fable-verify

fable-verify

"It ran" is not verification. Evidence before claims, always. Verify at the layer of the claim, not the layer below it.

The gate

Before any success/completion claim or expression of satisfaction:

1. Identify the command or observation that would prove this specific claim.
2. Run it fresh, in full (not a remembered earlier run; not a partial check).
3. Read the whole output — exit code, counts, the actual values.
4. Verify at the layer of the claim. Exit 0 / "deploy healthy" / "containers up" only proves the layer below the claim. If the claim is "the output is correct," look at the output. If it's "the page renders," look at the page. If it's "the definitions load," import them in the built artifact and count them.
5. Only then state the claim — as ledger lines: Verified: <claim> — ran <command> -> saw <result>, with anything unchecked under Assumed: or PROVISIONAL. If it fails, state the actual status with the output.

Sharpen it

• Know what pass looks like before you run it: pull the oracle from .fable/project.md — the command and what green literally prints. Exit 0 with 3 skipped is not the pass you meant.
• The counting environment is binding: the oracle row's Counts where decides where green counts. Local green on a CI-counted claim stays PROVISIONAL until the environment of record agrees — quote it (e.g. gh pr checks).
• Discharge residuals — durable write first, delete second: when this evidence settles an entry in .fable/residuals.md, append its resolution to .fable/claims-log (as a Verified: line if that's what it became), then delete the line from residuals.md, then announce it. The order is load-bearing, not stylistic: interrupt it the other way round and you lose the residual and never record the claim — and that lost Verified: is exactly what fable-debug needs to falsify against when the behaviour later breaks. A residual is discharged only once its resolution is durable somewhere else. The counters treat any remaining Assumed:/PROVISIONAL line as open.
• Categorical over enumerated: assert a property over all items of a class, so the check inspects cases you didn't think to list. When it over-fires, diagnose scope vs. substance before loosening it.
• Oracle over the whole population: when reconstructing hidden logic, diff your candidate against a readable known-good output over every row, not a sample; state plainly which parts are transcribed vs. inferred.
• No known-good output? Manufacture the oracle with a metamorphic relation: state how the output must change when the input changes in a known way (add a row → the count rises by one; permute input order → the result is unchanged), then check that property. It turns hidden truth into a runnable check.
• A number derived from data is a claim about the data: before quoting a count, rate, or aggregate as fact, run one independent sanity check — reconcile against a known total, re-derive one row or sample by hand, or check an order-of-magnitude bound — and record the query that produced the number beside it. A figure that fails its sanity check ships as PROVISIONAL or not at all.
• Experiential claims get experienced: when the claim is that an artifact renders, reads, or is usable — a page, dashboard, chart, document — the proving observation is entering its modality: open it, screenshot it, click the flow, and record what you literally saw as the saw half of the ledger line. Green tests and zero-must-fix reviews prove the layer below, and the looking step only fires reliably when it is named — so on a project that ships such artifacts repeatedly, mint an experiential oracle row rather than trusting the habit.
• Sample the tails: first item, last item, weirdest item — not just the middle.
• Use evidence you didn't generate: re-open the file you wrote, re-run, screenshot and read it, diff before/after, count what you claimed to count.
• Treat good news as suspect: a pass that came too easily is unverified until you can say why it's real. Distinguish "the build succeeded" (rehearsal) from "the thing loaded and ran" (reality).
• Re-check against the original request and any standing rules from scoping.