eval(pilot): demonstrate parity with the proven manual CV workflow

[akoita]

Parent: #4

Outcome

Demonstrate that DraftLoop can reproduce the value of the proven manual author-critic workflow with less repetitive user effort.

Scope

• Use a consented private human-authored application CV as the local baseline.
• Keep raw candidate material, job content, provider responses, and reference CV private.
• Commit only sanitized structural fixtures and aggregate measurements.
• Define the private thresholds for complete, professional, application-ready, and comparable before evaluating the run.
• Compare first draft, critic report, revised draft, and final human-approved output.
• Measure factual invariants, unsupported additions, required-section completeness, chronology, relevant-achievement recall, critical-requirement coverage, style violations, ATS/visual readiness, manual edits, review time, rounds, provider cost, and confidence.

Acceptance criteria

• Zero factual invariant violations and unsupported model-added facts.
• Required sections and chronology are preserved unless explicitly omitted.
• Relevant-achievement recall and requirement coverage meet the agreed private baseline threshold.
• Final output is judged usable for a real application after bounded human review.
• Results and limitations are recorded without publishing personal data.
• One consented case is reported as representative evidence, not general product validation; broader confidence requires later pilot cases.
• Adversarial and misleading-source cases remain part of the evidence.
• pnpm validate passes.

Dependencies

• v0.7 evidence-backed drafting milestone
• v0.8 independent review and readiness milestone