# BLACK_DOG_ORGAN_AUDIT Honest accounting of what Black-Dog actually does to organs vs what the architecture diagram promises. Read 2262 DAG entries; cross-checked against the three Black-Dog call sites in `src/main.cpp`. ## 1. Black-Dog code exists; it is BARELY wired `src/runtime/black_dog.cpp` + `include/black_dog.hpp` define `black_dog::Store`, `extract_features`, `hash_action_chain`, `get`, `update`, `save`. The store reads/writes a JSON file on disk. `bd_store` is invoked at exactly **3** call sites in the entire `--chat`/`--task` runtime: | line | function | writes conductance to DAG? | |--------|---------------------------|----------------------------| | 337-347| `run_chat_identity` | **NO** — `(void)cond_before; (void)cond_after;` at line 348 silently discards both. The store is updated; the DAG entry is not. | | 3008+ | `run_ariz_e2e` | YES — line 3331-3334 writes both fields. | | 3447+ | `run_chat_organ_route` | YES — line 3485-3486 writes both fields. | **Every other `--chat` path does NOT call Black-Dog at all:** - `run_native_terminal_task` (line ~2128) - `run_native_code_repair` (line ~1131) - `run_native_tool_call` (line ~1387) - `run_form_replay` (line ~2350) - `run_chat_json_repair` (line ~416) - `run_native_terminal_task_organ_first` (just added in PHASE-13.WIRE) — bypasses BD too Result: the high-traffic paths that actually shipped this sprint do not feed Black-Dog. Food/poison labels are written via shortcut at lines 378/731/1157 (computed inline, not via `bd_store.update`), and conductance never moves. ## 2. Per-organ Black-Dog accounting (all 2262 DAG entries) ``` organ n pass fail food≠0 poison≠0 cond_b≠0 cond_changed food_avg poison_avg physarium_7b_chat 1966 1465 501 1418 501 0 0 0.72 0.25 physarium_7b 139 127 12 135 135 119 136 2.15 1.58 phys05_claim_extractor 57 57 0 57 0 0 0 1.00 0.00 json_repair (legacy stub) 40 27 13 9 13 10 13 0.27 0.26 phys05_code_skeleton 28 28 0 28 0 0 0 1.00 0.00 phys05_json_repair 20 20 0 0 0 0 0 0.00 0.00 phys05_test_writer 0 – – – – – – – – phys05_triz_contradiction 0 – – – – – – – – phys05_renderer 0 – – – – – – – – phys05_cache_matcher 0 – – – – – – – – phys05_critic_lite 0 – – – – – – – – phys05_wound (NEW spec) 0 – – – – – – – – ``` ## 3. Three concrete failures the audit surfaces ### A. `physarium_7b_chat` — food/poison labels written, conductance never moves (1966 entries) `food_score` and `poison_score` ARE set (1418 nonzero / 501 nonzero respectively), but `conductance_before = 0` for every single one of the 1966 entries. The store is updated only inside `run_chat_identity`, which discards `cond_before/after`; all other paths skip BD entirely. **Implication:** Black-Dog has no influence on routing. The 86.9 % traffic that goes to `physarium_7b_chat` will keep going there regardless of whether it fails. The dog has no leash. ### B. 0.5B organs that DID fire — auto-pass, zero poison (105 entries) `phys05_claim_extractor` (57 calls): `pass=57/57`, `poison_nonzero=0`, `food_avg=1.00`. Same shape for `phys05_code_skeleton` (28/28) and `phys05_json_repair` (20/20). This means: **either the verifiers on those routes auto-pass, or the food/poison computation is "set 1.0 if return code 0" without checking anything semantic.** Looking at `run_chat_organ_route` line 730-731: ```cpp e.food_score = 1.0; e.poison_score = 0.0; ``` — hardcoded. Conductance does update via `bd_store.update` (cond_before nonzero appears 0 times because the keys are fresh on first call), but the food/poison fed in are placeholders. **The dog isn't tasting the food; it's just clicking the bowl.