Conducty Verify

Conducty Verify — The Evidence Gate

Run the command. Read the output. Then claim the result. Nothing else qualifies.

The Gate

Before ANY completion claim, status update, or expression of satisfaction:

1. IDENTIFY — What command proves this claim? (from the prompt's Verification field)
2. RUN     — Execute the command (Bash) fresh and complete
3. READ    — Full output, exit code, failure count
4. JUDGE   — Does the output confirm the claim?
              YES → State the claim WITH the evidence
              NO  → State what actually happened WITH the evidence
5. RECORD  — Log the result for checkpoint and improvement tracking

Skip any step and the claim is unsupported. "Should work" is not evidence. "The agent said it passed" is not evidence. Only output from a command you just ran via Bash is evidence.

Calibrated Verification

The plan assigns a review level to each prompt. Verification scales accordingly:

verify-only (Low complexity)

Run the verification command. Read the output. If it passes, the prompt is done.

IDENTIFY: P3 verification is `npm test -- --grep "auth"`
RUN:      npm test -- --grep "auth"
READ:     12/12 tests pass, exit 0
CLAIM:    "P3 verified — 12/12 auth tests pass"

spec-review (Medium complexity)

Run the verification command. If it passes, dispatch the spec compliance reviewer (see [[conducty-execute]]). Both must pass.

IDENTIFY: P5 verification is `cargo test session`
RUN:      cargo test session
READ:     8/8 tests pass, exit 0
THEN:     Dispatch spec reviewer → spec-pass
CLAIM:    "P5 verified — tests pass, spec compliant"

full-review (High complexity)

Run the verification command. If it passes, dispatch spec reviewer. If spec passes, dispatch quality reviewer. All three must pass.

IDENTIFY: P7 verification is `make test && make lint`
RUN:      make test && make lint
READ:     all pass, exit 0
THEN:     Spec reviewer → spec-pass → Quality reviewer → pass
CLAIM:    "P7 verified — tests pass, spec compliant, quality approved"

Checkpoint Verification

When verifying an entire group (used by [[conducty-checkpoint]]):

For EACH prompt in the group:
  IDENTIFY → RUN → READ → JUDGE
  Record: pass with evidence / needs-fix with failure output

Report: "Group A: 3/4 passed. P2 needs-fix (2 test failures in auth module)"

Never say "Group passed" without individually verifying every prompt.

Fix Verification

When verifying a fix prompt resolved the original issue:

1. Run the ORIGINAL verification command (the one that failed)
2. Run any ADDITIONAL check for the specific failure that triggered the fix
3. Both must pass
CLAIM: "Fix verified — original tests pass AND specific failure resolved"

Evidence Format

For consistency and machine-readability in [[Metrics]] / [[Prompt Log]]:

**P{N} Verification** [{timestamp}]
- Command: `{command}`
- Exit code: {0/1/...}
- Result: {pass/fail}
- Detail: {summary of output — test count, error message, etc.}
- Review: {verify-only/spec-review/full-review} — {outcome}

This format feeds into [[conducty-checkpoint]] health metrics and [[conducty-improve]] learning.

When to Apply

• Before marking ANY prompt as completed
• Before passing ANY prompt during checkpoint
• Before claiming ANY fix resolved an issue
• Before moving to the next parallelization group
• Before any expression of satisfaction about work state