Skip to content

v2.5.0 — scene-fit selection evaluation + anti-overfit signal

Latest

Choose a tag to compare

@benelser benelser released this 25 May 03:06
· 275 commits to main since this release

v2.5.0 — scene-fit selection evaluation

The grammar reached 29 scene types in v2.4.0. v2.5.0 adds the missing
piece: a CLI handle for the author-agent to query "given this subject,
which scene types fit?" + a hermetic test that proves the recommendation
doesn't default to the structure/compare/tension rut.

Added

  • docent scene-fit list [--json] — every scene type grouped by cognitive
    cluster, with one-line "reach for this when..." guidance per type.
  • docent scene-fit recommend <subject-id> [--json] [--top N] — rules-based
    mapper from survey signals (anchor discipline, load-bearing 5%, argument
    shape) to recommended scene types with rationale citing the matched needles.
  • docent hermetic-scene-fit — 10-fixture harness covering every cognitive
    cluster + 1 anti-overfit probe. All green or none.

Added — anti-overfit signal

scene-fit recommend raises warningOnDefault: true when the recommended
body collapses to the default rut (no signal-specific primitives — only
structure / compare / tension). The flag invites the author to
revisit the survey for a finding that should have surfaced a more specific
primitive.

Required — scene-set commitment

survey-template.md and survey-explainer.md gain mandatory § 2.6
"Scene-set commitment". The agent runs docent scene-fit recommend <id>
and commits to a scene set with rationale tying back to a survey finding.
Mirrors the v2.1.2 § 2.5 style-commitment contract.

Skill cascades (docent-pr, docent-ar, docent-explain) get step 3b for
the same purpose.

Verified

  • 11/11 scene-fit fixtures green (10 clusters + 1 anti-overfit probe)
  • 3/3 style hermetic unchanged
  • 4/4 gallery hermetic unchanged
  • tsc clean