Verify Goal

Purpose

The builder must not grade its own work by re-reading it. This skill runs the checks the goal file names, as processes with exit codes, and puts their output where a fresh evaluator can read it. On a host whose native goal evaluator reads only the transcript, this output is the only evidence that evaluator will ever see.

Procedure

1. Read the goal file. Refuse when status is not active. Refuse when a criterion row lacks a check or an expected result: send it back to the goal skill. Enforce the goal's token budget before every check, setting exhausted at the limit or blocked when a finite token limit cannot be measured. Do not start another check after either state. Use iterations_used for this entry; verification does not allocate another iteration, including when the current iteration is the last allowed one.
2. Run every deterministic check, one at a time, exactly as written in the check column. Capture exit code, stdout and stderr. Never edit the command to make it pass; never substitute a similar command.
3. Compare each result with the expected column. met when it matches, unmet when the command ran and did not match, error when the command could not run (missing tool, syntax error, timeout). An error is not an unmet criterion.
4. Run review checks only when every deterministic check is met. Name the artifact read and quote the passage that decides. A review verdict without a quoted passage is unmet.
5. Write the ledger entry in the Iterations section (format below), newest last. Quote at most twenty lines of output per criterion; keep the full output in the transcript.
6. Update status: met when every row is met; active when at least one row is unmet and none is error; blocked when any row is error. Preserve exhausted or a budget accounting block from step 1; unchecked rows are unmet, not evidence of completion.
7. Print the summary to the transcript verbatim: goal slug, iteration number, per-row verdicts, status. This line is what a native evaluator matches against the condition.

Zetetic Gates

Pillar	Gate	Failure action
Logical	verdict derives from this run's result, not a previous one	rerun the check
Critical	output quoted, not paraphrased	quote it or mark unmet
Rational	review only after deterministic checks pass	skip reviews, report unmet
Essential	status updated from the verdicts, never by hand	recompute from the rows

Output Format

### Iteration <n> (<date>) by verify-goal
| id | verdict | command | exit | evidence |
|----|---------|---------|------|----------|
| C1 | met | `pytest tests/auth` | 0 | `42 passed in 3.1s` |
| C2 | unmet | `git diff --stat main -- tests/auth/fixtures` | 0 | `1 file changed` |
status: active

Hand-offs

Condition	Next skill	Reason
every row met	none; status met	the contract is fulfilled
a row unmet	refine-goal	the next iteration needs a backlog
a row error	none; status blocked	the check itself must be fixed first

Anti-patterns

• Reading the diff and declaring a criterion met without running its command.
• Rewriting a failing command into one that passes.
• Reporting "tests pass" without the count and the exit code.
• Treating a timeout as unmet and iterating on code that was never measured.

Examples

Iteration 2: C1 met (pytest tests/auth, exit 0, 42 passed), C2 unmet (fixture diff shows one file). Status stays active; the unmet row goes to refine-goal, which will add a step to restore the fixture rather than relax C2.