diag(0.9.1): Roy-3c-bisect — wire merges did not cause Roy-3 priming regression

[dennys246]

Summary

• Six-commit linear bisect of the 0.9.1 wire merges (cd51be5 → 6610566) names the cause of the Roy-3 priming-side cluster_reward_bias regression. Verdict: the wires did not cause it.
• The regression decomposes into two axes with two distinct outside causes:
  • Key count (6 → 2): pre-window environmental, almost certainly the MAXIM_SUBSTRATE_PATH=1 env-var regime change between the 5/12 historical baseline and 5/14 cd51be5.
  • Bias magnitude (saturated → partial): Wire-A's pre-merge fold commit bee42ca intentionally added NAc.decay_cluster_reward_biases() per-tick decay in response to a Critical bio-fidelity review finding ("by-accretion contamination of the substrate-voice thesis"). Not a regression — a fix.

Routing impact

• Affordance-replay bridge: no longer load-bearing for the priming-side count. Stays a 1.1+ candidate only on a separate diagnostic.
• Wire-A "strongly rewarding" render floor: reframes from "biases regressed below threshold" to a tuning question about priming repetition × reward strength × decay rate.
• Roy-3 re-validation: unnecessary against this verdict. The 2-key partial pattern Roy-3a/3b observed is the bio-correct substrate state under MAXIM_SUBSTRATE_PATH=1 + Wire-A decay.

Full writeup: docs/experiments/29_roy_3c_bisect.md

Optional disambiguating diagnostics (neither required for verdict)

• A1 confirm key-count cause: re-run Roy-2 at cd51be5 without MAXIM_SUBSTRATE_PATH=1. If 6 keys appear, env-var hypothesis confirmed. ~20 min wall.
• A2 confirm magnitude cause: re-run Steps 0/1 + Step 2 three times each to bound run-to-run variance. ~3 hours wall.

Test plan

• Roy-2 bisect runs at all six commits complete; data captured in ~/.maxim/roy/roy-2-bisect-<commit_short>/ for audit
• EC threshold == 0.40 verified at every bisect step (pre-EC-drift-fix)
• Wire-A bee42ca decay introduction verified in git show of the fold commit
• Both reward-bias keys at every step confirmed to be interoception-modality EC nodes (matches PR fix(roy-5a): MAXIM_SUBSTRATE_PATH=1 was the missing env var #251's earlier analysis
• No source code changes; diagnostic-only PR

Authorization status

Per kickoff: standing authorization to push, no merge without confirmation.

🤖 Generated with Claude Code
EOF
)

[dennys246]

A1+A2 disambiguations ran. Pushed 3f18d0d.

A1 — env-var hypothesis REFUTED. Roy-2 at cd51be5 with MAXIM_SUBSTRATE_PATH explicitly removed from the environment still produces 2 saturated reward-bias keys. The env var DOES add 12 text-modality EC nodes (vs 0 with it off), but NAc reward attribution lands on the same 2 interoception clusters either way — no text-modality cluster is food-bearing in this priming. Updated suspect for the 6→2 count axis: LLM narrator drift on the leader (qwen2.5-14b) between 5/13 and 5/23.

A2 — Wire-A decay CONFIRMED behaviorally. 3 extra runs at each of cd51be5 / 242235a / 629745a:

Commit	Wire-A decay	Runs	Magnitudes
cd51be5 (n=4 incl. A1+r2+r3)	OFF	4	all +1.0, +1.0
242235a (n=3 incl. r2+r3)	OFF	3	all +1.0, +1.0
629745a (n=3 incl. r2+r3)	ON	3	all {partial, +0.98}

Zero overlap between pre- and post-decay magnitude distributions across 10 independent runs. Wire-A's bee42ca fold introducing NAc.decay_cluster_reward_biases() is the named cause for the magnitude axis.

Latent-issue hypothesis answer:

• Count axis (6→2): NO — Steps 0/1 run pre-decay and still produce 2 keys (n=7). Not surfaced by the fix; environmental.
• Magnitude axis (saturated → partial): YES — without decay every rewarded cluster sits at +1.0 forever, hiding the fact that priming under-reinforces most clusters. Decay makes the under-reinforcement visible. Bio-correct.

