midnightntwrk/midnight-expert/zkir-regression

ZKIR Regression Sweep

Run a curated set of verification claims against the current toolchain to detect behavioral changes. Each claim is verified through the full pipeline — claim classification, contract writing, compilation, execution, and PLONK proof verification.

Use this when:

• A new compiler version or checker package version is released
• An agent suspects unexpected behavior from the toolchain
• You want a confidence check before presenting Midnight-specific claims

Step 1: Determine Mode

If $ARGUMENTS is empty → full sweep (all categories).

If $ARGUMENTS contains a category name → targeted sweep (that category only).

Valid categories: arithmetic, types, state, privacy, zk-proof, transcript.

Step 2: Record Toolchain Versions

compact compile --language-version
compact --version

Record both for the report header.

Step 3: Run Claims

For each claim in the list below (filtered by category if targeted):

1. Load the midnight-verify:verify-correctness skill to classify the claim domain
2. Load the appropriate domain skill
3. Dispatch the sub-agent(s) indicated by the domain skill's routing table
4. Collect the verdict

This follows the same flow as the /midnight-verify:verify command — you are the orchestrator for each claim.

Claim List

Step 4: Compare Results

For each claim, compare the actual verdict against the expected verdict:

• Verdict matches expected → PASS
• Verdict does not match → FAIL
• Verification returned Inconclusive but expected Confirmed → FAIL (toolchain may be unavailable)

Step 5: Report

## ZKIR Regression Report
 
**Toolchain:** compact CLI vX.Y.Z, language version A.B.C
**Date:** YYYY-MM-DD
**Mode:** [full sweep / targeted: <category>]
**Ran:** N claims
 
### Results
 
| Category | Passed | Failed | Total |
|---|---|---|---|
| arithmetic | N | N | N |
| types | N | N | N |
| state | N | N | N |
| privacy | N | N | N |
| zk-proof | N | N | N |
| transcript | N | N | N |
| **Total** | **N** | **N** | **N** |
 
### Failures (if any)
 
**<claim-id>:** Expected <expected verdict>, got <actual verdict>
-Claim: "<claim text>"
-Actual result: [what the verification pipeline returned]
-Interpretation: [what this failure suggests about toolchain changes]

If there are zero failures, end with:

All N claims passed. Toolchain behavior matches expectations.

Adding New Claims

Add a row to the claim list table above. Each claim should:

• Be verifiable through the normal /midnight-verify:verify pipeline
• Have a deterministic expected verdict
• Test a specific, observable behavior
• Include the expected verdict qualifier (tested, zkir-checked, zkir-inspected, source-verified)