Distillation

🔴 (21:57) User supplied a repository search result with 100 matches, with additional matches available, under /home/byk/Code/institutional-transition-lab/research/; the displayed results were truncated.
🔴 (21:57) User supplied research/session-handover-2026-08-27.md references to commits: 58a9663b8a44cf2d7bbb89102f5c70c00cdf0d91; 79bde4d6c5cc187231c6108b8439c58cdb26b220 for cursor pagination and a regression test; 1d3af619e018f3e5050152c847cd5162219375a9 for robustness analysis, an event timeline, and a report; and 2741626670eeee678948916c7a237b1c483f88a6 for bounded retries after dropped HTTP responses.
🔴 (21:57) User supplied methodological instructions from research/session-handover-2026-08-27.md: retrieval and triage identify candidate institutional facts but do not establish them; scan the complete study period; and keep institutional coding blinded to performance outcomes.
🔴 (21:57) User supplied handover evaluation results: institutional precision 100%, institutional recall 90%; Terra/medium and Terra/xhigh were each tried three times without improving institutional recall; one strict result achieved 95.8% decision accuracy, 100% institutional precision, recall, kind, date, and minimum-evidence grounding, 3/3 Valkey recovery, and 75% strict edge recall.
🔴 (21:57) User supplied successful workflow evidence from research/session-handover-2026-08-27.md: CI run 33108057720; Luna benchmark-compatibility run 33108057688 with artifact 9661404982, digest sha256:7a6a80fcd5ff5eb0db24a955827071f827fc2fea89723fecd3cd76077eb8afc4; and full retrieval/coding run 33108057787 with artifact 9662866867, digest sha256:aa598eac82a2c4d6de072ca19e0c59291730fded4203a98d7b6a91f4ff9f59c8.
🔴 (21:57) User supplied that the full retrieval/coding run produced 224 schema-valid responses (99.56%) and four institutional responses.
🔴 (21:57) User supplied research/sol-reasoning-hard-case-protocol.md, associated with workflow run 33077982385, which asks whether gpt-5.6-sol improves institutional-event discovery or strict power-edge performance. Its inputs are cases/benchmarks/oss-institutional-events-v1.documents.json, cases/benchmarks/oss-institutional-events-v1.labels.json, and cases/benchmarks/oss-institutional-events-v1.hard-cases.json; primary metrics include decision accuracy, institutional recall, and strict edge performance.
🔴 (21:57) User supplied research/oss-governance-retrieval-development-v2.md results from workflow run 32977337091: of 16 documents newly selected by V2, Luna labeled 2 institutional, 2 control, and 12 no-event; institutional/non-institutional agreement was 202/206 (98.1%); only 2 of 8 provisional institutional flags survived audit as genuine institutional events, and one had incorrect direction and target.
🔴 (21:57) User supplied that research/oss-governance-retrieval-development-v1.md used workflow run 32975107169; its license query returned 108 results but was capped at 100; processing took 3.19 seconds per document; and the ranker was not given institutional-event labels, performance observations, or contract-validity outcomes.
🔴 (21:57) User supplied strict boundary guidance from research/oss-governance-retrieval-development-v1.md: choosing a registry design exercises an existing steering body’s authority, while normalizing license headers or adding a missing license does not necessarily constitute an institutional power change.
🔴 (21:57) User supplied research/oss-governance-enriched-audit-envelope-v1.md details: workflow 33053908119; implementation commit 3526b683c16fb130545034c3dbb795d0b4fca6c8; invalid material was quarantined, leaving 7 validated institutional flags and a 12-document institutional/abstain queue; grounding among validated evidence spans was 100%.
🔴 (21:57) User supplied research/oss-governance-coding-results-v1.2.md conclusions: product/control events are not institutional power changes; all substantive institutional candidates require source adjudication; no run supports admitting model-authored institutional edges directly into the canonical record; Luna coders both treated OpenTofu PR #727 as an effective license change, while Terraform PR #33661 was recognized as an effective license change but had coder disagreement.
🔴 (21:57) User supplied coding-run provenance from research/oss-governance-coding-results-v1.2.md: workflow runs 33086454759, 33060493337, and 33063470968; artifacts 9653837747 and 9642983578; commits 9ba7d77baec0b001649a4d455330a9fe298a05e9 and 786f33a7cb044d385f49f417c059c39299f70111.
🔴 (21:57) User supplied research/luna-terra-hard-case-comparison.md details: workflow run 32971528370; the slice contained 4 institutional events, 1 product control, 2 no-event documents, and 1 ambiguous document; no configuration produced a false institutional positive; all configurations detected the relevant transfer, Node.js/io.js reunification, and Redis license transition, but none classified the Valkey governance excerpt as institutional, instead producing no_event, abstain, or once control_event.
🔴 (21:57) User supplied that Redis was represented either as one modified-license edge or as separate added and removed license edges, with targets varying among Redis, Redis 7.4, and Redis software.
🔴 (21:57) User supplied an additional passing comparison result from workflow run 32971528362: 96% decision accuracy, 100% institutional precision, and 90% recall.
🔴 (21:57) User supplied research/luna-event-extraction-development-benchmark.md details: the benchmark asks whether a small, inexpensive reasoning model can identify institutional events; it contains 10 institutional events and 11 product-release controls; Run A used workflow 32967471044, commit fb3cfe3, artifact 9606315652; Run B used workflow 32968114572, commit db42696, artifact 9606553726; neither clean run produced a false institutional positive, and all 11 product controls were classified correctly in both runs.
🟡 (21:57) Assistant stated that the prior response had ended before the required artifact and that it was reconstructing a seven-record array from completed source audits, adding omitted source_audit data, and checking every accepted response and revised edge before returning it.
🔴 (21:57) User supplied a tool result showing that the reconstructed top-level JSON object contains exactly the keys "records", "schema_version", and "summary".

Distillation

Observations