** ### C. 5 of 8 0.5B organs are stillborn — 0 calls, 0 BD signal `phys05_test_writer / triz_contradiction / renderer / cache_matcher / critic_lite` and the new `phys05_wound` have never been invoked. They have no DAG entries, no food, no poison, no conductance. There is no poison data to harvest into surgery datasets. They are dark. ## 4. Where conductance ACTUALLY moves Only `physarium_7b` on the ARIZ route shows conductance behaviour: ``` ariz_e2e_44b2fc775a633665 physarium_7b cond 0.6416 → 0.6533 ariz_e2e_979dde753459b039 physarium_7b cond 0.6935 → 0.6948 ariz_e2e_0a66cdcb747bfbfd physarium_7b cond 1.2325 → 1.1260 ``` 136 entries out of 139 had `cond_before ≠ cond_after`. ARIZ pipeline is the only place where Black-Dog is actually a learning loop. Everything else is decorative. ## 5. The four bugs that block PHASE-13.BLACK_DOG_ORGAN_COLONY 1. **`(void)cond_before; (void)cond_after;` at main.cpp:348** — `run_chat_identity` discards the only BD signal it computes. Trivial fix: remove the `(void)` casts and write the values into `e`. 2. **`run_native_*` paths don't call `bd_store` at all.** Each of `run_native_terminal_task`, `run_native_code_repair`, `run_native_tool_call`, `run_form_replay`, `run_native_terminal_task_organ_first` needs: - `extract_features(input, route_name)` at entry - `bd_store.get(...)` to read prior conductance (for routing) - `bd_store.update(features, chain, food, poison)` at exit - DAG entry write with cond_before/cond_after populated. 3. **`run_chat_organ_route` hard-codes `food=1.0 poison=0.0` (line 730-731).** Real food/poison must come from the verifier (`food = verifier_pass ? 1 : 0; poison = capsule.aborted ? 1 : 0` etc.). Currently the 0.5B organs look perfect on every call because the verifier signal isn't being piped into the BD loop. 4. **5 organs have never fired.** Without DAG entries there is no poison to harvest, no surgery dataset, no QLoRA target. They need to be put in the path of `--chat` traffic on dispatcher routes that actually reach `--chat`. Today `Route::ariz / renderer / cache_matcher` are wired only in legacy `--task`. ## 6. Recommended order (BD2 → BD3 → BD5 → BD6) Before any organ "rewire" or QLoRA surgery: 1. **BD2 fix-and-instrument first.** Patch the four bugs above so every `--chat` DAG carries real food/poison/cond_before/cond_after. Without this, every subsequent number is unfalsifiable. 2. **BD3 poison harvest.** Once BD writes are real, point a script at `dag/runs/*.json` and stream all entries with `poison_score > 0` or `verifier_pass = false` into `data/organ_surgery//poison_train.jsonl`. 3. **BD5 learning curve probe.** Run 30 tasks × 5 rounds. The 5th round must show: lower fallback_to_7B count, lower poison sum, higher conductance on successful chains, faster wall. 4. **BD6 QLoRA only after BD5 establishes a real baseline.** ## 7. Stop conditions - No GREEN claim is allowed where any of these is true: - `food_score / poison_score / conductance_before / conductance_after` are all zero in the DAG entry - the DAG entry's `organ` field is `physarium_7b_chat` AND the route in question is supposed to be served by a 0.5B chain - the surgery dataset directory is empty for the organ being claimed "live" - The PHASE-13.WIRE chain (`run_native_terminal_task_organ_first`) is **not yet ready for production claims** — it does not call BD; its organs are untrained for the role. It is a baseline probe instrument, not a victory. Authoritative artefacts: - `reports/BLACK_DOG_ORGAN_AUDIT.md` (this file) - `reports/black_dog_organ_audit.json` - `reports/ORGAN_TRAFFIC_AUDIT.md` (prior — strict organ counts only) - `docs/ORGAN_CANONICAL_MAP.md` (prior — name reconciliation)