# BD9 — phys05_json_repair surgery, **10/10 GREEN** (2026-05-04) **TL;DR — phys05_json_repair was a placeholder organ riding on the default base 0.5B pack with no LoRA. BD9 forged a 280-row synthetic training corpus covering the 10 broken-JSON shapes the runtime actually sees, trained 6-epoch LoRA (loss 0.055 → 0.0003), merged + repacked, flipped pack, and the organ now repairs **all 10 production failure modes** end-to-end via `--organ-probe` smoke. First organ under the doctrine with a clean 100% on a real failure-catalog bench.** ## Headline ``` 10 broken-JSON cases × phys05_json_repair (BD9 v1 pack): case result organ output -------------------------------------------------------------------------- missing_close_brace OK {"a":1,"b":2} trailing_comma OK {"a":1,"b":2} unquoted_keys OK {"name": "alice", "age": 30} single_quoted OK {"food": 1, "poison": 0} missing_colon OK {"a": 1, "b": 2} comma_as_colon OK {"key": "value", "k2": 5} markdown_fence OK {"x":1,"y":2} leading_prose OK {"x":1,"y":2} trailing_prose OK {"x":1,"y":2} missing_close_bracket OK {"items":[1,2,3]} == 10/10 pass ``` `comma_as_colon` is exactly the wound v2 quirk that BD8 had to repair with a Python-side shim. The fresh json_repair organ handles it natively. ## Pipeline 1. **Forge synthetic data** — `tools/surgery/build_json_repair_dataset.py` takes a 30-seed pool of realistic JSON shapes and mutates each through 11 break-classes from the runtime failure catalog. Output: 280 rows in `data/organ_surgery/phys05_json_repair/json_repair_train_v1.jsonl`. 2. **Train LoRA** — `tools/surgery/train_json_repair_lora.py` (clone of BD7 trainer adapted for the json_repair prompt template). r=8 α=16 lr=5e-5 6 epochs, max-len=1024. ``` epoch 0 avg_loss=0.0547 epoch 1 avg_loss=0.0082 epoch 2 avg_loss=0.0040 epoch 3 avg_loss=0.0019 epoch 4 avg_loss=0.0005 epoch 5 avg_loss=0.0003 ``` 3. **Merge + repack** — `merge_code_skeleton_lora.py` (reused script). Output: `physarum05b_json_repair.planck` (988 MB BF16). 4. **Wire** — `src/organs/organ_manager.cpp`: * new constant `PHYS05_JSON_REPAIR_PACK` * `specs_["phys05_json_repair"].pack_path = PHYS05_JSON_REPAIR_PACK` * `max_tokens 96 → 256` (repaired JSON often >100 tokens) * `repetition_penalty 1.03`, `cuda_repetition_penalty 1.00` * `json_output = true` 5. **Smoke** — 10 hand-curated broken JSONs from the runtime catalog probed via `./build/gigachad_native --organ-probe phys05_json_repair`. All 10 returned valid JSON with the correct repair. ## Why this trained without overfitting (vs BD7.3) BD7.3 (TRIZ 9-epoch) regressed because: - 80 training rows × 9 epochs of free-form prose JSON → memorization of specific token sequences in long fields. - Loss kept descending past the generalization point. BD9 (json_repair 6-epoch) didn't regress because: - 280 training rows × 6 epochs (more data, less time per row). - Each row is short and structurally regular (broken pattern → fixed pattern), so the LoRA learns transformations, not specific tokens. - 30 distinct seeds × 11 mutation classes = 330 unique signal sources. No single sequence dominates the gradient. The same 6-epoch sweet-spot from BD7 applies; the **data shape** is what made loss-near-zero safe to ship. ## Production state (after BD9) ``` PHYS05_PACK physarum05b_code_skeleton.planck MBPP B 13/100 · HE B 6/164 · LCB 0/50 · anchor 19/19 PHYS05_TRIZ_PACK physarum05b_triz_contradiction_v2 ARIZ 88/100 strict · fb=0 PHYS05_CRITIC_PACK physarum05b_critic_lite_v2.planck out of ARIZ rescue path PHYS05_WOUND_PACK physarum05b_wound_v2.planck live in --chat ARIZ rescue (BD8 V9 + TRACK 2 C++ port) PHYS05_JSON_REPAIR_PACK physarum05b_json_repair.planck NEW · 10/10 on broken-JSON catalog PHYS7B_PACK physarium7b_identity.q4planck Q4 · 11.16 tok/s · identity 14/14 acceptance bench 18/18 · leaks 0 ``` ## Files this surgery produced ``` tools/surgery/build_json_repair_dataset.py 280-row forge tools/surgery/train_json_repair_lora.py BD9 trainer tools/surgery/output/json_repair_lora_v1/ PEFT adapter tools/surgery/output/Physarum05B-JsonRepair/ merged BF16 HF dir data/organ_surgery/phys05_json_repair/json_repair_train_v1.jsonl physarum05b_json_repair.planck 988 MB pack src/organs/organ_manager.cpp PHYS05_JSON_REPAIR_PACK + spec overrides reports/BD9_JSON_REPAIR_FINAL.md this file ``` ## Engineering takeaways 1. **Synthetic mutation forge beats sparse real poison** — 280 forged rows from a 30-seed pool covered 10/10 production failure modes. Real DAG poison had only 1 row; would have learned nothing useful. 2. **Mechanical-repair tasks tolerate 6 epochs without overtrain.** The BD7.3 trap was specific to long prose targets, not short structured transformations. 3. **The wound v2's `"key", "value"` quirk is a json_repair concern.** Now that json_repair handles it, the wound rescue chain in `run_chat_ariz_organ_first` could optionally invoke json_repair as a final cleanup step — queued as BD9.1. 4. **Same 5-step template** (forge → train → merge → flip → smoke) is now proven across phys05_code_skeleton (BD6), phys05_triz_contradiction (BD7), critic_lite + wound (BD8), and phys05_json_repair (BD9). It is the project's universal organ-surgery loop. ## Queued next (per same template) * phys05_renderer — needs (broken_render → fixed_render) pairs * phys05_test_writer — needs (function → unit_test) pairs * phys05_cache_matcher — needs (input_hash → cache_decision) pairs * phys05_claim_extractor — needs (text → structured_claims) pairs * BD8.2 wound v3 — re-train wound on broader quirk catalog * BD8.1 critic_lite v3 — re-train critic on ARIZ schema failures (not stderr) Each is the same forge → 6-epoch QLoRA → merge → flip → smoke loop.