Skip to content

v1.1.0: visible decision citations + mandatory red-team compliance loop

Latest

Choose a tag to compare

@OC-NeuralSense OC-NeuralSense released this 30 Jul 00:57
· 1 commit to main since this release

Two additions on top of v1.0.0, both about trust and coverage: seeing why the system made a decision, and closing a real gap in what it actually checks.

Visible decision citations

Every skill, agent, and gate now attaches a decision_record to its output by default: the exact methodology/<file>.md#<section> that justified each finding, the checks that ran, any rule that was set aside and why, and a status. This is shown by default, not withheld until asked. It cites this project's own independently written methodology only, never the two private source books; see docs/book-grounding.md for the exact boundary.

Mandatory full-checklist red-team compliance loop

The existing three-lens review (structure, prose, soundness_and_reader_fit) left real, measurable gaps against the full methodology/checklists.md index: house style, most word choice, all length and rhythm, half of sentence-internal grammar, section apparatus, and usage judgment were never checked.

A new agent, red-team-reviewer, now runs after every skill, agent, and workflow stage at standard depth and above. It checks the complete ten-level checklist and audits the producing component's own coverage claim rather than trusting it. A new gate, GATE-COMPLIANCE, and a bounded loop (orchestration/policies/red-team-policy.yaml, max_iterations: 5) route findings back to their owning stage and re-check; if issues remain after the cap, the loop stops and reports every remaining finding honestly rather than a silent pass. GATE-RELEASE now requires GATE-COMPLIANCE green. quick depth stays exempt and explicitly non-exhaustive, exactly as before.

Like every other component, red-team-reviewer never reads or references the private source books.

Verification

npm run validate (16 components), npm test (121/121), npm run build, npm run release:check (6/6 gates), plus a manual walkthrough of the new instructions on a planted-defect sample passage confirming red-team-reviewer catches exactly the class of issue (grammar mechanics, house-style tells) the old three-lens review misses.