Skip to content

YNM 1.4.0 — Bounded Evidence Review Research Release

Latest

Choose a tag to compare

@RobLe3 RobLe3 released this 20 Aug 13:30
· 2 commits to main since this release

YNM 1.4.0 — Bounded Evidence Review Research Release

YNM 1.4.0 is the current public research implementation of YES / NO / MAYBE, a bounded evidence-backed project review methodology.

YNM turns broad review conclusions into explicit propositions, connects them to traceable evidence, preserves unresolved questions as MAYBE, records reviewed and unreviewed scope, and separates review conclusions from authority to act.

What 1.4 adds

  • adaptive YNM-0 through YNM-3 execution
  • stronger affirmative-evidence requirements
  • bounded negative claims and proposition decomposition
  • evidence sufficiency and operational convergence boundaries
  • explicit validity boundaries and YNM-BRP-1
  • improved evidence and authority containment
  • optional provider-neutral acceleration interfaces

Measured behavior

On the controlled primary-executor fresh holdout, YNM 1.4 PORTABLE achieved 100% material recall, 0.9722 precision, zero unsupported claims, 100% required-MAYBE recall, complete evidence traceability, and zero authority violations.

These results apply to the tested executor and fixtures. They are not universal guarantees.

Important limitations

This is a Research / Experimental release.

  • Execution remained expensive: approximately 2.34× input tokens, 2.82× output tokens, 2.40× elapsed time, and 2.90× tool calls relative to control in the primary holdout.
  • Equivalent improvement did not safely reproduce on every tested executor.
  • YNM-BRP-1 missed one frozen uncertainty-preservation criterion.
  • Automatic activation remains insufficiently observable.
  • Ruflo/RuVector acceleration effectiveness was not established.
  • Independent interoperability remains MAYBE.

YNM should be used as bounded, auditable decision support, not as a universal correctness oracle, certification system, or autonomous publication authority. All frozen adverse findings remain unchanged.

Package integrity

Deterministic runtime content signature:

0552f68ea5129c0a1f4decc41a50fdef60b50a0b0c1f28cac9e2c85b9e3754af

The generated runtime package contains 62 files. Evaluation fixtures, adjudication outputs, historical state, and research-only tooling remain source-repository evidence and are excluded from the installed package.