Distillation

Date: Sep 8, 2026

• 🟡 (18:20) Downloaded ZIP verified at exactly 423046 bytes with SHA-256 1f6be91fcd98e43e99d82d5f274536f6c23616e7e6c711e26324a2cc4b06ef06. Archive members and uncompressed sizes: documents.json 445940; luna-a.json 56076; luna-b.json 52967; adjudication-package.json 149490; preliminary-package.json 107543; preliminary-summary.json 707; preliminary-summary.md 543; summary.json 703; summary.md 542; terra-route.json 391727; terra.json 43744; v1.2-adjudication-package.json 156038; v1.2-preliminary-package.json 109772; v1.2-preliminary-summary.json 788; v1.2-preliminary-summary.md 618; v1.2-summary.json 784; v1.2-summary.md 617; v1.2-terra-route.json 425012; v1.2-terra.json 50036.
• 🟡 (18:21) Archive-member SHA-256 hashes: documents.json=2a62ab05687982c112a29f6126227c26c3ff8defc89ca64662753aaf85c3c0ce; luna-a.json=ee86739a7585fd9759500489e00af3e8bddae9ec3908890c86950a29fdee3d36; luna-b.json=ed7db9330b84f933a88fa567c759273054e4cf40b15330ecf09a4ef75ad5d2e8; adjudication-package.json=a413e78efae5ac26ade2418aab291ca1d4c26d5e6d23c4291d1934c5cef5228b; preliminary-package.json=27e5ff6523c143eeeecf925b2909caffc1a600b30ec35f0be42e2a20248917fb; preliminary-summary.json=9270803dce6ed5c729d87428875c98846a183222226db315a236b1930f631bff; preliminary-summary.md=d132a6ddf44923d576e2ef556bc66dfc5b02621c462c8bf0a7f678a82a106fce; summary.json=00bff3e365dd1a2c9a7151c88d085460a1b4f5685fdc345b4501e186825ef5eb; summary.md=4335755ee9cf9ddb342ddeb897d24f0468b3e6cf249e597e7096a859c3d5dd93; terra-route.json=f74eefa18e435dd08dd6c550869b7b547e8d29301ffebe35c4f4586a0818d4de; terra.json=7377b3d28f1cbae141cfea03e942d01cbc6688608e4a87e88f72d79d9de70cb1.
• 🟡 (18:21) Versioned archive-member SHA-256 hashes: v1.2-adjudication-package.json=14644759f53071adfe1173d4e25bb01931b8c1591fa8a48295dbd0556a754769; v1.2-preliminary-package.json=0b7a142526ff54b02e673bfc35a5e6a2410ad7fb5e11ac56ba5113da3a90e6a7; v1.2-preliminary-summary.json=23f9c8847b7e3217de043777d5efeda61c4c75cacaf5f73e71b9a0b186b5e1b2; v1.2-preliminary-summary.md=fd391b76d3f979eb62522fed7fea2cb3091d4eeaffc0f87aad4722263ab501d7; v1.2-summary.json=60a1dbd0142855803975ab0e914d8f5ca8553e71ebccc604dcf5bff7aad2ba49; v1.2-summary.md=346921fb0be7e7487db20a541b4962a9c5ceb09970f61f6c5cc7df45368ca155; v1.2-terra-route.json=648dc1db88c7e2d1dceb96784c8532805260a5abb0ffb3e05c554aefc764e96b; v1.2-terra.json=77a397c057577deb14c59437af31513c4db8bfbb7ec47c0fee511b5302a5937c.
• 🟡 (18:21) documents.json document bundle metadata: coding_id="oss-governance-enriched-coding-v1", coding_schema_sha256="bfc735a7d8602d1f236a88c2d23b195f0600a8bae86fb8c8936afe7883a41b8e", design_status="performance_blinded_model_informed_development", model_input_fields=["entity_id","publisher","published_on","source_type","sources","evidence_bounds"], outcome_data_used=False, schema_version=1, and transition_dates_excluded=True. Raw SHA-256 was 2a62ab05687982c112a29f6126227c26c3ff8defc89ca64662753aaf85c3c0ce; canonical JSON SHA-256 was d70f7f795847968339a645b6dbd6d4428af6bd6044de44fa2602d4c4343b0ee9; record count was exactly 40, all 40 unique.
