Skill Harness: published receipts

sqlite-expert

Verdict: CANT_TELL_YET

Cut sub-reason
null
Unmeasured sub-reason
underpowered
Value class
null
Wrong instrument
false
Declared synthetic control
false

Double-ceiling NO-GO (2026-07-09/10): Null passed 14/14 epochs across two screens and a paired k=8 run; discordance d-hat=0.00 (Jeffreys 95% CI [0.00, 0.26]); pre-stated GO/NO-GO gate returned NO-GO. Full-vs-Null is structurally unmeasured on this task class — published as CANT_TELL_YET, not a fabricated no-benefit CUT.

Cost beside evidence

Cost triple
Leg Figure Detail
standing_tokensREFUSED (not_instrumented)Case study reports USD spend (~$6.17), not the standing/fired/aux token triple.
fired_tokensREFUSED (not_instrumented)Full-arm input tax noted as +28k cache-write tokens per epoch in prose; not a fired-cost triple measurement.
aux_tokensREFUSED (not_applicable)

Clause-level evidence grade

Clause evidence: UNMEASURED (no_extraction: clause evidence requires an extraction output; produce one with skill init --out <file>)

Measurements

Keys the standard defines but this receipt does not carry are listed as absent rather than filled in.

Measured values and typed refusals
Key Figure Detail
p01.0 (6/6 epochs)Two Stage-0 Null screens at 3/3 each before the paired run.
full_pass_rate1.0 (8/8 epochs)
null_pass_rate1.0 (8/8 epochs)
p_winREFUSED (underpowered)At d≲0.5 no effect is detectable inside N_max=40; Full-vs-Null is structurally UNMEASURED regardless of budget.
discordance_rate0.0x=0 discordant epochs; Jeffreys 95% CI [0.00, 0.26] — upper bound still inside the structurally-unmeasured region of the sizing table.
go_nogoNO_GO
hazard_entry_nullabsent from this receipt
hazard_entry_fullabsent from this receipt
null_completion_rateabsent from this receipt
full_completion_rateabsent from this receipt
silent_violation_rateabsent from this receipt

Evidence admissibility

Status: admissible

Paired k=8 run ingested through write-time evidence admissibility machinery; all cited numbers verified against raw .eval logs and the evidence store.

Instrument identity

Figures produced under different instrument identities are not comparable, so the generation stamp travels with the figures.

extractor_model
anthropic/claude-sonnet-5
prompt_fingerprint
claude-code-2.1.197
schema_fingerprint
2f76c933

Source of record

prose_path
docs/case-studies/double-ceiling-structurally-unmeasured.md
date
2026-07-09
notes
Pre-registered apparatus shakedown + d upper bound as NO-GO datum; not a benefit measurement. Subject skill was a thin SQLite helper (cut-candidate); 14/14 Null epochs at ceiling across screens + paired run.