| name | eval-harmonize |
| description | Phase 3: Reconcile a draft evaluation with its critique, producing a final trace and updated YAML. |
| disable-model-invocation | true |
| argument-hint | ["resource-id"] |
Phase 3 — Harmonize Evaluation
Reconcile the draft evaluation and critique for resource $ARGUMENTS,
producing the final trace and evaluation YAML.
You are a neutral adjudicator. You have two inputs: a draft evaluation and
a critique of that draft. Your job is to determine, criterion by criterion,
whether the draft verdict stands or whether the critique's evidence requires
a change — always grounded in the RDP criteria.
Steps
-
Read all inputs.
- Read
data-drafts/$ARGUMENTS/$ARGUMENTS.yaml (the draft YAML).
- Read
data-drafts/$ARGUMENTS/draft.md (the Phase 1 trace).
- Read
data-drafts/$ARGUMENTS/critique.md (the Phase 2 critique).
- Read
docs/criteria.md for the full rubric.
-
Adjudicate each criterion.
For each criterion (A.1.1 through E.1.2):
- If the critique found no contradicting evidence: the draft
verdict stands. Note this briefly.
- If the critique found contradicting evidence: evaluate whether
the evidence actually changes the verdict per the RDP criteria.
Consider:
- Is the contradicting text actually about the data (not software)?
- Does it apply to the resource's data as a whole or only a subset?
- Is the critique's interpretation of the criteria correct?
- Could the draft and critique both be partially right?
- Record your reasoning for each decision.
-
Identify irreconcilable contradictions.
If the draft and critique present evidence that genuinely cannot be
harmonized (e.g., two pages on the same site state different licenses
with no clear resolution), flag this as a contradiction that requires
human review. Set was-controversial: "true" in the YAML if this
occurs.
-
Write the final trace.
Write data-drafts/$ARGUMENTS/trace.md combining evidence from both
the draft and critique into a single coherent trace. Use the standard
trace format from CLAUDE.md. Include:
- All navigation paths from both phases (renumber sequentially).
- For each criterion: the final verdict, the evidence that supports it,
and (if the critique raised a challenge) a brief note on why the
challenge was accepted or rejected.
-
Update the evaluation YAML.
Update data-drafts/$ARGUMENTS/$ARGUMENTS.yaml to reflect the final verdicts.
If verdicts changed from the draft:
- Update
license, license-type, license-link if needed.
- Update
license-issues to match the final criterion verdicts.
- Update
license-commentary to reflect the final findings.
- Keep
provisional: "true" — only a human reviewer removes this.
- Update
last-curated to today's date.
-
Validate.
Temporarily copy the YAML to data-sources/$ARGUMENTS.yaml, run
make check, then remove it. The YAML stays in data-drafts/ until
a human curator promotes it. Fix any validation errors.
-
Write the harmonization report.
Write data-drafts/$ARGUMENTS/harmonize.md in this format:
# Harmonization Report: {resource-id}
- **Resource**: {name}
- **Harmonized**: {YYYY-MM-DD}
- **Harmonizer**: Claude (automated, Phase 3)
- **Inputs**:
- Draft: data-drafts/{id}/draft.md
- Critique: data-drafts/{id}/critique.md
## Criterion-by-Criterion Resolution
### {criterion ID} — {short description}
- **Draft verdict**: {verdict}
- **Critique challenge**: {summary of challenge, or "none"}
- **Final verdict**: {verdict}
- **Reasoning**: {why the draft was upheld or overturned}
## Contradictions Requiring Human Review
(List any irreconcilable issues, or "None.")
## Changes from Draft
| Field | Draft Value | Final Value | Reason |
|-------|------------|-------------|--------|
| ... | ... | ... | ... |
(Or "No changes from draft." if the critique did not overturn anything.)
## Final Score
| Criterion | Verdict | Stars |
|-----------|---------|-------|
| ... | ... | ... |
| **Total** | | **N** |
-
Report to the user.
Summarize:
- Final star score
- Whether any draft verdicts were overturned by the critique
- Whether any contradictions need human review
- Remind them the evaluation is
provisional: "true" and needs review
Rules
- Follow all guidelines in CLAUDE.md under "Evaluation Guidelines."
- Your role is adjudication, not new research. Do NOT fetch any web pages.
Work only from the draft trace and critique files.
- Evaluations are about data and data access only. Ignore software licensing.
- Quotes in the final trace must be verbatim text originally fetched in
Phase 1 or Phase 2.
- The YAML must pass
make check before reporting success.
- If in doubt about a criterion, favor the interpretation supported by
more direct evidence (closer to the license page, more explicit text).