• 🟡 (18:21) The 40-record governance corpus comprised exactly 18 OpenTofu records (opentofu-github-issue-{1353,2109,2573,258,340} and opentofu-github-pr-{1010,1107,1152,1650,2830,2953,2959,3473,3559,4018,727,823,990}), 12 Terraform records (terraform-github-issue-{22468,34139,35563} and terraform-github-pr-{21175,21345,22332,22745,33661,34096,34103,34847,38385}), and 10 Valkey records (valkey-github-issue-{2961,3289,3441,4276,4508,4509} and valkey-github-pr-{1390,1788,2927,345}), sourced from the corresponding GitHub issue or pull-request URLs.
• 🟡 (18:21) Corpus evidence bounds showed source_text_truncated=True only for valkey-github-issue-4276, whose captured GitHub body was exactly 12000 characters. patch_selection_truncated=True applied to opentofu-github-pr-2830, opentofu-github-pr-2953, terraform-github-pr-33661, terraform-github-pr-34847, and valkey-github-pr-1788. All records had patch_unavailable_count=0; all other displayed records had source_text_truncated=False.
• 🟡 (18:21) Notable patch evidence in the corpus included OpenTofu governance files CHARTER.md, GOVERNANCE.md, CONTRIBUTING.md, contributing/FAQ.md, multiple TSC/*_NOTES.md, TSC_SUMMARY.md, rfc/20251021-databricks-backend.md, .github/workflows/release.yml, version/official.go, and .github/scripts/compare-release-versions.sh; Terraform included LICENSE, module-source/getter files under internal/addrs/, internal/configs/, and internal/getmodules/moduleaddrs/, plus AI governance instructions .amazonq/rules/governance.md, .claude/governance.md, .cursor/rules/governance.mdc, .windsurf/rules/governance.md, .clinerules, .continuerules, .github/copilot-instructions.md, .rules, AGENTS.md, CLAUDE.md, and GEMINI.md; Valkey included GOVERNANCE.md, CONTRIBUTING.md, MAINTAINERS.md, COPYING, and numerous src/*.c files.
• 🟡 (18:21) v1.2-preliminary-summary.json reported: record_count=40, adjudication_count=35, canonical_ledger_records=0, full_response_agreement_count=0, provisional_core_agreement_count=5, luna_a_invalid_count=5, luna_b_invalid_count=2, luna_class_agreement=0.675, luna_strict_edge_agreement=0.55, terra_completed=False, and terra_invalid_count=None. Route-reason counts were abstention=4, body_patch_relation_disagreement=11, bounded_evidence=6, class_disagreement=13, edge_disagreement=18, event_field_disagreement=13, invalid_response=6, and low_confidence=5.
• 🟡 (18:21) v1.2-summary.json preserved the preliminary counts and agreement metrics but recorded terra_completed=True and terra_invalid_count=1. Package top-level keys were records, schema_version, and summary; record keys were canonical_ledger_eligible, full_response_agreement, luna_a, luna_b, provisional_core_agreement, record_id, route_reasons, source_url, and terra_advisory. There were exactly 40 unique records and 35 routed records.
• 🟡 (18:21) Repository search in /home/byk/Code/institutional-transition-lab found more than 100 matches for provenance, blinding, grounding, and canonicalization concepts. Key safeguards include src/institution_lab/source_retrieval.py rejecting forbidden events, transitions, transition_dates, and performance_outcomes, requiring transition_dates_excluded is True and outcome_data_used is False; src/institution_lab/source_enrichment.py enforcing the same frozen-manifest conditions; and scripts/copilot-code-governance.mjs rejecting bundles that are not performance blinded.
