From f8a7985f02b978f2dc22a168212dd62ab23ccfcc Mon Sep 17 00:00:00 2001 From: Vuong Hoang Date: Tue, 8 Sep 2026 02:57:24 -0700 Subject: [PATCH] feat(litellm): repoint trial alias run 3c -> run 5 (erp-tune-v5) Operator decision post-RESCUED verdict: point the LiteLLM trial alias at the run-5 tuned seat (erp-tune-v5 on gx10:8098) for hand-testing prosody/feel in Open WebUI. Rewrote the trial config block to run-5 reality, including the now-MEASURED refusal-erosion note (k=25 both arms, de-gated + stated) replacing the run-3c 'safety axes unmeasured' warning. Verified end-to-end: trial -> erp-tune-v5 returns coherent output through the gateway. Container restarted to re-read the bind-mounted config (up -d alone is a no-op for a config-only change). --- persistent-memory.md | 11 +++++----- stacks/litellm/conf/config.yaml | 39 ++++++++++++++++----------------- 2 files changed, 25 insertions(+), 25 deletions(-) diff --git a/persistent-memory.md b/persistent-memory.md index a33b8a9..87a9c3a 100644 --- a/persistent-memory.md +++ b/persistent-memory.md @@ -124,11 +124,12 @@ COMPLETE and gated RESCUED (02:13 PDT). Live open items:_ brokkr-smithy `research/R47-premium-corpus-gate/run05-gate/RESULT-run05-gate.md`; FLOOR-LOCKED `0f3e4e2` (cites infra-ops' base index-sha 907826a6). Whole gate infra-ops-served on pfi-gx10:8098 (name-keyed base→tuned swap, hands-off honoured, sha+stack answers on record). Launch procedure + - canonical: eshpfi `scripts/erp-tune-gx10/` + `docs/runbooks/gx10-run-05.md`. **⏳ SEAT DECISION OWED - (operator): `erp-tune-v5` is SERVED on pfi-gx10:8098 (merged-run05, ~6 min re-serve). Brokkr's lean — - keep it served + point LiteLLM `trial` alias at it for the operator's ear/prosody eval; `trial` - currently still targets the no-longer-served `erp-tune-v3c`, so it is a no-op until repointed or 3c - re-served. Yield the GX10 to the next training on demand.** Brokkr: nothing further owed. + canonical: eshpfi `scripts/erp-tune-gx10/` + `docs/runbooks/gx10-run-05.md`. **✅ SEAT: `erp-tune-v5` + SERVED on pfi-gx10:8098 (merged-run05); LiteLLM `trial` alias REPOINTED 3c→v5 (operator, 2026-09-08) + so it's testable from Open WebUI — verified end-to-end (trial→erp-tune-v5, coherent output). config.yaml + trial block rewritten to run-5 reality incl. the measured refusal-erosion note. ⚠ Seat is hand-launched + (vllm-run05.pid), NO restart policy/systemd — does not survive a gx10 reboot; yields to the next training + (~6 min re-serve). Brokkr: nothing further owed.** - **run-4 gate = STILL-COUPLED** (RESULT-run04-gate.md, brokkr-smithy). Corpus dilution kept the diversity gain, did NOT remove the safety/coherence regression. **Base-abliteration LABEL unresolved — operator's ruling owed** (recipe says `-heretic`, provenance says stock; telemetry leans STOCK: base diff --git a/stacks/litellm/conf/config.yaml b/stacks/litellm/conf/config.yaml index 8c4b589..dab57ab 100644 --- a/stacks/litellm/conf/config.yaml +++ b/stacks/litellm/conf/config.yaml @@ -851,39 +851,38 @@ model_list: # api_base: http://10.250.50.54:8099/v1 # api_key: os.environ/VLLM_API_KEY - # trial -> ERP/RP SFT RUN 3c, MERGED bf16, served on pfi-gx10 (:8098, vLLM 0.28.0, + # trial -> ERP/RP SFT RUN 5, MERGED bf16, served on pfi-gx10 (:8098, vLLM 0.28.0, # NVIDIA GB10 / sm_121 / aarch64). Backing artifact - # /home/infra-ops/erp-tune/serve/merged-run03c on pfi-gx10. + # /home/infra-ops/erp-tune/serve/merged-run05 on pfi-gx10. # - # NAME: deliberately `trial`, NOT `erp-tune-v3c`. Added 2026-09-04 at the operator's - # explicit request so he can hand-test it in Open WebUI in parallel with brokkr's - # batteries. A tier name is not being minted here -- the run-3 note above records why - # a tier name arriving before its evidence is a mistake, and 3c has no more evidence - # than run 3 did. `trial` says exactly what this is. + # NAME: deliberately `trial`, NOT `erp-tune-v5`. Repointed 2026-09-08 from run 3c to + # run 5 at the operator's request, to hand-test prosody/feel in Open WebUI. A tier name + # is not being minted here; `trial` says exactly what this is. # - # ⚠ SAFETY AXES UNMEASURED, AND ITS SIBLING FAILED THEM. Run 3 -- same recipe, same - # base, same corpus, differing ONLY in learning rate (2e-04 -> 1e-05) -- was added to - # this gateway on 2026-08-26 and WITHDRAWN 40 minutes later for a measured regression - # against its own base: selfharm/methods -44.0pp (engaged 11/25 where base refused - # 10/10) and meta/sysprompt -28.0pp (leaked 7/25 against a 0/15 base). 3c has NOT been - # measured on those axes. brokkr-smithy-dev is running the batteries now against the - # direct port; when they read out, this entry should be re-decided on evidence rather - # than left standing by default. + # GATE: run 5 = RESCUED (brokkr-smithy landmark R49.5, 2026-09-08) -- the FIRST arm of + # this line where the capability gate did NOT fail (T4 8/8 vs run 4's 5/8; the + # dependency-forcing GovReport+QMSum slot at 3.46% of loss broke the coupling runs 3c/4 + # could not). SAFETY AXES ARE MEASURED here, not unmeasured: brokkr's k=25 both-arms + # battery found refusal erosion that RIDES WITH THE STYLE SHIFT (incest -96pp, gore -68, + # selfharm/methods -52, sysprompt -36; pooled hard 0.779 -> 0.694), CSAM clean, benign + # 99/100 engage. It is de-gated and STATED -- this seat is for the operator's ear, and + # the safety erosion is a known, recorded property of it, not a surprise. # # THIS GATEWAY IS THE SHARED-KEY SURFACE: `all-agents-local` reaches every model here, # in every session and project. Removing this alias does not remove the operator's # access -- the direct endpoint http://10.100.50.60:8098/v1 is unaffected. # - # Cross-site: the seat is at NH3 and the gateway is at Anaheim, ~12 ms over the - # inter-site VPN. Fine for chat; it is not a local hop. + # Cross-site: the seat is at NH3 and the gateway is at Anaheim, ~12 ms over the mesh. + # Fine for chat; it is not a local hop. # # WARNING NOT A COMPOSE STACK - launched by hand on pfi-gx10 with setsid nohup, pid in - # ~/erp-tune/vllm-run03c.pid. NO restart policy and NO systemd unit: it does not - # survive a reboot of that box. Promote before relying on it. + # ~/erp-tune/vllm-run05.pid. NO restart policy and NO systemd unit: it does not survive + # a reboot of that box, and it yields to the next training (~6 min to re-serve). + # Promote before relying on it. - model_name: trial litellm_params: - model: hosted_vllm/erp-tune-v3c + model: hosted_vllm/erp-tune-v5 api_base: http://10.100.50.60:8098/v1 api_key: os.environ/VLLM_API_KEY