anti-hallucination-gates

Anti-Hallucination Gates

The most common agent failure is declaring victory too early. This skill is the gate between "I think it works" and "evidence proves it works".

The Iron Rule

No claim of "done" / "fixed" / "passing" without a concrete artifact path proving it.

A claim like "I fixed the bug" is incomplete. The proper form:

"Fixed by commit abc123. Test tests/foo.test.ts::handles-edge-case now passes (output captured in .harness-anchor/test-2026-05-28.log). Build still passes (.build/last-build.log exit 0)."

Default-FAIL Evaluation Contract

Adapted from Anthropic's Code with Claude 2026 reference impl:

"Every criterion starts false; the agent can't mark it passing without opening evidence first."

Standard criteria (adapt to project)

Criterion	What counts as evidence
Compile / build passes	Build log path + exit code 0, or compiler stdout/stderr captured
Type-check passes	tsc --noEmit (or equivalent) output
Tests pass	Test runner output showing N passed, 0 failed, 0 errored
Coverage obligations (non-trivial / risk-bearing features)	The spec/code obligations are exercised by tests the run actually executes — for features with real logic or risk constructs, a /test-plan report (.harness-anchor/coverage-<ts>.md) with no open gaps. A green suite that skips the risk path is a false pass.
Static analysis	Lint report file, OR explicit "no warnings" line in output
Test integrity (when tests changed in the same diff)	The claim names why each test change happened (new coverage vs adjusted expectation) — /verify reports this as ### Integrity; a weakened assertion that makes failing behavior "pass" is a red flag
Deliverable committed & reproducible (when claiming "done / shipped / committed")	Full git status reviewed (the whole tree, not just state files) and the committed HEAD builds — not only the dirty working tree. A feature marked pass whose source isn't committed is not delivered: HEAD won't reproduce the evidence.
Manual smoke	Concrete steps + observed result (only if the above can't cover)

Filling out evidence in feature_list.json

{
  "id": "doc-list-rendering",
  "status": "pass",
  "evidence": {
    "timestamp": "2026-05-28T13:45:21Z",
    "commit": "abc123def456",
    "artifacts": [
      ".harness-anchor/build-2026-05-28T13-44.log",
      ".harness-anchor/test-2026-05-28T13-45.txt"
    ],
    "notes": "All 47 tests pass. Build clean. clang-tidy: 0 warnings on changed files."
  },
  "completedAt": "2026-05-28T13:45:21Z"
}

evidence: null ⇒ status cannot be "pass". This is enforced by feature_list.schema.json.

Calibrated Uncertainty (2026 consensus)

When evidence is missing or partial, do NOT bluff. The 2026 LLM-hallucination research consensus is:

"Aim for calibrated uncertainty: systems that transparently signal doubt and can safely refuse to answer when unsure."

Templates for expressing uncertainty:

• "I am uncertain whether the test passes because I didn't run it. Recommend npm test -- --testPathPattern foo to verify."
• "The build appears to succeed based on no error in the output, but I did not see an explicit success line. Recommend cmake --build .build && echo BUILD_OK."
• "This fix may resolve the issue, but I have not reproduced the original failure. To gain confidence, run ./repro.sh before and after the change."

These statements earn user trust. Confident-but-wrong claims destroy it.

Anti-patterns

Do NOT say:

• "Should work now."
• "The build should pass."
• "Tests should be green."
• "This looks correct."
• "These are old / unrelated changes." (said about uncommitted files without running git diff to confirm — verify before you dismiss; this is how a pass's own source gets left behind)

Say instead:

• "Build passes (evidence: ) / Build pending verification (recommend: )"
• The test status alongside a captured artifact path.

The "should" detector

When you find yourself typing "should", "probably", "I think", or "looks like" about a behavioural claim, pause. Either:

• Run the command to verify, then state the observed result, OR
• Explicitly mark the claim as unverified per the templates above.

Pre-claim checklist

Before saying ANY of {"done", "fixed", "ready", "complete", "passing", "working"}:

- [ ] I ran the build/compile, observed exit code 0
- [ ] I ran the test suite, observed N passed / 0 failed
- [ ] the passing tests exercise the spec's edge cases / risk paths — not just the happy path (run `/test-plan` if unsure)
- [ ] if tests changed alongside source: I explained each test change (new coverage vs adjusted expectation)
- [ ] I have a file path for each piece of evidence
- [ ] feature_list.json is updated with evidence object + timestamp + commit
- [ ] if I claimed "committed / shipped / delivered": I reviewed the **full** `git status` and confirmed the committed HEAD builds (not just the working tree)

If any box is unchecked: state uncertainty explicitly, do NOT flip status to pass.

When to invoke

• The user asks "is it done?"
• You're about to update feature_list.json status to "pass"
• You're about to say "the fix should work"
• The PostToolUse hook injected warnings — do NOT silently ignore them; surface and address per self-correction-loop

Looking up evidence commands for unfamiliar frameworks

When the project uses a test/lint framework you don't have committed to memory (Catch2, doctest, ruff, deno test, etc.) — invoke the docs-lookup skill before constructing the verification command. The lookup result becomes part of the evidence trail.

Bluffing the command and not actually running it is the anti-pattern this skill exists to prevent — that's why lookup precedes execution here.

Related

• verification-before-completion (superpowers) — the same Iron Law (no completion claim without fresh evidence). This skill is its harness-anchor counterpart and superset: it adds the on-disk evidence record (feature_list.json) and the /verify subagent. One verification run satisfies both gates — capture the evidence once, don't re-verify.
• feature-state-keeper — actual writes to feature_list.json
• test-coverage-design / /test-plan — derive the coverage obligations this gate checks (are the right things tested, and actually run?)
• self-correction-loop — what to do when evidence shows failure
• /verify command — full automated verification pass via verification-runner subagent