• 🟡 (18:21) Model authorization is explicitly limited: src/institution_lab/model_policy.py, scripts/copilot-extract-events.mjs, tests/test_model_policy.py, and cases/model-selection/event-triage-v1.json require exclusion of canonical_power_graph_serialization. The frozen model-selection contract identifies artifact ID 9652609349 with SHA-256 44ea0132eb7416e2c4ad530e66b4e5c672700fd0eb2d9db6be5e5cb196bcd2d0.
• 🟡 (18:21) /home/byk/Code/institutional-transition-lab/pyproject.toml defines project institutional-transition-lab version 0.1.0, Python >=3.12, Hatchling build backend, package src/institution_lab, and CLI entry points: institution-lab-oss, institution-lab-llm-eval, institution-lab-llm-triage, institution-lab-retrieve-oss-sources, institution-lab-enrich-oss-sources, and institution-lab-governance-coding. Optional analysis dependencies are duckdb>=1.3,<2, httpx>=0.28,<1, numpy>=2.2,<3, pandas>=2.2,<3, pyarrow>=19,<22, pydantic>=2.11,<3, pyyaml>=6,<7, ruptures>=1.1.9,<2, and scipy>=1.15,<2; lab uses jupyterlab>=4.4,<5 and jupytext>=1.17,<2; dev dependencies are pytest>=8.3,<9 and ruff>=0.11,<1. Ruff uses line length 100, target py312, and lint rules ["E","F","I","UP","B","SIM"].
• 🟡 (18:21) /home/byk/Code/institutional-transition-lab/package.json defines private ESM package institutional-transition-lab-reports version 0.1.0, requires Node >=22.12, depends on @observablehq/notebook-kit exactly 2.3.0, and provides scripts reports:preview="notebooks preview --root docs" and reports:build="node scripts/build-reports.mjs".
• 🔴 (18:21) User’s project protocol states that the target is a time-t forecast of the next performance regime based on current regime/trajectory, recent decision-right changes, cultural/operational mediators, entity and market controls, and measurement uncertainty—not an after-the-fact inclusivity score. Required outputs are P(next regime | information available at t), expected transition time or hazard curve, expected level/slope changes for each outcome, uncertainty, and historically comparable cases.
• 🔴 (18:21) User’s project protocol treats institutional state as a versioned graph: nodes are people, boards, teams, shareholder classes, foundations, or constituencies; edges are appoints, removes, funds, promotes, informs, approves, merges, or vetoes; scope may be company, segment, repository, subsystem, or decision class. An institutional event is defined as power_graph(after) - power_graph(before), separating: 1. leader changed/rights did not; 2. rights changed/leader did not; 3. both changed; 4. neither changed.
• 🔴 (18:21) User’s protocol keeps outcome semantics separate: company segments are modeled independently where possible, with company-wide results explicitly weighted; company outcomes include demand share, value-pool share, relative enterprise value, and revenue/margin. Open-source adoption, ecosystem, contributor, and delivery-health dimensions remain separate; stars and raw commit counts are weak auxiliary measures rather than primary outcomes.
• 🔴 (18:21) User’s protocol defines a staged causal chain: power-graph change -> information/culture behavior -> delivery capability -> market or ecosystem performance. Westrum-like public GitHub indicators must be labeled behavioral proxies, and public release data is only DORA-adjacent unless deployment frequency, production lead time, change-failure rate, or recovery time are actually observed.
• 🔴 (18:21) User’s transition-detection protocol requires outcome/control-only inputs and excludes leadership names, event dates, and event labels. Candidate methods are: 1. piecewise-linear segmentation; 2. PELT or binary segmentation with robust costs; 3. Bayesian change-point models returning transition-date distributions; 4. hidden Markov or switching-regression models for noisy high-frequency data. Methods and hyperparameters are frozen on development cases; cross-method agreement is reported, and disagreement yields a transition-date interval.
• 🔴 (18:21) User’s protocol requires excluding incomplete final periods; testing a declared grid of warm-up periods, minimum phase lengths, and improvement thresholds; requiring sufficient pre-transition history and post-transition follow-up; and reporting every raw candidate, eligible-specification denominator, support rate, and rejection reason. A detector setting enters a candidate’s denominator only if the date could legally be selected under its warm-up and minimum-segment constraints.
