Distillation

🟡 (18:24) /home/byk/Code/institutional-transition-lab/scripts/build-reports.mjs imports readdirSync and rmSync from node:fs, spawnSync from node:child_process, and fileURLToPath from node:url.
🟡 (18:24) scripts/build-reports.mjs derives repositoryRoot from ../, docsDirectory from ../docs/, cacheDirectory from ../docs/.observable/cache, and notebookCli from ../node_modules/@observablehq/notebook-kit/dist/bin/notebooks.js, all relative to import.meta.url.
🟡 (18:24) Report-build cache handling in scripts/build-reports.mjs: Observable interpreter results are keyed by cell source rather than files read inside a cell, which can leave reports stale when a pipeline refreshes a CSV without changing notebook source. The implemented fix removes only docs/.observable/cache using rmSync(cacheDirectory, {recursive: true, force: true}) before building.
🟡 (18:24) scripts/build-reports.mjs discovers reports by reading docsDirectory, retaining filenames ending in .html, sorting them, and mapping each to docs/${name}.
🟡 (18:24) scripts/build-reports.mjs invokes the Observable Notebook Kit CLI synchronously through process.execPath with arguments [notebookCli, "build", "--root", "docs", "--empty", "--", ...reportFiles] and options {cwd: repositoryRoot, stdio: "inherit"}. It throws result.error when present and otherwise exits with result.status ?? 1.
🟡 (18:24) /home/byk/Code/institutional-transition-lab/research/oss-governance-enriched-audit-envelope-v1.md is titled Enriched OSS governance audit envelope v1; its status is development safety mechanism; not holdout evidence, and its run date is 2026-08-27. (meaning August 27, 2026)
🟡 (18:24) The enriched OSS governance audit design uses a V2 source corpus of 225 bounded GitHub issue and pull-request bodies. Luna-medium compresses the corpus, but an earlier audit found that model confidence and grounded quotations do not establish whether a proposal merged or whether claimed rights match the final patch.
🟡 (18:24) The fail-safe envelope is defined as the deterministic union of every result from frozen governance, steering committee, or veto queries and every schema-valid Luna institutional_event or abstain response.
🟡 (18:24) For pull requests, the enrichment workflow caches current state, merged_at, merge commit, base and head SHAs, the complete changed-file listing, and up to 20 relevant files / 50,000 patch characters. Governance-like filenames remain visible even when GitHub omits their patch; all bounds and missing evidence are explicit.
🟡 (18:24) Enriched audit run references: GitHub Actions workflow 33053908119 at https://github.com/BYK/institutional-transition-lab/actions/runs/33053908119; implementation commit 3526b683c16fb130545034c3dbb795d0b4fca6c8 (short form 3526b68) at https://github.com/BYK/institutional-transition-lab/commit/3526b683c16fb130545034c3dbb795d0b4fca6c8; artifact 9639245355.
🟡 (18:24) The audit retrieval produced 434 unique candidates and the frozen 225-document triage corpus. Luna returned 222 schema-valid responses; 3 responses were quarantined, leaving 7 validated institutional flags and a 12-document institutional/abstain queue. Grounding among validated evidence spans was 100%.
🟡 (18:24) The fail-safe union contains exactly 40 records, partitioned as: 1. Narrow query only — 28; 2. Model queue only — 8; 3. Both — 4.
🟡 (18:24) Of the 40 enriched records, 14 are issues and 26 are pull requests. Of those PRs, 23 merged and 3 did not. Every PR file listing was complete relative to GitHub’s reported changed_files count.
🟡 (18:24) The enriched bundle retained 110 governance-relevant file records; 4 PRs had no relevant file under the frozen filename/patch vocabulary.
🟡 (18:24) No governance-relevant file had an unavailable patch in the audit run. Three individual patches reached the 50,000-character aggregate bound.
🟡 (18:24) Five PRs hit either the 20-file or 50,000-character selection bound: 1. OpenTofu #2830; 2. OpenTofu #2953; 3. Terraform #33661; 4. Terraform #34847; 5. Valkey #1788. Truncation is recorded per PR and per file, requests expanded review, and is not treated as evidence that omitted content is irrelevant.
🟡 (18:24) Safety-net check: Luna again labeled Valkey #345 as no_event, but the record entered the envelope through narrow_query, was confirmed merged, and included 3 relevant files. The deterministic union therefore repaired the known model false negative without tuning the model or consulting performance outcomes.
🟡 (18:24) OpenTofu #2830, Terraform #33661, Valkey #1390, and Valkey #2927 remain in the enriched envelope through the model queue. All except Terraform #33661 also entered through narrow queries. Their merged state and patch evidence can now be reviewed against body-only predictions.
🟡 (18:24) Interpretation boundary: the run validates enrichment mechanics and the known-case safety net but does not estimate population recall. The retrieval vocabulary, ranker, union, and patch rules are model-informed development choices.
🟡 (18:24) The prescribed next step is a blinded coding pass over all 40 enriched records. The coder must see source evidence but not project performance, fitted phases, or transition dates.
🟡 (18:24) The blinded coding pass should record: 1. proposal versus effective state; 2. body/patch conflicts; 3. canonical actors; 4. rights; 5. targets; 6. directions; 7. explicit abstentions.
🟡 (18:24) Records with unavailable or truncated evidence require expanded source retrieval before adjudication, and only adjudicated graph diffs may enter the event ledger.

Distillation

Observations