# Handoff — 2026-08-05, revision b **Supersedes `docs/HANDOFF-2026-08-05.md`.** Session A in that file — P1-0c — ran, stopped at its source gate, and is now closed as superseded. Session B — P1-1a — is unaffected and **must not be restarted or interrupted**; it was never blocked by the gate. Everything either agent needs to know about *state* is in `docs/PROJECT-STATE.md`. Everything about *task* is in its prompt file. Nothing else needs to be explained in the chat window. --- ## What happened, in three sentences P1-0c was told to **extract** six values verbatim from the pinned prototype. Three of them are rows that C11, C12 and D6 **create** — a prototype pinned before a correction cannot contain the row that correction adds — so the assignment stopped on a category error rather than on a project blocker. The fourth, A3, produced a real finding: the pin recomputes both codes per run, which falsifies the premise that a permanent natural key was there to be extracted. The agent's stop was correct against its instructions and its report is accurate. **The prompt was wrong, not the execution.** Re-running P1-0c reproduces the stop exactly. --- ## Session A′ — P1-0d · the corrective assignment Fresh session. Paste this as the opening message: > Read `docs/PROJECT-STATE.md` first — it is the only source of project state in this repository, > and every other document is a record of a moment, not current status. > > Then execute `docs/AGENT-PROMPT-v5-P1-0d.md` exactly as written. It supersedes > `docs/AGENT-PROMPT-v5-P1-0c.md`; do not execute that file. `docs/p1-0c-status.md` is accurate, > is reused, and is not re-derived. > > Documentation only. **Deliverables 1–3 have no dependency on any seed row — do them first and let > them land.** Then transcribe the rows already found at `9359c67`, then author the three v5.0 rows > under per-field provenance, then resolve A3 by measurement. > > **`absent` is a legitimate result for a field. It does not stop the row, and a row does not stop > the assignment.** Inference from neighbouring rows, counts, or the annex's marginal domains > remains forbidden. ## Session B — P1-1a · unchanged, keep running Nothing about P1-1a changes. If it is mid-flight, leave it. If it needs a nudge, the only new fact it needs is: > `MatchGroupID` loses its CodeLedger bucket — it is a run-scoped grouping label, not an entity. > `BRCode` keeps its bucket; issuance stays P1-1b. Create the column and the surviving bucket only. --- ## What P1-0d will hand back `docs/p1-0d-status.md`, an updated `PROJECT-STATE.md`, `docs/p1-0-c20-authored-rows.md`, the extracted rows inside the drafts, provenance markers on every column, B2/B3/B4/B12 specified, and the A3 collision-measurement table. **The line to check first is the list of fields it marked `absent`** — with an authored-row model, that list is the real result, exactly as "what it could not find" was under the extraction model. **The second line to check** is whether C9 and C15 are ticked in `docs/project-todo.md`. They close on transcription alone and require no decision from anyone; if they are still open after P1-0d, the assignment did not do deliverable 4. ## B1 — decided 2026-08-05, and P1-1a needs the one-line version The reference-counted default was **rejected**. B1 closed by specifying the storage model the lifecycle draft never stated. If P1-1a is anywhere near the tombstone path, give it this: > **Byte storage is per capture. A byte object belongs to exactly one capture event and is never > shared between captures, even byte-identical ones. `RetainedContentHash` is a duplicate-detection > index and is never a storage pointer — no storage-pointer column, no path derived from the hash, > no two `Document` rows resolving to the same storage asset ID.** Reference counting is not > implemented. §3 eligibility is unchanged and is now the second barrier, not the only one. Purge > is capture-scoped storage hygiene and is never named, logged or presented as erasure. Full reasoning and the rejected alternatives: `docs/p1-0-b1-decision.md`. ## Do not hand over - `docs/AGENT-PROMPT-v5-P1-0.md`, `-P1-0b.md` and now `-P1-0c.md` — executed or superseded, history. - `docs/HANDOFF-2026-08-05.md` — this file replaces it. - `docs/p1-0-review-findings.md`, `docs/FDR-to-Sibyla-decisions-for-Luis*.md` — historical, and both open with a "do not sign off yet" that was true on 2026-08-04 and is not now. - Any summary of state written in a chat window. If it is worth an agent knowing, it belongs in `PROJECT-STATE.md`. ## The process fix, so this does not recur Two lines belong in `AGENTS.md` at the next edit: 1. **An assignment states, per value, whether it is *extracted* or *authored*.** A value a decision creates is never listed as an extraction. This one mislabelling cost four unrelated deliverables. 2. **A gate is scoped to its deliverable, never to the assignment.** A missing seed row stops that row. It does not stop provenance markers, governance specification, or decision folding that never touched it.