Compare the new analysis with its previous publication
Type: types/instruction.md
Identify consequential coverage or qualifications lost between the previous publication and the new analysis, without repeating source verification.
Commission
The workshop coordinator loads this instruction before commissioning the comparison and supplies it to the reviewer. This is a workshop-only check, not a new stage in the standard analysis workflow. Opening this instruction does not itself authorize a worker launch.
Supply the target, exact previous and new publication paths and identities,
any recorded source/boundary differences, and an owned report path under
evidence/comparisons/. Select the previous publication explicitly; do not
silently substitute a legacy review, another system or another run. If only a
legacy review is available, return that gap for the coordinator's decision.
Keep both inputs unchanged and keep this comparison out of fresh analysts'
source-first inputs.
Compare
Read both publications and identify differences that could change a reader's understanding of the system. Concentrate on lost mechanisms, conditions, limitations and qualifications, not wording changes or preservation of every old detail. Use supporting members where the public accounts need context.
Distinguish an explained scope or source-revision change from an unexplained loss. A changed revision alone does not explain a particular omission. Treat old claims as comparison material, not established truth. Do not ask the new analysis to restore an unsupported claim merely because the old one made it.
Do not reanalyse source, inspect worker traces, acquire evidence, run the target, repair publications or integrate. Source-grounded correctness remains the existing verifiers' responsibility. When the publications cannot settle a consequential difference, flag it for a decision rather than expanding the comparison into another verification pass.
Return
Write only the commissioned comparison report. Keep it short:
- Identify both inputs by run/publication identity and immutable Git revision or member hashes; record the comparison boundary.
- Return no consequential loss found, difference needs a decision, or comparison unavailable when a required publication cannot be read.
- For each consequential difference, cite the relevant passages/member paths, explain its significance, and distinguish explained changes from unresolved ones. State any coverage limit.
The coordinator resolves flagged differences before closing the target's comparison. A positive comparison is not semantic acceptance, integration permission or clearance of an existing content or operational-incident hold. Missing publications are not a pass. Recompare if either selected input changes.