Skip to content

PR8A: Signed planner benchmark evidence tooling - #36

Merged
min9lin9 merged 30 commits into
boulder-native-planner/pr7-preview-integrationfrom
boulder-native-planner/pr8a-benchmark-tooling
Aug 26, 2026
Merged

min9lin9 merged 30 commits into
boulder-native-planner/pr7-preview-integrationfrom
boulder-native-planner/pr8a-benchmark-tooling

Conversation

@min9lin9

Copy link
Copy Markdown
Owner

Summary

  • plan benchmark with explicit trust-root/study-root inputs, Ed25519 provenance verification, descriptor-pinned no-follow evidence reads
  • Canonical study identity and report recomputation; strict common planner-output normalizer with independently verified source trust
  • Deterministic HOLD-first promotion decisions; repository never contacts providers or executes study runs
  • Recorded v0.2 PR8B field-study outcome: HOLD (36 scored / 27 eligible / 7 execution failures / 3 critical caps)

Merge gate

  • bun test: 403 pass / 3527 assertions
  • bunx tsc --noEmit
  • 222-file pack dry-run

Stacked on PR7. This completes the RFC v0.2 implementation program.

Adds plan benchmark with explicit trust-root/study-root inputs, Ed25519
provenance verification, descriptor-pinned no-follow evidence reads,
canonical study identity and report recomputation, the strict common
planner-output normalizer with independently verified source trust, and
deterministic HOLD-first promotion decisions. Adds the plan benchmark
subcommand, fixtures, and benchmark plan documentation. The repository
never contacts providers or executes study runs; PR8B operational
evidence stays external, and the recorded v0.2 field-study outcome is
HOLD (user action request included).

Merge gate: bun test (403 pass), bunx tsc --noEmit, 222-file pack.
Comment thread test/planner-output-normalizer.test.ts Fixed
Comment thread test/planner-output-normalizer.test.ts Fixed
Comment thread test/planner-output-normalizer.test.ts Fixed
Burt and others added 28 commits July 19, 2026 01:44
…ulder-native-planner/pr8a-benchmark-tooling
Self-contained src/v2 kernel (JCS digests, effect gating, authority
verifier, lifecycle) with boulder v2 execute CLI, frozen v2-kernel
fixtures, and authority-vector corpus. ADR 0003 clause on root
AGENTS.md is scoped from byte-freeze to gate-outcome claims, with
counterevidence recorded (committed planner-stack docs predated it).

Refs: docs/adr/0003-v2-kernel-gates.md (K1 bounded; K2-K4 separate gates)
Byte-first reader (64 KiB limit, BOM/UTF-8/duplicate-key rejection
before JSON.parse), canonical digests, and exact-issue validation for
the frozen k2a-f contract foundation (fixtures/k2a-f v1).
… remediation, executor evidence

Adds the pre-execution safety gate (signed receipts), scope-attribution
receipts, score workflow state machine (blinding/lock/reveal), study
remediation policy wired additively into planner-benchmark, and the
common executor lifecycle/receipt evidence schema. Optional contract
fields only; existing bundles unchanged.
Isolated sandbox runner, independent byte-level oracle, and capture
pipeline producing the k0r evidence manifest. Evidence artifacts under
evidence/k0r/ are regenerated against the settled tree in a follow-up
commit.
Root and subdir guides updated for the planner stack, v2 kernel, k2a-f,
k0r harness, and ops router; new scoped guides for fixtures/ and
evidence/. Documents the tree as it exists per amended ADR 0003 clause
(no gate-outcome claims).
Adds reference/Boulder_Reference_Implementation_Strategy v0.2 (read-only
reference) and README --dry-run corrections. Moves the ReFoundation
planning zip from docs/ to plans/ so it stays out of the npm package
surface; doc-registry entry removed. The two Korean session/gap docs
stay untracked: they are gitignored local-only notes by design.
… tree

Reclassifies packed files after the v2/k2a-f/planner landings (runtime
107, fixture 44, public-doc 66; totals 249/250), drops the quarantined
ReFoundation zip from the packaged set, classifies the three new
AGENTS.md docs consistent with existing shipped AGENTS.md files, and
refreshes the readiness-v0 gate baselines from live evaluations.
Q3 conditions met: Q2 dispositions applied (8 commits), v1 deltas classified additive/expected, ADR 0003 clause amended with counterevidence, baselines regenerated, suite green except seven k0r evidence-contract tests whose recapture is gated on the ADR 0004 human exit flow (evidence/k0r/ held untracked as needs-evidence). 10732cb remains the historical AS-IS reference; Boulder-9-3-plus stays the read-only comparison control for that revision.
Traces every strategy section 9 fitness function to an enforced test
suite or a namespaced planned REF-E/REF-PR experiment so prose cannot
silently become a claim of enforcement. New executable guards: no
runtime-host package imports anywhere in src/, k2a-f sibling-only .js
imports, plan/planner cohort explicit-.js specifiers, no process.exit,
zero runtime dependencies, v2 K2-K4 anti-overclaim, and doctor probe
non-mutation. Also fixes the one pre-existing cohort violation the new
guard caught (plan-receipts type import missing .js).
- Replace boulder guide with ko html version bound to evidence harness
- Add k0r baseline generator, canonical helpers, issue-exit, reconcile
- Regenerate package inventory and readiness baselines for landed tree
- Make namespace-hostile tests pass under production bwrap CI: route
  package dry-run via writable staging, guard fresh-user K0R contract
  tests behind BOULDER_CI_BWRAP, adapt nested-bwrap probe
…ue path

Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
…guard

Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
Ultraworked with [omo](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: sisyphus-dev-ai <sisyphus-dev-ai@users.noreply.github.com>
@min9lin9
min9lin9 merged commit 7c117ab into boulder-native-planner/pr7-preview-integration Aug 26, 2026
2 of 3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants