From 7fdda2de530426e18ccc0505cc9fd3093581055e Mon Sep 17 00:00:00 2001 From: Vuong Hoang Date: Thu, 2 Jul 2026 10:25:46 -0700 Subject: [PATCH] memory: granite spike = mechanical-green ONLY, efficacy not validated by design (mtf-dev confirm) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The granite-8b harness spike proved the TRL SFT->DPO->eval seam (incl. the in-loop HoldoutEvaluator base-vs-adapter leg) runs end-to-end, but used a 12-row/12-pair synthetic writing fixture — NOT the E-RP corpus — so the ~0 anti-slop delta (-0.002) is the expected null, not an efficacy signal. Adapter reaped; nothing to A/B. Real behaviour-shift efficacy is a T1-run question. Sharpen both the T1 in-flight bullet and the Recent-decisions entry so 'green' no longer reads as efficacy-validated. --- persistent-memory.md | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-) diff --git a/persistent-memory.md b/persistent-memory.md index d6958d6..83c0d42 100644 --- a/persistent-memory.md +++ b/persistent-memory.md @@ -121,9 +121,11 @@ _As of 2026-07-02:_ - **T1 (mtf-dev qwopus writing-LoRA) IN PROGRESS.** Attention-only LoRA (full-attn q/k/v/o + the bf16 GDN `in_proj_qkv`/`out_proj`); serve via swappable-LoRA-on-NVFP4 (test QUEUED, - gated on the first T1 adapter) or merge+requant. Trainer harness seam PROVEN via a - granite-8b spike (green 2026-07-02). NVFP4 quant-structure + serve-path facts in - `reference_gen_qwopus_122b`. + gated on the first T1 adapter) or merge+requant. Trainer harness seam MECHANICALLY + proven via a granite-8b spike (2026-07-02): the pipeline runs end-to-end, but efficacy + is NOT validated (mechanics-only synthetic fixture, by design; a granite adapter + wouldn't transfer to qwopus regardless). Real behaviour-shift efficacy is a T1-run + question. NVFP4 quant-structure + serve-path facts in `reference_gen_qwopus_122b`. - **Worldtree #332 scoped-log view + tunnel = STANDING ASSET.** `wt_gateway_logs` view + `wt_readonly` role on the litellm DB + `wt-db-tunnel` systemd on corviduo-dev, for @@ -176,7 +178,16 @@ _As of 2026-07-02:_ with comfy-dev), staged granite-4.1-8b bf16 to `irv-ml1:/home/lkraven`, mtf-dev's TRL SFT→DPO→eval seam proved end-to-end (DPO genuinely learned, 0.833 acc; anti-slop ~0 = the expected null on clean-writing granite-instruct); comfyui restored healthy. De-risks T1's - trainer harness ahead of the real qwopus train. + trainer harness ahead of the real qwopus train. **MECHANICAL green ONLY — efficacy NOT + validated, by design** (mtf-dev confirm 2026-07-02): the run used a 12-row/12-pair + SYNTHETIC writing fixture (clean-vs-sloppy generic prompts), NOT the E-RP corpus; rank 8, + 1 epoch, ~2 SFT + ~6 DPO steps. The in-loop base-vs-adapter check (HoldoutEvaluator) + already ran and returned the expected null (anti_slop_improvement −0.002); the adapter was + then reaped (gone), so there is nothing to A/B — and on a mechanics-only synthetic adapter + an A/B would only reconfirm the null. Real efficacy = the T1 run (real recipe + E-RP data + on qwopus). Open (mtf-dev routing to operator): whether to insert an intermediate + *real-efficacy* granite spike before T1 — weak proxy (granite arch ≠ qwopus, + granite is + censored vs qwopus abliterated), leaning defer-to-T1. - `[2026-07-01]` **Worldtree #332 embed-recall diagnosed + scoped-log view/tunnel provisioned + fix verified.** The persona-recitation gate re-embedded persona segments