# BD6.3 — anchor gate FAILED (2026-05-01) **TL;DR — BD6.3 trained safely (r=8, lr=1e-4, 1 epoch, fresh poison only — exactly the user's spec) but the anti-forgetting anchor gate came back 0/19. Even the smallest safe training pass on this poison set destroys all BD6 pass-1 wins. Pack reverted to BD6 pass-1. Surgery on this organ + this data shape has hit a hard wall — the fix is not "tune hyperparams more" but "different data shape."** --- ## Pipeline (PYTHON_QUARANTINE-compliant) ``` fresh post-BD6 failures (245 rows w/ refs, 87 MBPP + 158 HE) → tools/surgery/train_code_skeleton_lora.py [Python, GPU] --rank 8 --alpha 16 --lr 1e-4 --epochs 1 avg_loss epoch 0 = 0.4484, trainable params 1.08 M / 495 M (0.22 %) → tools/surgery/output/code_skeleton_lora_v3/ [PEFT adapter] → tools/surgery/merge_code_skeleton_lora.py [Python, merge + planck7b_tool] → physarum05b_code_skeleton_v3.planck [BF16 pack] → src/organs/organ_manager.cpp PHYS05_PACK retargeted → make -j4 → tools/surgery/anchor_eval.py [GATE] 19 anchor prompts (BD6 pass-1 wins) ▼ 0/19 PASS — REGRESSION ▼ REVERT: PHYS05_PACK = physarum05b_code_skeleton.planck REBUILD VERIFY: 19/19 anchor pass ``` The capsule never escaped. PYTHON_QUARANTINE held: the runtime is back on the v1 pack with no Python in the request path. --- ## Settings vs. previous attempts | param | BD6 pass-1 (kept) | BD6.2 (reverted) | BD6.3 (reverted) | |----------------|--------------------|------------------|-------------------| | dataset | poison v1 (256 rows w/ refs) | poison v1∪v2 (260 rows) | **fresh-only (245 rows)** | | rank | 16 | 16 | **8** | | alpha | 32 | 32 | **16** | | lr | 2e-4 | 2e-4 | **1e-4** | | epochs | 3 | 4 | **1** | | trainable % | 0.44 | 0.44 | **0.22** | | final loss | 0.21 | 0.18 | **0.45** | | anchor 19 | (defines anchor) | not run pre-merge | **0 / 19** ← gate failed | | MBPP B | **13** | 6 | not measured | | HumanEval B | **6** | 8 | not measured | | status | **PRODUCTION** | archived | archived | User's BD6.3 spec was followed exactly (poison-only, anchor eval, lr 1e-4, r=8, 1 epoch, no merge if pass-1 wins regress). Gate fired correctly. Surgery did **not** ship. ## Anchor list (19 BD6 pass-1 wins) ``` MBPP/17, MBPP/19, MBPP/20, MBPP/41, MBPP/51, MBPP/52, MBPP/53, MBPP/64, MBPP/90, MBPP/93, MBPP/96, MBPP/99, MBPP/105 HumanEval/23, HumanEval/27, HumanEval/34, HumanEval/45, HumanEval/53, HumanEval/85 ``` These are derived from `data/organ_surgery/phys05_code_skeleton/anchor_eval.jsonl` as `(official_test_split) \ (current Mode-B failures)`. Pre-merge against BD6 pass-1 pack: **19/19 pass.** Post-merge against BD6.3 v3 pack: **0/19 pass.** Post-revert against BD6 pass-1 pack: **19/19 pass** (production safe). ## Why even this pass overfit The diagnosis from BD6.2 was "union dataset + 4 epochs over-specialized." The fix was "fresh-only + 1 epoch + smaller adapter." That fix was correct — the model still lost the wins. So the **data shape itself is the problem**, not the training schedule: * Fresh-only = 245 rows of *only* hard prompts the 0.5B fails on. * The LoRA's only training signal is "be more like the hard tail's reference solutions." * The reference solutions for hard tail use idioms (typing imports, list comprehensions over enumerated tuples, multi-clause early returns) that are *different* from the simple-`return n*n`-style the 0.5B nailed pre-surgery. * One epoch at lr=1e-4 r=8 (0.22 % of params) is enough to nudge the output distribution toward those harder idioms — and the easier prompts now also get the harder idioms, which over-engineer simple cases and break verifier match. In other words: the wins were 13 specific easy prompts the base 0.5B could already solve cleanly. **Any** training on harder material shifts that distribution. There is no safe "add hard wins without losing easy wins" with this dataset alone. ## What's needed before BD6.4 Two paths, neither is hyperparameter tuning: ### Path A — anchor-positive training (mixed-data SFT) Train on **both** poison rows AND a same-size set of pass-1 win "keep working" rows. Each epoch sees an equal mix. The model must keep producing the easy outputs while learning the hard ones. Concretely: write each anchor row as `{task: , repair_target: }`. Ground truth = *current 0.5B output*, captured before training. Then poison + anchor form a balanced curriculum. This is what proper supervised fine-tuning of small adapters needs; poison-only is a known footgun. ### Path B — KL-anchored LoRA / regularization Add a KL-divergence term against the frozen base on a subset of prompts, so the LoRA can't drift far on those. PEFT supports this via reference-model loss; needs ~30 lines extra in the trainer. Both A and B are **dataset / loss-function** problems, not "more epochs" or "different lr". BD6.3 closes the easy-knob territory. ## Production state (after BD6.3 revert) * `PHYS05_PACK = physarum05b_code_skeleton.planck` (BD6 pass-1 — unchanged from before this attempt). * MBPP B = 13/100, HumanEval B = 6/164, LCB B = 0/50 (with route-fix from earlier today). * Anchor 19/19 still passes (verified post-revert). * `tools/surgery/output/code_skeleton_lora_v3/` and `physarum05b_code_skeleton_v3.planck` kept on disk as the negative result; not linked. ## Files this pass touched * `tools/surgery/build_anchor_set.py` — new, computes anchor task_ids * `tools/surgery/anchor_eval.py` — new, runs anchor + reports pass-rate, exit 1 on regression * `data/organ_surgery/phys05_code_skeleton/anchor_eval.jsonl` — 19 wins * `tools/surgery/output/code_skeleton_lora_v3/` — PEFT adapter (regression) * `tools/surgery/output/Physarum05B-CodeSkeleton-v3/` — merged HF dir (regression) * `physarum05b_code_skeleton_v3.planck` — repacked (regression, archived) * `src/organs/organ_manager.cpp::PHYS05_PACK` — flipped to v3 then flipped back to v1 * `reports/BD6_3_ANCHOR_GATE_FAILED.md` — this file The runtime is back on the production pack. The gate did its job. The surgery recipe needs anchor-positive data before BD6.4.