Build Log

Transcript Backfill — From 1 Real Session to 28

July 18, 2026

Ran the Crucible on the backfill approach, corrected two Council-flagged risks structurally (a second disambiguation signal beyond date, and a separate always-correctable mapping file instead of mutating sealed session JSONs), then built the matcher and found three real bugs in its own dry run before applying anything.

ForgeKit OSArc IntelligenceTranscript BackfillData Coverage

sessions newly matched to real transcripts (up from 1 pre-existing)

131

sessions correctly flagged ambiguous rather than guessed

25/3/0

of the 28 matches: complete / malformed-1-line / broken-chain source integrity, independently verified

Timeline

Start

Charge received: expand transcript coverage

Zeb: 'let's charge another session and see if we can go through a lot more of these sessions for data'

+10m

Real data surveyed

196 real local transcripts found spanning 2026-06-16 to 2026-07-18; confirmed no existing slug-to-uuid mapping exists anywhere

+20m

Scoping question asked and answered

Zeb: build the full matcher, backfill as many as reliably possible — not a small proof batch

+25m

Crucible checkpoint asked and run

Zeb confirmed this warranted Council review before mechanically editing 100+ sealed records

+40m

Council converges on two corrections

Same-day alone is insufficient; never mutate sealed session JSONs directly. Zeb: 'agreed on both counts'

+70m

Matcher built with both corrections structural

classifyFromScoredCandidates requires a content/duration signal beyond date; applyMatches writes only to a separate mapping file

+75m

First dry run appears to hang

Background bash task exceeded 120s with no output; killed and diagnosed rather than waited on indefinitely

+85m

Root cause found: quadratic file re-reads

loadRawSnippet was re-opening the same transcript file from disk on every one of up to 120 calls per candidate — fixed to read each file once

+95m

Second real run: 19 seconds, clean results

36 high_confidence, 123 ambiguous, 2 no_candidate — but two uuid collisions spotted by eye in the output

+100m

Collision detection built and tested

Two pairs of sessions (forgehome-intelligence/forgehome-recipe-book, forgekit-os-history-epilogue/forgekit-os-history-narrated-deck) both correctly downgraded to ambiguous

+110m

Third bug found: _TEMPLATE.json swept in as a fake session

no_candidate list included 'session-name-YYYY-MM-DD' — the schema template file's placeholder slug/date, caught before it could contaminate results

+115m

Final dry run: 28 clean, non-colliding matches

161 real sessions, 28 high_confidence, 131 ambiguous, 1 no_candidate, 1 already_linked

+120m

Results presented, applied to the mapping file

Zeb reviewed and approved; 28 matches written to a separate, never-sealed-JSON-touching file

+130m

Source integrity independently verified for all 28

25 complete, 3 with exactly 1 malformed line each (real but minor transcript artifacts), 0 broken chains, 0 errors — wired into the matcher itself as a --verify step, not a one-off manual check

What shipped

forgekit-os/scripts/arc/match-transcripts.mjs — correlates session JSONs to real transcripts via same-day date (candidate filter) + content-keyword overlap + duration plausibility (mandatory disambiguation signals), with structural collision detection and _TEMPLATE.json exclusion

classifyFromScoredCandidates — pure decision function proving same-day alone can never produce high_confidence, and multiple equally-strong candidates are never silently disambiguated

resolveCollisions — downgrades BOTH sessions to ambiguous when two independently claim the same transcript uuid, never picks a winner

verifyMapping / --verify CLI flag — checks every mapped transcript's real sourceIntegrity via the proven loadClaudeCodeTranscript parser and persists the result back into the mapping file

forgekit-os/arc-interpretations/transcript-match-map.json — the separate, always-correctable mapping file (never the sealed session JSONs) recording 28 real matches with content scores, matched keywords, and verified source integrity

npm run arc:match — CLI command (dry run by default, --apply to write, --verify to re-check integrity)

20 new tests covering the pure matching/collision/apply/verify logic — all passing (108/108 total across npm run test:os)

“agreed on both counts. we can move”
— Zeb, confirming both Council corrections (second disambiguation signal, separate mapping file) before the matcher was built