[Droptops]

What

• eval/lib/sampling.js (1 line): add claude-fable-5 to shouldOmitAnthropicSamplingParams. Fable 5 rejects any explicit temperature (API 400 "temperature is deprecated for this model"), the same restriction the map already handles for Opus 4.7. Without this, run-execute-boundary.js fails 25/25 fixtures at the API layer.
• eval/public_notes/aar_v0_8_execute_boundary_fable5_readout.md: full readout of the 2026-06-10 pre-registered AAR v0.8 EXECUTE-boundary regression run against claude-fable-5 — both runner conditions, per-case strip, verbatim failure texts, SHA-256 audit anchors, deviations log, claim boundary. Raw/scored result files stay local per .gitignore policy; the memo's SHAs are the anchors.
• eval/package-lock.json: 2-line version-field sync to package.json 0.6.2 (npm regenerates this for anyone; no dependency changes).

Why

The EXECUTE-boundary case set had only been run against sonnet-4-6. This run adds the fable-5 data point under the frozen v0.8 skill ZIP (SHA verified pre-inference) with the committed scorer, unmodified.

Headline results

	skill (frozen ZIP)	script_default (CE SKILL.md)
primitiveMatch	20/22 = 90.9%	21/25 = 84.0%
B-OverEnum / hallucinatedToolIO / checklistExposed	0 / 0 / 0	2 / 0 / 3

vs sonnet-4-6 ZIP run: 82.6%, also 0/0/0. Fable-5 fixes EB-08 and EB-20, reproduces EB-14 (EX-R06-class form mismatch).

What a reviewer should know

• EB-24B is the finding that matters: fable-5 executed straight through the planted body/attachment mismatch ("Sent. … dispatched.") where STATE_BLOCKER was expected — both Type 9 pair halves ruled EXECUTE, the design's INVALID-as-design configuration. Sonnet had flagged the mismatch (failing only on form). Flagged in the memo as the priority for any r-series follow-up.
• No gate-pass claim: §10 thresholds were not pre-registered against this run; 90.9% vs the proposed ≥90% is reference-only.
• Sampling caveat: fable-5 cannot be temperature-pinned, so the run is model-managed sampling; the runner's recorded temperature: 0 field reflects the script variable, not the wire (pre-existing quirk shared with the Opus 4.7 path, disclosed in the memo).
• Scorer parity: no failure in either condition was a blocker-lexicon gap; the memo classifies each failure (form-mismatch vs genuine miss vs truncation-confounded). scorerParity.js, the cases file, and all regexes are untouched.
• One aborted all-error attempt (pre-fix) is kept as an error record with its own SHA, per protocol.

🤖 Generated with Claude Code