canon: Discernment Transfer Ladder method + rulebook-transfer evidence (COO probe)#264
Conversation
…dence (COO probe) Adds canon/methods/discernment-transfer-ladder.md (target_repo: outcomes-driven-development) operationalizing rulebook-transfer; appends a bounded second-probe evidence block to rulebook-transfer (down-tier execution to Haiku at ship-grade; first data on the chain arm of star-vs-chain). Proposal; ratify the exact text.
Canon Quality — Frontmatter Schema ✅All 49 file(s) in Validator: |
Canon Quality — Homepage Surfacing ✅49 essay(s) scanned. Soft report — never blocks; the hard field gate is the Frontmatter Schema job. All published essays resolve to the homepage feed. Report: |
Canon Quality — P0010 Retrieval-Readiness
|
Canon Quality —
|
Two portable-governance items from the 2026-06-30 COO distillation session. Proposal —
models-do-not-mutate-canon; ratify the exact text.Adds
canon/methods/discernment-transfer-ladder.md— the empirical method that operationalizesrulebook-transfer(reference + held-out truth → measure on two axes → write the missing discernment down → re-run + regression-guard the reference → train/test on unseen cases → stop when a capability won't transfer). Frontmatter routestarget_repo: outcomes-driven-development.Edits
canon/principles/rulebook-transfer.md— appends a bounded second probe to the Evidence section: down-tier execution to Haiku reaches ship-grade under an explicit lens (articulability ladder observed v1→v5); first data on the chain arm of star-vs-chain (mid-tier stewardship transfers but is lossy; frontier authoring is surgical). Device tier still untested.Full eval evidence: kirigami-mcp PR (cross-model fidelity run).
Note
Low Risk
Documentation-only canon changes with no runtime or auth code paths; main effect is governance reasoning and how tier delegation is documented.
Overview
Adds
canon/methods/discernment-transfer-ladder.mdas a provisional Tier-2 method that turnsrulebook-transferinto a measured loop (reference vs held-out truth, structural + semantic gaps, rulebook compensation, regression on the reference tier, unseen-case validation) and argues for a shallow-star authoring topology driven by human stewardship time. It also spells out how evaluation/ratification can move down-tier only as delegated, bounded stewardship aligned withmodels-do-not-mutate-canon.Updates
canon/principles/rulebook-transfer.mdby appending a second probe (COO canon, 2026-06-30) to the Evidence section: stronger down-tier execution when lenses are explicit (Haiku v1→v2 compensation), plus first chain-arm data on stewardship (mid-tier fixes are lossy vs frontier authoring), with device tier and full star-vs-chain resolution still open.Reviewed by Cursor Bugbot for commit 00009ab. Bugbot is set up for automated code reviews on this repo. Configure here.