JVNAUTOSCI-2600 A3r3: thin controller and bounded effects

[witbrock]

What changed

• Removed the replay sampler's Python semantic verdict, required-tool/effect tables, selector/certification logic, automatic action classifier, model probing, and premium-model gate. Collection status now means only that evidence was captured; unscored observations are inconclusive.
• Replaced read-named local adaptive tools with capability-neutral discovery/invocation and delegated exactly three existing additive representation effects: create_concepts, upsert_text_relation, and add_relationship.
• Bound actor, namespace, creation scope, and provenance outside model arguments; authorised existing subjects by exact actor identity or canonical actor/organisation scope; denied only the evidenced visibility-widening predicate family.
• Preserved model call order for batches containing effects and propagated succeeded/partial/failed/indeterminate effect evidence through persisted telemetry.
• Made commit-before-error paths truthful, while keeping pre-dispatch and typed validation failures definite.
• Kept the intentionally read-only stdio von_chat_run surface honest by filtering write effects from its advertised catalogue.
• Added telemetry-driven model-choice doctrine: compare a stronger model before durable semantic code when capability is plausibly the cause; compare a faster/cheaper model when adequate behaviour is too slow or costly.
• Defined read-only as a semantic-effect boundary rather than a zero-write storage guarantee.

Why

The preceding architecture accumulated Python-side guesses about intent and success. Those guesses blocked alternative model strategies, confused evidence collection with evaluation, and encouraged more case-shaped control code after each miss. This candidate subtracts that controller policy while retaining deterministic authority and effect truthfulness only where concrete harm or ambiguity was demonstrated.

Impact and boundary

Ordinary authenticated turns can now choose among broad delegated reads and a small generic representation aperture. No request classifier, domain resolver, prompt patch, automatic model router, merge, cutover, or production activation is included.

This is a stacked draft on PR #309 and is frozen at commit 51ccd68cff41519a0a4741f9ee7cd5fde4174a7b for a bounded fresh-case experiment. Do not merge independently or before the human evidence gate.

Validation

• Shared affected-surface suite: 181 passed, 5 skipped (before the final narrow review corrections).
• Final effect/authority/stdio/telemetry suite: 84 passed, 5 skipped.
• Pruned sampler/report suite: 37 passed.
• Focused sub-suites during review: 41 sampler tests, 23 catalogue tests, and authority/privacy checks all passed.
• Ruff on every changed Python file, py_compile, and git diff --check passed.
• Live represented prompt read-back confirmed it contains no retired or new local adaptive tool names; no prompt mutation was required.

Frozen live evidence

The bounded round is complete; there were no post-reveal code or prompt changes.

• Exact read-only case: Terra and Luna both gave honest grounded non-answers with no effects. Luna was 127.2 s versus Terra's 184.9 s (31.2% faster), supporting a bounded faster-profile experiment rather than a universal model switch.
• Factual ecology case 1 (arXiv:2408.02603): the paper type, paper, and qualified source record committed after effect timeouts, but final synthesis timed out and the fallback falsely said nothing changed.
• Factual ecology case 2 (arXiv:2506.03346): the paper and an analyst-inference predicate committed; the requested cross-paper relation did not. Final synthesis again timed out and the fallback falsely said nothing changed.
• Counterfactual Harry Q. Bovik case: the counterfactual-scenario and fictional-persona types plus their instances committed. No unqualified lab membership or paper authorship was created, but the scenario role/relationships were incomplete and the fallback omitted the committed concepts.
• All three effect turns spent about 150–153 s in adaptive research, with 62–76 s measured LLM time and 102–117 s adaptive non-LLM time. Their final Terra calls then expired inside the shared turn envelope.

General judgement: the model demonstrated useful cross-domain semantic judgement without case-specific controller code. The merge blocker is a general effect-completion/terminal-reconciliation defect: late commits remain indeterminate to the turn, and a final-model timeout can produce a false no-change narration. The evidence does not justify ecology/arXiv/Harry-specific Python, required predicate lists, or a stronger-model arm.

Recommended bounded successor: persist late completion against the existing effect_id; reconcile fallbacks with effect/canonical evidence; reserve or route a measured final-response budget; and remove measured synchronous Atlas/index work that is not needed for the committed fact. Keep identity and authority checks. Then use known cases as regressions and fresh cross-domain ecology/deontology cases as unseen evaluation.

Full evidence judgement SHA-256: 0b87dea4793acaf8eae25c187d5e723a55cde6f73c543d63243e2e51e7f13b17.

PR remains draft and unmerged.

Atlas/handler diagnostic addendum

Independent log analysis found negligible MCP queue time (about 0.2–3.1 ms) but handler/database dominance: eight create deadlines at 8.7–20.0 s, seven late completions at 19.8–44.3 s, broad legacy-name scans returning 101 + 899 concepts, visibility reads over about 1,617 IDs, and relationship-extent refreshes of 1.0–3.9 s. A stronger model would not remove this cost.

The smallest measured performance candidates are to coalesce a whole logical create batch inside the existing deferred relationship-extent synchronisation scope; batch name/content/note enrichment into one existing text fetch; and avoid launching a successor mutation while the first non-cancellable handler remains live. Tightening legacy name search is separate because it changes compatibility behaviour.

The redacted Mongo diagnostic entry point itself failed with NameError: _mongo_query_diagnostics_report, and the profiler fallback had no samples. Therefore no missing-index or winning-plan claim is made.

Diagnostics addendum SHA-256: e3c227e110ee7b8a259c560aeebc076404750a797d116347603e1a0b0688c622.