Skip to content

ANCHOR-ONE-GATHER-01 — ARCHITECT VERDICTS

Seat: Architect (Fable, fresh window). Date: 2026-08-14. Disposes of: the Driver's ANCHOR-ONE-GATHER-01 inventory (Q-P2P-009 → 011, F-01 → F-12). Ratified by Tim, 2026-08-14, this session, after plain-language examination. Q-P2P-010 was delegated to the Architect by Tim in writing ("it's your call") with a stated leaning to option (b); the Architect held (a) and Tim's conditional approval of (a) stands. The reasoning is recorded at V-P2P-010 so a future reader finds a decision, not a silence. Inventory accepted as the design's evidence base. Population searched before absence claims, one name-blindness caught and repaired mid-job, every figure sourced, offline computations marked. No finding rejected.

V-P2P-009 — Provenance carry-across: NOTHING CROSSES. Abandonment in writing, censuses first.

Ruling (Tim, 2026-08-14): option (c). No pre-existing provenance crosses into the clean Pith. The new Pith's provenance is stamped only from its own ingestion forward.

The two classes at issue — (a) synced_at on 125,920 component rows and (b) the dated versions/ snapshot series (≥144,674 objects, earliest 2026-05-09) — are our own observation metadata, not register truth. TGA publishes no corpus vintage, so they are irreplaceable by nature; doctrine 1 therefore requires either sacredness or deliberate written abandonment. The evidence for abandonment: the size-diff census over the 20.06% enumerated found zero differing series across 5,823 codes with ≥2 snapshots — five months of snapshots that observed nothing changing. Class (c), the 2026-03-09 enrichment provenance, is already dead under the KN ruling.

Conditions attached:

  1. The two zero-cost censuses named in F-08 run before the abandonment ruling is filed, so the ruling names exactly what it abandons: (i) completion of the versions/ size-diff census over the remaining ~80%, plus the unrun same-length sample; (ii) the per-code join of the training/ prefix against the D1 15,169 code set. Zero TGA requests; read-only against the Prototype morgue. Brief: GATE6-CLOSE-01 (staged alongside this document).
  2. The design does not wait on the censuses. Under (c) nothing crosses either way; the clean-Pith schema freezes now. The census result attaches to the abandonment filing when it lands. If the census finds materially differing series — real observed change in the window — that is new information and returns to Tim before the abandonment is filed.
  3. This closes North Star gate 6 in two halves: the answered half (nothing predates 2026-04-14; vintage is ours alone) is settled by the inventory; the unanswered half becomes a number when GATE6-CLOSE-01 returns.

Ruling (Architect under Tim's written delegation, 2026-08-14): option (a), with (c)'s legal item sequenced alongside as calendar-waiting work per the drip amendment.

Anchor one holds, per companion volume: file listing, source uri, sha256, byte size, per-volume licence field (nullable until the legal review pass populates it), and release applicability range. No document bytes enter the backplane. The schema reserves the document store so the bytes have a home the day legal clears.

Why the Architect held (a) over Tim's leaning to (b), recorded honestly: the filed stop condition (CVIG-LANDSCAPE-RECON-01 Finding 4) names storage as the restricted act — "no CVIG content enters any RTOpacks store until a written legal position exists" — and the amendment adds "Site licence does not launder document licence." Tim's observation that we "just stored it and didn't really do any processing" correctly describes the Prototype acquisition, which predates the stop condition and stays where it is, untouched, in the morgue. But the clean backplane is the commercial product's store; ingesting NC-licensed bytes into it before a written legal position is the act the stop condition forbids. The cost of compliance is near zero: the 562-file corpus is regenerable in ~15–25 minutes at concurrency ≤2 once legal clears (CVIG-ACQUIRE-01-CENSUS:446–448).

Conditions attached: the legal position needs an actual lawyer, not a seat's reading of licence text. The licence-field population pass may read the morgue's existing 562 files (read-only reference) rather than re-fetching. Anchor one ships four and a half datasets; the fifth arrives when legal does.

V-P2P-011 — Sequencing: RAW-FIRST. Parsing is a later, re-runnable pass, per dataset, as each parser proves.

Ruling (Tim, 2026-08-14): option (a).

Edition 1 of the backplane is raw + provenance, proven complete by register reconciliation against TGA's own facets — the instrument that exists and has run (TGA-REGISTER-RECONCILIATION pattern). Parse passes follow as separate editions, per dataset, gated on each parser's own proof:

  • Quals — QUAL-PARSE-01 is proven at the highest bar in the estate; its parse pass runs immediately after edition 1, re-resolved against the fresh raw, which also closes its one caveat ("against our mirror only — TGA was never checked").
  • Units — after PARSER-FIX closes (the first-criterion-after-element-heading defect) and the NRT-PARSE-02 corpus launch clears its three fix rounds. Anchor one does not wait for either.
  • Packaging semantics (0116), RTOs, CVIG extraction — no proven parser exists; each is a future parse pass against stored raw, written by doctrine 7 against the shape its consuming tile demands.

