Route live research through llm-connect and record pilot findings

Assistant: codex
Assistant-Model: gpt-6-astra
Assistant-Session: 01a0cffe-206f-7323-89d3-b20ca4cf4bf9

research/reviews/LLMConnectPilot.md

@@ -0,0 +1,60 @@
+# llm-connect authority pilot
+
+Date: 2026-09-24. Reviewer: codex; initial editorial review, not independent validation.
+
+## Execution
+
+Per user direction, live execution now calls the deployed llm-connect `/execute`
+service. No provider key was retrieved. A temporary Kubernetes port-forward exposed
+`activity-core/llm-connect` locally; no deployment was changed. Requested profile:
+`custodian-triage-balanced`; resolved provider/model: OpenRouter / Gemini 2.5 Flash.
+Credentials and routing remain service-owned. Client retries are disabled; internal
+service retry count and dollar cost are unknown.
+
+Receipts under `research/bench/infospaces/` record source/prompt/config digests,
+requested settings, actual model, usage and finish reason:
+
+- `carniato-authority-02`: one response, `length`, 4,470 total tokens. Rejected.
+- `carniato-authority-03`: four responses, all `stop`, 12,638 total tokens. Workflow completed.
+- `local-first-authority-02`: one response, `length`, 12,489 total tokens. Rejected.
+
+Total reported usage across these six service requests: 29,597 tokens, including
+failed attempts. These totals do not measure dollar cost. The `-01` failed direct
+provider runs remain intact. No successful local-first generated artifact exists.
+Generated output remains unreviewed in the bench; this editorial disposition does
+not rewrite the original run or convert assisted critique into independent review.
+
+## Review of the completed Carniato draft
+
+Artifacts: [candidate](../bench/infospaces/carniato-authority-03/artifacts/generated/candidates.md),
+[relations](../bench/infospaces/carniato-authority-03/artifacts/generated/relation-proposals.md),
+[assisted critique](../bench/infospaces/carniato-authority-03/output/evaluations/source-review.md).
+
+Disposition: **revise candidate; reject proposed graph relations; admit nothing**.
+
+- Workspace P01 maps provisionally to the existing Immediate Feedback candidate.
+  Its family and scale labels do not satisfy the repository vocabulary contract.
+- The candidate's chat, social, checkbox and form examples have no independently
+  located implementations. They are illustrative hypotheses, not known-use evidence.
+  Its recurrence assertion is unsupported by this bounded source packet.
+- The proposed `requires` edges target C01/C02 claim IDs as if they were candidate
+  IDs. These are evidence references, not valid pattern relation targets.
+- Server authority is contextual in this source. A mandatory server-authority
+  prerequisite for immediate feedback is not established, and conflicts with the
+  local-first comparison hypothesis in the earlier manual review.
+- The assisted critique approved recurrence and both relation justifications.
+  This is a concrete false-positive example for our admission process: automated
+  self-review cannot supply missing evidence or validate graph identity.
+
+The two narrowly sourced claims are useful starting points, but the larger dynamics,
+failure modes and examples extend beyond the excerpt and need separate evidence.
+No FEP identities, canonical pattern files or graph edges were added.
+
+## Follow-up owned by FEP-WP-0002-T02
+
+Before another local-first run, select explicit research-appropriate execution
+settings and diagnose repeated truncation within a bounded call/token plan; simply
+raising the cap again is not justified by these results. Refine prompts to prohibit
+invented known uses, distinguish claims from pattern identities, and require
+unsupported recurrence to remain unknown. Verify those rules on this captured
+failure before the next live run. T03 still owns cross-source case walkthroughs.