Blind A/B Human Review

Reviewer cockpit for output quality decisions

Compare visible Variant A and Variant B outputs, fill the decision file, then run adjudication. The answer key is intentionally hidden from this page.

Cases5

Ready0

Pending5

Invalid0

Review Flow

Open reports/output_blind_review_pack.md or this kit and compare Variant A vs Variant B for each case.
Record choices in reports/output_review_decisions.json without opening the answer key.
Use winner_variant A or B, confidence from 0 to 1, and a short reason for every case.
Set reviewer_attestation truthfully after choices are recorded and before adjudication.
Run python3 scripts/yao.py output-review-import --input <reviewer-decisions.json> --blind-review-attested --run-adjudication after choices are recorded.
Refresh python3 scripts/yao.py review-studio . before asking for release approval.

Case 01

skill-package-contract

Awaiting

Turn this repeated workflow into a reusable team skill package.

Rubric

has-entrypoint
Output names the routeable Skill entrypoint.
1.0
has-interface
Output includes neutral interface metadata.
1.0
has-report
Output points reviewer to generated overview report.
1.0
has-resource-boundary
Output preserves entrypoint/resource separation.
1.0

Variant A

Create a routeable SKILL.md, agents/interface.yaml, reports/skill-overview.html, and a focused eval plan. Keep the root entrypoint lean, move durable guidance into references/, scripts into scripts/, and evidence into reports/.

Variant B

I can write a prompt for that workflow and include a short checklist.

Case 02

output-eval-expectation

Awaiting

Upgrade this production skill so we know whether the generated output is better than baseline.

Rubric

has-baseline-comparison
Output explicitly compares with-skill and baseline outputs.
1.0
has-assertion-grading
Output includes objective assertion grading.
1.0
has-scorecard
Output produces a scorecard report path.
1.0
has-failure-taxonomy
Output records failure taxonomy.
1.0

Variant A

Add Output Eval Lab cases with baseline_output and with_skill_output, run assertion grading, report with-skill vs baseline pass-rate delta, and record failure taxonomy plus next fixes in reports/output_quality_scorecard.md.

Variant B

Add more examples and run the trigger eval again.

Case 03

ir-before-packaging

Awaiting

Package this skill for OpenAI, Claude, Agent Skills, and generic targets.

Rubric

has-ir-first
Output requires Skill IR before packaging.
1.0
has-core-ir-fields
Output lists core IR contract fields.
1.0
has-targets
Output names requested runtime targets.
1.0
has-semantic-preservation
Output says targets must preserve the capability contract.
1.0

Variant A

Export folders for each platform and adjust files as needed.

Variant B

Export Skill IR first with job_to_be_done, trigger_surface, workflow, resources, eval_plan, risk, and governance. Then compile or package targets from the IR so OpenAI, Claude, Agent Skills, and generic packages preserve the same capability contract.

Case 04

near-neighbor-boundary

Awaiting

I only need a one-off summary of these notes, not a reusable process.

Rubric

declines-skill
Output refuses unnecessary skill creation for one-off work.
1.0
names-near-neighbor
Output labels the request as near-neighbor instead of owned work.
1.0
requires-reuse-signal
Output asks for repeat-use evidence before packaging.
1.0

Variant A

Create a SKILL.md and a reusable workflow anyway so future notes can use it.

Variant B

Do not create a skill for this one-off request. Treat it as a near-neighbor: answer the summary directly unless the user confirms repeated use, shared ownership, or a reusable output contract.

Case 05

file-backed-governed-package

Awaiting

Turn the attached release brief source into a governed skill package.

Rubric

uses-file-backed-evidence
Output names file-backed source evidence.
1.0
has-governance
Output preserves governed ownership metadata.
1.0
has-output-contract
Output preserves output and rollback boundaries.
1.0
has-trust-and-scorecard
Output requires trust and output scorecard artifacts.
1.0
does-not-invent-evidence
Output forbids invented launch evidence.
1.0

Variant A

Draft a release announcement with the changelog, support notes, and owner name.

Variant B

Use the file-backed fixture as source evidence, then create a governed skill package with SKILL.md, agents/interface.yaml, owner, review cadence, input_files, output contract, rollback boundary, trust report, and reports/output_quality_scorecard.md. Mark missing launch metrics as missing evidence instead of inventing them.

Decision Template

Use this shape in reports/output_review_decisions.json; leave a case blank when the reviewer is not ready.

{
  "schema_version": "1.0",
  "reviewer": "",
  "reviewed_at": "",
  "decisions": [
    {
      "case_id": "skill-package-contract",
      "winner_variant": "",
      "confidence": null,
      "reason": ""
    },
    {
      "case_id": "output-eval-expectation",
      "winner_variant": "",
      "confidence": null,
      "reason": ""
    },
    {
      "case_id": "ir-before-packaging",
      "winner_variant": "",
      "confidence": null,
      "reason": ""
    },
    {
      "case_id": "near-neighbor-boundary",
      "winner_variant": "",
      "confidence": null,
      "reason": ""
    },
    {
      "case_id": "file-backed-governed-package",
      "winner_variant": "",
      "confidence": null,
      "reason": ""
    }
  ]
}