# Reality Lab: three controlled AI workflow tests

Prepared for owner review, 11 September 2026 (Brisbane). Not approved for public release. Open index.html for the readable report.

This package contains 34 unique synthetic cases, exact model prompts and reference answers, all nine attempt records, reports, frozen scorers and a dated protocol amendment. Eight attempts returned completed output. Their 92 repeated case checks matched the authored references; the failed initial quote attempt remains in the denominator of pipeline reliability (92/102 attempted case checks). These are not 102 independent customer documents.

Provider-reported completion costs total US$0.02957075. A further US$0.10 remains reserved with unconfirmed billing. Model-request times exclude human setup, checking, corrections and delivery. No human baseline, independently validated answer key, measured delivery cost, customer sale or time-saving claim is supplied. The quote study did not meet its original three-successful-run criterion.

## Reproduce without buying API calls

Unzip this folder. With Node.js 24 installed, open a terminal here and run:

    node verify-evidence.mjs

No package installation, API key or network call is required. The command validates the recorded hashes and reruns every score, including the failed attempt. To inspect a single run:

    node scripts/score-workflow-trial.mjs quote-accuracy experiments/results/2026-09-11/c010d172-4207-41dc-bfa2-55912c13a98f.json

Do not show expected.json to someone completing a blind baseline. The answer keys are included for independent audit; if already viewed, use a separately preregistered new fixture set for a blind comparison.

## Provenance and redaction

The source records were exported from the private dashboard database. The package removes the internal owner field and preserves the exact prompt, model output, settings, provider/generation metadata, timings and costs. PROVENANCE.json records the source-export and packaged hashes. SHA256SUMS.json covers all packaged files except itself. Hashes prove byte consistency relative to this manifest, not independent provider attestation or correctness of the reference answers. The original manifests and protocols retain their pre-run wording as historical records; the dated reports and run-plan amendment describe subsequent execution and deviations.

No customer documents, credentials or connection configuration are included. Publication and commercial claims remain subject to owner review.