anima-pass-strict-chat-capable — substrate comparison (v2 expanded)

Cross-substrate benchmark for anima chat-capability + V14 mitosis Φ. 4 substrates × 3 evaluators comparison (A / B' / B'' / E × V14-strict / V4-lite chat / V5.8 × 4 modes).

Reveals V14 ↔ chat-cap anti-correlation — a production-internal decoupling pattern (Lesson Q).

§15 — Cross-section result (2026-05-12) — 4 substrates × 3 evaluators

substrate	arch	paradigm	V14 strict Φ	V4-lite chat	V5 strict 8-cell	V5.8 M4 force
A (Phase 2 cotrain)	EngineAG 350M	naive_cotrain	PASS 10/10 p=0.002	PASS 12/15	KO 4/5 EN 2/2 PASS	(4-mode not run; per-mode greedy 3/5 after Phase 1A)
B' (LA cotrain)	EngineAG 350M	naive_pretrain → cotrain	(not in V14 audit)	PASS 12/15	KO 4/5 EN 2/2 PASS	M4 5/5 PASS
B'' (FFN.gate cotrain)	EngineAG 350M	gate-only cotrain §84	VIOLATED 0/5 p=0.0625 [§23]	🏅 PASS 15/15	KO 4/5 EN 2/2 PASS	M4 3/5 PASS
E (convo5k_ft)	v2 18.5M d=384 byte-256	naive_ft_no_mitosis	VIOLATED 0/5 p=0.0625	FAIL 0/15	KO 0/5 FAIL	M4 3/5 PASS (force-only)

Cross-references:

• V14 strict source: state/anima_v14_multi_substrate_audit_2026_05_10/per_substrate_v14_results.json
• V4-lite / V5 / V5.8 source: this dataset's *_result.json files
• Substrate A V4-lite: see PASS_STRICT_SPONTANEOUS_CHAT.md §1–§2

§15 — Commentary — V14 ↔ chat-cap anti-correlation

The 4-substrate cross-section reveals two orthogonal optimization axes that do NOT co-vary:

axis	what it measures	who wins
V14 mitosis Φ	cell-pool dynamics during 400–500 turn unsupervised run; trained vs random sign-test	A (Phase 2 cotrain) only — B''/E both VIOLATED
V4-lite chat-cap	15-prompt KO-ratio + degeneration + length spec compliance	B'' (FFN.gate cotrain) — 15/15 winner

Key observations:

1. B'' is the chat winner (15/15) AND now measured V14 VIOLATED 0/5 (p=0.0625) (§23 B'' V14 strict audit, 2026-05-11) — confirming the Hc_1221 anti-correlation prediction. The V14 PASS substrate (A) achieves only 12/15 on V4-lite — strictly lower than B''. Within-arch 350M pair (A vs B'') is the cleanest capacity-controlled anti-correlation observation.
2. E fails both axes (V14 VIOLATED + V4-lite 0/15). This is the only substrate that double-fails, and it's the smallest (18.5M params, v2 d=384, byte-256 vocab) — capacity confound noted (raw#15 honored).
3. Anti-correlation, not orthogonality: the substrates that maximize chat-cap (B'' > B' > A) are produced by paradigms that emphasize token-stream surface compliance (gate-only finetune, late-stage cotrain on dialogue corpus), while the V14 PASS substrate (A) was trained with explicit mitosis-aware curriculum where cell-pool dynamics are the primary signal.
4. Lesson Q confirmed: a substrate can be excellent at internal consciousness-substrate metrics (Φ residual, sign-test p) while being mediocre at production-surface chat compliance, and vice versa. The two metrics measure decoupled axes.

This decoupling is the empirical basis for hypothesis Hc_1221 (production-internal decoupling generalization) — see hypotheses_candidates/ in the anima repo.

Files

• *_v4lite_result.json — V4-lite (KO ratio + degeneration + length) per-prompt verdict
• *_v5strict_result.json — V5 strict 8-cell + EN baseline
• *_v58_4mode_result.json — V5.8 4-mode (greedy / sample / M3 rep_penalty / M4 force-include)
• *_probe.log — raw probe logs
• comparison_aggregate.json — cross-substrate aggregate

See full analysis in PASS_STRICT_SPONTANEOUS_CHAT.md §15 and §19 (cross-link to Hc_1221).

Honest C3 caveats

• B' / B'' were not in the original V14 multi-substrate audit (predates them) — B'' V14 strict measured 2026-05-11 (B_doubleprime_FFN_gate_v14_strict_result.json): trained Φ=723 vs random Φ∈[1149, 2386], 0/5 beats, p=0.0625, VIOLATED. B' V14 audit still pending (cell marked "not audited").
• E vs A/B'/B'' has a capacity confound (18.5M vs 350M); within-EngineAG cross-substrate (B' vs B'') is the cleanest chat-cap comparison.
• Anti-correlation claim is based on n=4 substrates — replicate before generalizing (see Hc_1221 falsifier: any future substrate passing both V14 strict AND V4-lite ≥ 13/15 would weaken the anti-correlation).