This resolves the doctrine-7 tension the TM named as this design's first question: the raw capture is the mirror (doctrine 2 — the pipeline is the asset, reconstructibility demonstrated at birth), while doctrine 7's demand-driven schema governs the parsed shapes tiles consume. Capture is dataset-shaped — TGA's shape, not ours. Parse passes are demand-driven. The clean-room brief's own words carry forward as the operating rule: "if we parse on the way in we bake today's mistake into tomorrow's database."

V-P2P-012 — North Star amendments approved in principle (canon edits, five gates at filing).

Ruling (Tim, 2026-08-14): amend both.

  1. §1 — "cross-model triangulation" is replaced by the verification vocabulary the record actually contains and paid for: dual-instrument agreement, second-witness cross-column comparison, prediction-then-measure against frozen baselines, guard polarity proven offline before the first call, independent tripwires that fire on parses wearing a success flag, digest re-hashing at every gate. (Driver F-12/§2.1; the existing claim returns zero hits across the 17-document ingestion set.)
  2. §5 — "through the proven parsers" is reworded to match V-P2P-011: raw captured first; parsing as separate re-runnable passes through parsers proven per dataset, named as such.

Both amendments are dated and go through the normal five-gate canon path at the filing pass.

Findings dispositions (F-01 → F-12)

  • F-01 — RESOLVED, with the Driver's caution intact. The ~3,985 figure is sourced: it is Tim's ruling of 2026-08-14, recorded in where-we-are.md (project files) with the full breakdown — 3,887 current + 76 re-registration pending + 22 suspended — from Tim's own read of the live register. The Driver's search population was the repo, where where-we-are.md is not yet filed; the miss is legitimate and validates the Driver's own method note about populations. Filing-pass action: the figure receives a dated attribution line in the repo (Tim, live TGA register, 2026-08-14) so the North Star's citation resolves. The caution stands: our own instruments cannot reproduce a live org count (/search/organisation returns a ~240 shuffled pool); the figure is a founder observation, dated, not a pipeline measurement.
  • F-02 — ADOPTED into the design. Every dataset's population statement is written before the first call, from TGA's own facets at run time, and reconciled after (design §6).
  • F-03 — DECIDED in the design (design §5): whole register at level 1; levels 2–3 for Current components in edition 1; historical-release content as a follow-on pass. Reasoning recorded there.
  • F-04 — ADOPTED as hard runner constraints (design §4): partition by Type/Id × Status/Id, orderby=Code asc mandatory, totalCount never count, boundary-key progress guard.
  • F-05 — ADOPTED: CVIG metadata enumeration walks all historical releases (574 pairs), never latest-only, never the zip endpoint.
  • F-06 — NOTED, out of scope for anchor one. Delivery-notification history is regenerable and disposable under doctrine 1; the volume precedent (1.29M rows) is carried as sizing evidence.
  • F-07 — ADOPTED: data-sources.md is not a source for anchor one. Its correction or retirement is a filing-pass item, not this design's.
  • F-08 — ACTIONED as GATE6-CLOSE-01 (see V-P2P-009).
  • F-09 — ADOPTED: the runner is Workers/Node fetch; shell curl and Python urllib are in the forbidden set (design §4 guards).
  • F-10 — ADOPTED: sizing is against our own write path, not TGA's rate. The new pipeline's write path is measured before the corpus run (design §6).
  • F-11 — ADOPTED verbatim: nrt-atom-schema.md §6 build gates + G21 + G23 + the boundary-key guard are lifted into the design's verification section and will be lifted into the build briefs.
  • F-12 — ACTIONED at V-P2P-012.

Open item carried from the Driver's least-sure line: the possibility of an organisation-side reconciliation filed under an unmatched name. The cheapest check — asking Alex whether any document reconciles tga_organisations against a live TGA organisation facet — rides with the GATE6-CLOSE-01 relay as a one-line question, costing nothing.

Least sure, and what would make it wrong: V-P2P-010's reading that the stop condition binds the clean backplane. If Tim's legal advice comes back that ingest-and-derive without redistribution is clearly permitted for NC-licensed material, then (a) cost us one edition cycle of delay on CVIG bytes for nothing — an acceptable price for not crossing a filed rule on a seat's own licence reading. Everything else in this document is a ratified ruling or a disposition of a sourced finding.

— Architect seat (Fable), 2026-08-14.