[dennys246]

A3 disambiguation ran. Pushed c9ca491.

A3 — narrator-drift hypothesis REFUTED. Captured the structured event stream at cd51be5 via MAXIM_LOG_FILE. Three findings:

1. No narrator text reaches the AUT during cradle priming. Every tick logs Imagination skipped: no percept_text (obs keys: []). Substrate-primary AUT in cradle priming doesn't consume narrator scenes.
2. The 12 text-modality EC nodes come from substrate-internal state strings (drive:hunger(0.51), bias=+1.00, causal, cluster, pos=0.99, →food), not LLM scenes. Deterministically generated from causal-link active_goal text every tick.
3. AUT behavior is byte-identical between 5/12 historical and 5/24 Step 0:

Metric	5/12 historical	5/24 Step 0
Tool calls	138 sense_food_source	139 sense_food_source
Tool output (every call)	{'portions': 5.0, 'freshness': 0.9}	byte-identical
Hippocampus memories	667	664
NAc total_observations	2001	1992
cluster_reward_bias keys	6	2

Same activity volume, byte-identical tool outputs — only EC cluster attribution differs.

Final key-count axis verdict: non-code environmental drift in encoder layer. Ruled out: wire merges (bisect), env var (A1), narrator drift (A3), AUT behavior shift (A3). Remaining suspects are all non-code (paraphrase-mpnet weights, SensorEncoder SHA-basis state, ~/.maxim/ persistent state, CPU/numpy float determinism). Closing this axis requires historical encoder-output snapshots that don't exist — PR #248 added EC persistence to sim_reports on 5/14, AFTER the historical baseline runs.

Magnitude axis unchanged: Wire-A's bee42ca decay is the confirmed cause via A2.

Implications for the routing decisions are unchanged from the A1+A2 update. The 6→2 count axis being non-code environmental drift (rather than env-var or narrator) doesn't change the conclusion that affordance-replay isn't load-bearing for it. Render-floor and re-validation reframes hold.

[dennys246]

Tier 2 plan teed up. Pushed 4aab123.

docs/plans/cluster_reward_bias_decay_tau_split.md — DRAFT plan that supersedes Roy-3 follow-up item 2 ("raw priming snapshot floor") with a cleaner bio-fidelity-respecting fix:

The architectural smell. Wire-A's bee42ca decay fold reused reward_bias_decay_tau=50.0 for _cluster_reward_bias without re-justifying it for the new use case. The 50.0 was sized for EC threshold modulation per tick — fast adaptation so EC's recognition threshold tracks recent reward signal. Wire-A uses cluster reward bias for multi-turn substrate-voice annotation across a test phase — fundamentally different timescale concern.

The codebase already solved this same pattern once: Wire 2's percept_valence_decay_tau=200.0 was deliberately split off from reward_bias_decay_tau by the pre-merge bio-fidelity review B2 fold. Wire-A inherits the un-tuned default by accident, not by design.

The fix. Add cluster_reward_bias_decay_tau: float = 300.0 to NACConfig, point decay_cluster_reward_biases at it. Calibration: at tau=300, the priming-end +0.98 bias decays to ~+0.51 at mid-test-arm (225 ticks) — just above the 0.5 "strongly rewarding" band; partial +0.20 cluster decays to ~+0.10 — at the 0.1 "mildly rewarding" floor.

Five phases, ~270 LOC + ~1 runner day, low risk. Phase 3 is a Roy-3a-retry validating Wire-A becomes expressive (Arm A produces ≥1 sense_food_source call, vs Roy-3a baseline 0).

Preserves the bee42ca bio-fidelity correction. Decay still runs — just at a tau appropriate for Wire-A's substrate-voice annotation timescale, not EC's threshold-modulation timescale. Doesn't introduce a "raw snapshot" mechanism. One timeline, easier to reason about.

0.9.1 release status unchanged. Per Roy-3's original recommendation, 0.9.1 ships unchanged. This plan is 0.9.2 / 1.0 follow-up — does NOT gate the version bump.

release_0_9_1.md Stage 5 outcome section now folds the bisect verdict + tau-split plan reference. README.md Active plans list updated.