• 🔴 (18:21) User’s protocol classifies phases as growth when the lower credible slope is economically positive, decline when the upper credible slope is economically negative, and stagnation otherwise; economic significance must be predefined per metric because a statistically detectable tiny slope does not automatically count as meaningful growth.
• 🔴 (18:21) User’s event-coding protocol covers the whole study period, uses primary sources, keeps reviewers blind to later performance regimes, preserves ambiguity through confidence and alternative codings, and distinguishes announcement date, effective date, and an optional date when implementation became observable.
• 🔴 (18:21) User’s estimation sequence is ordered: 1. exploratory lagged cross-correlations and visual overlays; 2. within-case event studies estimating pre-trends and post-event level/slope changes; 3. conditional regime-transition probability/hazard models; 4. quasi-experiments using forks, reunifications, staged rollouts, and rights changes with credible comparisons; 5. hierarchical cross-case models preventing large entities from dominating by observation count. Primary outcomes, lag families, and estimators are predeclared, and sensitivity analyses are grids rather than cherry-picked cells.
• 🔴 (18:21) User’s prediction protocol evaluates out of sample by: 1. training on earlier periods and forecasting later periods; 2. leaving an entire company/project out; 3. freezing one prospective case before its next institutional transition; 4. comparing with trajectory-only, sector-only, and naive-persistence baselines. Metrics are calibration, Brier score or log loss, rare-transition precision/recall, transition-timing error, post-transition-slope error, and decision value under explicit costs.
• 🔴 (18:21) User’s first four falsifiable hypotheses are ordered: 1. power-graph diffs improve held-out forecasts more than leader-title changes alone; 2. predictive power is followed by Westrum-like behavior and delivery-health changes before external performance; 3. veto, appointment, and information-right changes predict distinct transitions after controlling for ownership concentration/current performance; 4. governance differences after shared-ancestry forks predict contributor-diversity, delivery-health, and adoption divergence beyond common prehistory.
• 🔴 (18:21) User’s project success stages are ordered: 1. reliably detect regimes before reading institutional histories; 2. improve held-out transition forecasts using power-structure data; 3. identify observable mediators that move before outcomes; 4. find repeatable, controllable interventions rather than immutable founder/ownership traits; 5. validate recommendations prospectively.
• 🔴 (18:21) User’s protocol states that only success stages 3–5 support advice about how to “switch from an extractive to an inclusive institutional configuration”; stage-2 association is predictive but is not yet an intervention recipe.
• 🟡 (18:21) /home/byk/Code/institutional-transition-lab/schema/core.sql defines 10 DuckDB tables: 1. analysis_run for run/config/code provenance; 2. entity; 3. segment; 4. source with content_sha256, authority, and archive path; 5. metric_definition; 6. observation; 7. phase; 8. transition with earliest/mode/latest dates, regimes, probability, and method; 9. institutional_event with announcement/effective dates, coding status, confidence, and reviewer_blinded_to_outcome; 10. event_source; 11. power_node; 12. power_edge; 13. event_transition_join. institutional_event requires at least one of announced_on or effective_on and constrains confidence to [0,1].
• 🟡 (18:22) /home/byk/Code/institutional-transition-lab/src/institution_lab/event_inventory.py implements outcome-independent YAML event inventories. EventInventoryRecord stores event_id, entity_id, event_kind, event_role, title, announcement/effective dates, affected scope, power change, coding status, confidence, and reviewer blinding. EventSourceRecord stores event/source IDs, publisher, publication date, evidence role, and URL. load_event_inventory() rejects duplicate event IDs and events without sources, validates dates and EventKind, constructs InstitutionalEvent, defaults confidence to 1.0, and generates source IDs as event-source: plus the first 20 hexadecimal characters of SHA-256 of the URL.
• 🟡 (18:22) /home/byk/Code/institutional-transition-lab/tests/test_event_inventory.py loads cases/projects/terraform-opentofu.events.yaml and asserts exactly 12 events, exactly 13 sources, event roles exactly {"institutional","control"}, all event reviewers not blinded to outcomes in this seed inventory, and all source URLs beginning with https://.