agentleFS
svy04svy04/ballast/skills/verify-gate/SKILL.md

verify-gate

Skill73 starsChanged 3 months agoMITView on GitHub

Treat research results and model knowledge as drafts until verified. Use when researching, quoting numbers or sources, recording knowledge as fact, or before declaring a task successful.

7 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

1 rule. 0 are common, 0 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 4 edited versions we read keep these.

  • 0

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Write the gate's result down where the claim lives — what was argued against it, which sources were opened, what survived.

    line 30

“Common” is measured: kept by at least 6 in 10 of the 4 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Verify gate

Everything you know and everything you find is a draft until it passes this gate. The failure mode this prevents: plausible statements hardening into "facts" through repetition.

Labels — attach one to every claim

LabelMeaningAllowed use
confirmedPassed the gate belowMay be stated as fact
observedYou (or a tool run) directly saw it happen, single instanceState with "observed on <date>"
assumedInference; plausible but unverifiedMust carry the label in-text
hearsayA source said it; no primary source seenMust name the source
unknownNo basisSay "unknown" — never fill the gap

Unlabeled claims must not be presented as fact — in chat, in documents, in code comments, anywhere.

The gate (required before confirmed)

  1. Refute first. Actively search for evidence against the claim before collecting support.
  2. Primary sources, ≥2, independent, named. A blog quoting a blog is one source, not two.
  3. Sample size stated. n=1 is an anecdote — label it observed, not confirmed.
  4. Limits stated. Where does the claim not hold?
  5. Date stamped. Verified knowledge expires — re-verify anything older than 90 days before relying on it.

Write the gate's result down where the claim lives — what was argued against it, which sources were opened, what survived. A confirmed label with no retrievable record is assumed wearing a badge.

Second-model verification (optional, recommended)

Different vendors have different blind spots. If the project has .claude/ballast.verifier.json (this skill reads it — the rules hook does not; a copy-ready example ships in the ballast repo at rules/ballast.verifier.example.json):

{ "command": "your-verifier-cli --check" }

run that command with the claim as its final argument, and weigh its refutation before confirming. Without an external verifier, you may still reach confirmed — but only via primary sources you actually opened, and note (self-gated) next to the label.

Health-check the verifier. The first time a session leans on it, confirm the command actually runs. If it fails, say so once and proceed (self-gated) — a configured verifier that silently never ran is the worst outcome, because the label implies a scrutiny that never happened.

Two things to get right when reading what comes back:

  • An empty, truncated, or off-topic reply counts as a failed run, not a pass — say so and fall back to (self-gated).
  • A refutation is rarely all-or-nothing: read the per-claim verdicts and any restatement the verifier proposes, and confirm the narrowed wording rather than rejecting the whole claim.

Success is also a claim

"Task done" passes the same gate: verify the thing actually happened (file exists, test passes, output is correct) before reporting success. An enthusiastic user reaction is sense-data, not verification — log it as a hypothesis and design a check.


Part of ballast — foundations first, every solved path kept, done means checks pass.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.