Verify before asserting

Verify Before Asserting

Overview

Models state wrong facts with full confidence — a fabricated citation, an out-of-date threshold, a misremembered dosage or limit. In a high-stakes context that is a liability, not a typo. A high-stakes external fact is not "known" until verified against an authoritative source. Memory does not count.

This is a discipline skill. Follow it even when you're confident — confidence is exactly the failure mode. Violating the letter of this rule is violating its spirit: a fact you're "sure of" still needs a source.

This is distinct from verification-before-completion (verify your own work is done) and karpathy-guidelines (surface assumptions before coding). This skill is about external facts you're about to assert.

When this fires (and when it doesn't)

Fires when both are true:

1. You're about to assert a specific external fact — a figure, threshold, rate, date, dosage, limit, config default, statutory/section reference, API version, standard clause, or named rule.
2. Being wrong carries real cost: legal, financial, medical, safety, regulatory, security, or production-reliability (a vendor limit, config default, or API version baked into shipping code).

Does NOT fire for low-stakes or common knowledge — general explanations, brainstorming, naming, code that's trivially testable. Example: "how does a hash map handle collisions, and suggest a variable name" needs no verification. Don't verify trivia — that's paralysis, not diligence. If genuinely unsure whether stakes are high, a quick check is cheap — verify.

The discipline

1. Retrieve an authoritative source. Order of authority (general):
   • Primary / official source (the law, the standard, the spec, the label, the vendor doc).
   • Issuing/regulating body or official registry.
   • Official guidance from that body.
   • Secondary commentary is a lead, not authority. Blogs, summaries, forum answers, news, another model — they point you to the source; they never replace reading it.
   • Model memory — never sufficient on its own for a high-stakes fact.
2. Cite precisely — to the section / clause / page / version, not "the docs say."
3. Check status — current, superseded, draft, or proposed? A draft/proposed fact may not be relied on as final.
4. Check currency — what date/version does it apply to? Has it changed since?

If you cannot reach an authoritative source (no web/tool access, doc unavailable): do not assert the fact as known. Label it explicitly unverified, say what you'd check, and stop — an unverified high-stakes fact is not an answer. Where a value drifts (vendor limits, rates), prefer reading it from a live authoritative source at runtime over hardcoding a copy.

Stakes gate

digraph verify {
    "About to assert a specific external fact" [shape=box];
    "Would being wrong cost real harm?" [shape=diamond];
    "Authoritative source reachable?" [shape=diamond];
    "Current / non-draft?" [shape=diamond];
    "State it, with citation" [shape=box];
    "Label unverified; do not rely" [shape=box];
    "Flag as draft / superseded" [shape=box];
    "State normally" [shape=box];

    "About to assert a specific external fact" -> "Would being wrong cost real harm?";
    "Would being wrong cost real harm?" -> "State normally" [label="no"];
    "Would being wrong cost real harm?" -> "Authoritative source reachable?" [label="yes"];
    "Authoritative source reachable?" -> "Label unverified; do not rely" [label="no"];
    "Authoritative source reachable?" -> "Current / non-draft?" [label="yes"];
    "Current / non-draft?" -> "State it, with citation" [label="yes"];
    "Current / non-draft?" -> "Flag as draft / superseded" [label="no / draft"];
}

Rationalizations — STOP

Excuse	Reality
"It's common knowledge"	Common knowledge is often out-of-date knowledge.
"I'm pretty sure it's X"	Sure-from-memory is the exact failure mode. Confidence ≠ correctness.
"I'll add the citation later"	A fact without provenance is a guess with a decimal point.
"A reputable blog/firm/model says so"	Secondary sources are leads. Open the primary one.
"I remember this section/version number"	Citing a locator you haven't opened is fabrication.
"It was right last year/version"	Thresholds, limits, and labels change. Re-check currency.
"No source handy, but it's probably fine"	"Probably fine" is unverified. Label it, don't assert it.

All of these mean: open the authoritative source and cite it, or label the claim unverified.

Output

State a verified high-stakes fact as: value/claim | authoritative source (precise locator) | status | applies-as-of | date checked.

Worked example: Lambda max timeout = 900 s | AWS Lambda Developer Guide, "Lambda quotas" | current | as of 2026-06 | checked 2026-06-01

Reference instance

verify-au-tax-law is the worked, domain-specific instance of this skill: it fills the authority hierarchy with Australian tax sources (legislation.gov.au, the Federal Register, ATO guidance) and the status gate with registered-vs-draft. Build new domain instances the same way — inherit this discipline, supply the domain's real authorities.