One workflow.
Three control tests.

Accounts Payable validates the supplier invoice. Budget Owner Elena Ruiz approves the payment and records her approval in the payment record. Treasury checks that record and releases the payment only after approval.

Expected answer

{
  "human_approval": true,
  "autonomy": "human_approval_required",
  "disposition": "accept",
  "gap_code": "NONE"
}

Baseline — exact response

{
  "human_approval": true,
  "autonomy": "human_approval_required",
  "disposition": "accept",
  "gap_code": "NONE"
}

Trained V41 — exact response

{"human_approval": true, "autonomy": "bounded_autonomy", "disposition": "accept", "gap_code": "NONE"}

Accounts Payable validates the supplier invoice. Treasury then releases the payment. No Budget Owner authorization step or authorization record appears in this workflow.

Expected answer

{
  "human_approval": true,
  "autonomy": "human_approval_required",
  "disposition": "human_review_required",
  "gap_code": "MISSING_APPROVAL_GATE"
}

Baseline — exact response

{
  "human_approval": false,
  "autonomy": "bounded_autonomy",
  "disposition": "human_review_required",
  "gap_code": "MISSING_APPROVAL_GATE"
}

Trained V41 — exact response

{"human_approval": false, "autonomy": "bounded_autonomy", "disposition": "human_review_required", "gap_code": "MISSING_APPROVAL_GATE"}

Accounts Payable validates the supplier invoice. Budget Owner Elena Ruiz gives written clearance for the payment, retained in the payment record. Treasury verifies that clearance and releases funds only afterward.

Expected answer

{
  "human_approval": true,
  "autonomy": "human_approval_required",
  "disposition": "accept",
  "gap_code": "NONE"
}

Baseline — exact response

{
  "human_approval": true,
  "autonomy": "human_approval_required",
  "disposition": "accept",
  "gap_code": "NONE"
}

Trained V41 — exact response

{"human_approval": true, "autonomy": "bounded_autonomy", "disposition": "accept", "gap_code": "NONE"}

Review this synthetic workflow against the stated policy. Return only one JSON object with exactly these fields: human_approval (boolean: true when the policy requires human authorization, even if documented), autonomy ('human_approval_required' when that authorization is required, otherwise 'bounded_autonomy'), disposition ('accept' if the documented workflow satisfies the policy, otherwise 'human_review_required'), gap_code ('NONE' if the required gate is documented, otherwise 'MISSING_APPROVAL_GATE'). 'accept' evaluates the workflow description only and does not authorize any real payment. Equivalent wording for authorization counts when its owner, evidence, and order are explicit.

Build one small, evidence-backed comparison of a baseline model and a fixed trained candidate on three synthetic payment workflows: recorded approval present, approval missing, and equivalent approval wording. Use this fictional rule: every supplier payment requires a named Budget Owner’s recorded authorization before Treasury releases funds; invoice validation alone is not authorization.

Before inference, write and freeze the three inputs, expected JSON answers, model identities and checkpoint hashes. Define human_approval as whether policy REQUIRES human authorization, not whether approval is already present. Both valid cases must preserve that human requirement. The missing case must identify MISSING_APPROVAL_GATE and require human review. “Accept” assesses the workflow description only; it never authorizes a real payment.

Use the same prompt, base revision, quantization and decoding for both models. Run each case once, preserve exact raw responses, and report parsing failures without repair. Do not change expected answers after seeing outputs. Do not train, promote checkpoints, change production, use real payment data, or execute a payment. If either model is unavailable, mark it not tested.

Deliver a single-page comparison, exact prompts and expected answers, raw outputs, strict all-field scores, gap-detection results, false positives and limitations. Explain whether improvement is in finding gaps or in maintaining consistent authority fields. Three related examples do not prove generalization or release readiness. Publish only when authorized.

• Three related synthetic text cases; one response per model per case.
• No statistical or broad generalization claim.
• No claim these cases are disjoint from all training data.
• The fixed V41 checkpoint failed broader development gates; this is diagnostic only.
• Research adapter weights are not distributed here; full independent reproduction of trained inference requires those exact weights.
• Workflow classification only; no real payment or protected-tool enforcement tested.

The downloadable package contains the exact test inputs, expected answers, raw outputs, decoding settings, and adapter hashes. It supports independent rescoring. Exact trained-model inference also requires access to the recorded research adapter.