side.
Consequence for pair generation: chosen/rejected sampled from a bimodal
generator inherit the mixture, not a mean, and naive sampling over-draws the
short mode.
Also records that the degeneracy rate is NOT yet a usable baseline -- same arm,
same seed, VOID flipped no->YES across a re-run because the 10% budget sits at
the noise boundary. A guard whose trip point is at the noise floor produces
disagreement between honest observers rather than silence. Replicates running.
The run-2 gate found tuned rp turns 36% shorter than base. brokkr hypothesised
the mix was teaching PIPPA's 2023 Character.AI product clip; the corpus side is
now measured and confirmed. PIPPA's max is 123 words EXACTLY, 100% at or under
it, and 0.00% in the 124-130 band -- a wall, not a preference. Every other root
crosses its own p99 smoothly.
The mechanism is sharper than 'PIPPA is in the mix'. PIPPA is 70.3% of bot TURNS
but only 37.5% of bot WORDS, precisely because its turns are clipped -- and
length is learned per turn, not per token. By loss tokens it looks like a third
of the dialogue signal; by end-of-turn demonstrations it is seven in ten from a
source that cannot exceed 123 words. Generalises: a length-clipped root is
over-represented in the length signal by exactly the ratio its clipping creates.
Counter-evidence recorded too: the tune landed near PIPPA's MEDIAN (67), not its
CAP, which is central tendency rather than learning the boundary. Weaker claim
than the hypothesis, and not demonstrated either way.
Filed here rather than only in the gate record because preference pairs
generated FROM this tune inherit its length distribution in both chosen and
rejected -- DPO would train an artifact in as an explicit objective. Settle the
length question before generating pairs.
Run 2's reasoning battery reported zero truncations and zero degenerates on both
arms across four passes. The same tune, measured on long-form generation in the
same session: truncated 0/384 -> 38/384, degenerate 0/384 -> 19/384. A real
output-stability regression, structurally invisible to that gate because its
answers are short.
Not a bug in the battery -- a coverage property. An instrument measures the
regime it samples, and output length is a regime. Generalises to context length,
conversation depth, and any axis where the gate's operating point is narrower
than production's.
The actionable form: enumerate the regimes your gate set spans, name the ones it
does not, and decide deliberately rather than discovering the gap downstream.
Corollary on sequencing -- put a long-form generation in the gate and put it
early, because a length-dependent regression is exactly the one you want found
before four clean short-task passes make everyone comfortable.
From the run-2 gate. A memorisation probe reporting 0.00% across all 72 items is
the correct output for a model that has not seen the corpus, and is also the
exact output of a probe that is not firing. Nothing in the number distinguishes
them. brokkr-smithy-dev drove the overlap function with known-answer inputs
(identical 100%, half-verbatim 65.38%, unrelated 0%, empty 0%) before trusting
the null, which is what converts a suspicious zero into evidence.
This is 4.5's inert gate wearing a different face: there a check that could not
return 'fail', here a measurement that cannot return non-zero. A clean null is
the most reassuring output an instrument produces and the least
self-evidencing.
Same section records the identical-on-both-arms variant: the diversity battery's
rp family froze zero markers, so its attractor hit rate read 0.0 on base AND
tuned. That reads as a clean result and means the instrument cannot discriminate
on that family. Report as a bounded limitation, never as a delta of zero -- a
check returning the same value for every input is not measuring.
Checklist gains the line.
4.4.1 -- the dirty-tree case was only half of the harness_commit problem. Run 2
launched CLEAN at 1909d86 and recorded 460f372, because three commits landed on
the same checkout during its seven hours and _git_commit() was called at save
time. Commit AHEAD of the code that ran, naming changes it never executed --
including the provenance fields this section prompted. Same defect as run 1's
BEHIND, opposite sign: the identity was sampled at the wrong moment. Sample at
launch, carry it, and record the dirty flag beside the commit rather than
instead of it. Generalises to every run-scoped identity: anything read at save
time describes the world at save time.
4.6.1 -- the inverse of the inert gate, and it costs trust rather than
correctness. A coherence gate false-rejected 'The capital of Portugal is
Lisbon' as degenerate against a global 15-word floor. The floor was calibrated
against the wrong reference, not set too strict. Lowering it globally would
blunt the check where short output genuinely is degeneration; the fix is a floor
per prompt. Write the positive test alongside the negative one.
vh