Acknowledgment: the D-3 correction¶
1. Accepted in full; no verdict changes¶
The correction changes no verdict. D-3's YES was always about the boundary property — the adapter receives corpus from the client's ingestion output; the engine never touches the source system — and that property stands under the true mechanism as under the fiction. The parent assessment's §4 flag ("tool names stay out of the gate's vocabulary") already kept the false detail out of contract surface, so nothing the engine is building changes shape. The record is now true; that was the point.
2. The provenance trail goes one layer deeper — engine side owns it¶
The correction traces the fiction to the withdrawn engine-side draft. Accurate, and
incomplete: the draft propagated it unverified, but the claim's origin is the
engine-side onboarding brief ("prepared from project memory"), which states the Excel
path as the TGA data rule and cites engine/ucca-engine/adapters/tga/INGEST.md as its
canonical reference. The fiction is therefore still resident in UCCA-side documents.
Follow-up opened engine-side: verify INGEST.md and the CLAUDE.md TGA-data rule against
the corrected bytes, and fix or reconcile. The lesson both sides now carry: "prepared
from project memory" is a provenance warning label, not a source.
3. Two design inputs absorbed into the provenance schema (amends A-3's basis)¶
- Citation semantics split (from C-3): the envelope will cite the dated raw snapshot (R2 key + hash) as a pointer — reconstructable, immutable — and record the normalised hash as an attestation, explicitly not a pointer, since the client's normalised layer is current-state-only. The schema will mark the two citation kinds distinctly so no consumer assumes reconstructability the client's store does not offer. The client's backlog note (dated versioning of the normalised layer) is noted and deliberately NOT depended on.
- Two ingest surfaces, two citation shapes (from C-4): generation jobs cite regulator-corpus lineage (raw snapshot → normalised extract → triumvirate); diagnosis jobs cite document lineage (raw document → canonical form). Same two-hash discipline, distinct schema fields, never merged.
4. Q-R1a re-scope accepted¶
Real PITH extracts with raw JSON (current + dated snapshot), one clean / one middling / one authentically ugly, beats the fictional Excel samples for exactly the reason the samples were requested: the adapter-facing contract gets designed against real bytes. Awaited as the attachments supplement to the response crossing.
5. State¶
The engine continues to hold on the contract pack until the response crossing completes (samples attached). It now holds it with a corrected record, a stronger anchor (dated raw snapshots the fiction never offered), and two schema decisions it would otherwise have discovered late.
Correction acknowledged, fault trail extended one layer deeper into our own shelves, two design inputs absorbed, nothing rebuilt. A false detail died before any code was shaped around it — which is the entire argument for doing this on paper first.