Shay 82c58dd85e docs(analysis): finding — speculative epistemic ceiling does NOT blunt learning

7-agent architecture audit + adversarial verification refutes the intuition that
all-speculative packs cap the model's learning. The reasoning/learning loop is
blind to epistemic_status (recall_realized ignores it; speculative facts chain to
deductive-closure fixed point). Lexicon-status promotion is compile-and-display-only
(runtime epistemic state comes from grounding_source, not the lexical row). The
verified/as_told distinction is honesty-disclosure-only and intentional. All three
coherence-promoters are dormant (proof-carrying demo-only, energy-policy default-off,
review never sets coherent). Conclusion: do NOT promote-to-coherent for learning
(cosmetic at best, false-verified wrong=0 hazard for facts); the only genuine lever
is wiring the sound, fail-closed proof-carrying path — a separate deliberate decision.

2026-06-14 14:58:10 -07:00

5.6 KiB

Raw Blame History

Finding: the speculative epistemic ceiling does NOT blunt learning

Date: 2026-06-14 Status: substrate finding (refutes a compelling-but-wrong intuition; corrects the "B1 coherent-promotion" plan). Method: 7-agent architecture audit + adversarial verification (run wf_f7c37d46-16f). TL;DR: All cognition/logos pack entries are speculative, and the engine's determine path can never label a conclusion verified — but neither fact blunts learning. The speculative state is an honesty/disclosure mechanism, by design, and the reasoning/learning loop is blind to it. The plan to "promote entries to coherent so the model can learn" is unwarranted; the lexicon variant is purely cosmetic and the fact variant risks manufacturing false verified.

The question

Audit found 25/28 packs have no epistemic_status (→ default speculative), and generate/determine/determine.py::_basis only returns "verified" when every ground is coherent. Hypothesis: this caps the model and is why it underperforms at learning from contemplation/practice. Tested rigorously. Refuted.

What is actually true (verified, high confidence)

Axis A — promoting a lexicon entry to coherent is compile-and-display-only

language_packs/compiler.py:75-92 maps a coherent lexical row → EpistemicState.DECODED, stored per-surface on the manifold (compiler.py:357 → vocab/manifold.py:114).
No reasoning/recall/realize/learning path reads that tag. The only non-test reads (compiler.py:406, :559) merely copy it forward; workbench/logos.py reads it for the Studio audit badge (display).
The runtime's per-turn epistemic state comes from grounding_source, not the lexical row: any pack-resolved lemma → grounding_source="pack" (chat/pack_grounding.py:461,489) → DECODED (core/epistemic_state.py:133-144). So a speculative and a coherent pack lemma are runtime-indistinguishable. (Adversarially upheld; no counterexample.)

Axis B — every realized fact is `SPECULATIVE` by construction

Told facts: generate/realize/realize.py:217 (status = EpistemicStatus.SPECULATIVE, "COHERENT is never a default (ADR-0021)"). Derived facts: realize.py:372. Also generate/proposition.py:264, generate/realize/quantitative.py:108.
determine._basis (determine.py:88-91) gates verified on those RealizedRecord grounds — never the lexicon. So lexicon promotion can never reach it.

Axis C — three coherence-promoters exist; all are dormant at runtime

Proof-carrying (ADR-0218): teaching/proof_promotion.certify_promotion + vault/store.py:417,486-507 apply_certified_promotion (flips COHERENT at :507). Built, sound, independently-re-verified, fail-closed — zero runtime callers (only tests + demos/proof_carrying_promotion).
Energy-policy (ADR-0148): vault/store.py:361 promote_eligible_entries, wired into the live turn boundary (chat/runtime.py:2324,2509) but default-off (config.vault_promotion_enabled = False). Promotes on energy cooling, not coherence/entailment.
Curator review: teaching/review.py:274 runs live on correction turns but its epistemic_status is a passed-in param defaulting SPECULATIVE; the live call sets no status. Computes no coherence.

So verified is genuinely unreachable at runtime — by deliberate conservatism, not a bug. pipeline.py:379's COHERENT proposal branch is dead relative to the live path.

Axis D — the learning loop is blind to `epistemic_status` (decisive)

generate/realize/recall.py:19-62 recall_realized ignores epistemic_status entirely. SPECULATIVE consolidated facts are immediately re-readable as premises, so deductive closure climbs to fixed point with the speculative ceiling fully in place. The ceiling does not impede chaining or learning.
The verified/as_told distinction is observational only — it selects a surface string (render.py:31) and gates nothing else.
Intentional and sound: consolidate.py:22, realize.py:331-334 — "a sound inference never upgrades the standing of its speculative premises." idle_tick and contemplation are proposal-only/SPECULATIVE by construction (consolidate.py:21-22, runtime.py:871).

Conclusion

The all-speculative state is not why the model struggles. The learning loop works (monotone climb to deductive closure) regardless of it. The real learning wall is elsewhere (recognizer/coverage), not epistemics.
Do not promote entries to coherent "to help learning." The lexicon variant is cosmetic (no runtime consumer); the fact variant would make speculative content admissible-as-evidence — a wrong=0 hazard (false verified) — for zero learning benefit.
The only genuine epistemic lever is wiring the proof-carrying promotion path (Axis C #1) so a deductively-proven derived fact can become coherent → the engine could honestly emit verified for what it proved (vs. was told). That is a real honesty/capability upgrade and it is wrong=0-safe (fail-closed, independently re-verified) — but it was deliberately left unwired, so turning it on is a ratified-decision, scoped separately in docs/handoff/proof-carrying-wiring-scope-2026-06-14.md.

Cross-references

ADR-0021 (epistemic status = revision position), ADR-0218 (proof-carrying promotion), ADR-0148 (energy-policy promotion), ADR-0206 (epistemic disclosure), docs/issues/proof-carrying-coherence-promotion.md, and the parallel honesty finding docs/analysis/holonomy-resonance-proof-not-robust-2026-06-14.md.

5.6 KiB Raw Blame History