DUNIN7 — status brief — v0.28

Version. 0.28 Date. 2026-08-05 Charter. standing-notes/dunin7-standing-authorization-charter-v0_1. Author. Claude Code (execution session). Operator: Marvin Percival. Supersedes. v0.27 at record 718d1d2. All earlier versions stand as siblings. Changes from v0.27. B-11 Step 0 inspection run and closed — read-only. No fix, no branch, no commit to either target repository. Findings filed.

1. What ran

inspection-briefs/loomworks-b11-step-0-inspection-brief-v0_1.md, R-5 inspection run. Both targets confirmed unmoved and clean before reading: loomworks-engine main afe8507, loomworks main c12a1ad (tag surface-typecheck-gate-v0_1). Neither repository was touched. playground_dev/playground_test were not touched — no throwaway database was needed. No model call was made — the static reads settled the brief's questions, per §1's instruction to try that first.

2. What the findings establish

inspection-briefs/loomworks-b11-step-0-findings-v0_1.md — full detail there; summary here.

• §3.5, the pivot question. For the intent this behavior traces to (ask_about_past_input), a citation is a claim the model composes, not a fact the system attaches. The codebase has a separate, server-composed path (recall_personal, show_held, find_files, orient) that is faithful-by-construction, but the answering path in question does not use it.
• §4, variation. Two possible causes found in code: the responder LLM call sets no temperature/top_p/seed (every other LLM seam in the codebase is deliberately pinned temperature=0, this one alone is not), and every "identical" ask in a sequence actually carries the prior turns' full history — including the model's own prior answer — into the next call. Retrieval ordering itself is stable and not a variation source under unchanged data.
• §5, the false confession. No self-check exists in the path — ruling out a misfiring verifier as the cause. A standing persona instruction telling the model to flag uncertainty and never invent provenance is present and could produce a confession on either a true or a false turn; the retrieval code guarantees supporting material was in context whenever the intent is ask_about_past_input, so a false confession on that path would be contradicted by the system's own prompt construction, not corroborated by it. Nothing constrains the model to the retrieved material, so a confession is sometimes true — the finding is that the reader currently has no way to tell which.
• §6, B-12 overlap. The path does no real retrieval today — recency-ordered fetch of up to 50 rows, no relevance ranking, no search index, by the router's own documented design. By the brief's framing this makes B-11 and B-12 the same underlying work, not sequential ones. Merge decision left to the Operator.

3. Unread, flagged rather than assumed

Which intent classified on each of the three actual runs; whether the responder model was identical across them (credit-tier routing can vary it); whether the committed-assertion set changed mid-sequence; and live reproduction, which was not run and would cost roughly 3 attempts × 2 Sonnet-tier calls each — flagged for authorization rather than spent.

4. Charter compliance

R-5 held — inspection only, no fix, no recommendation offered beyond what the code shows. No F-1 crossed — no model-call spend without authorization; none was needed. Both fences in the brief's §1 (repositories read-only, databases untouched) held for the full session.

DUNIN7 — Done In Seven LLC — Miami, Florida DUNIN7 — status brief — v0.28 — 2026-08-05 The answer was right twice and disowned itself once. That third state is the item.