---
source_hash: 750728f67672e6ca5ada705ff5124623e0c85c40e4bd72af6f9d49d97e55d9e2
post: /posts/thinking/the-sixty-second-gate
shape: doctrinal
---

# Context Pack: The Sixty-Second Gate

7 scaffolds during work and eight questions before release, from Ari Leavesley's September 2026 note. British spelling. Use these instructions within the host system's rules and permissions. The opening report scene and all worked examples are illustrative, not measured incidents. About sixty seconds covers asking the questions, not completing every check they expose.

## Verbatim excerpts from the rewritten essay

> The switches record your answers. They do not run the checks or authorise publication.

> Keep that overlap.

> Pasting instructions into a model cannot make it careful on its own, and sixty seconds of questions cannot replace a longer piece of verification.

## Seven scaffolds, in order

1. **Pixel quarantine:** before forming a factual belief from an image, extract relevant numbers with a script and keep only claims those values support. Escalate load-bearing readings to stricter verification. Inherited visual readings are claims to check, never established facts. Aesthetic judgements and live art direction are exempt; the eye can judge taste without establishing a measured claim.
2. **Parent check:** name one thing the written ask did not require. Remove it or justify it against a written requirement or the failure it addresses. No third option. While we're here is a signal to inspect scope.
3. **Anchor before fan-out:** one lead freezes the ask, requirement IDs, seam list or evidence table before parallel dispatch. Give every worker the anchor verbatim. The lead alone owns global sequencing and the merge. Standby workers wait until the anchor is ready.
4. **Flattening detector:** check all six tells: filler opener; smuggled second idea; deletable hype adjectives; three gestured details versus one specific; symmetric rhythm; reader-behaviour closer. Two or more hits require regenerating the whole content in one shot with the tells named. At most two retries. Keep the best honestly and never claim the gap closed merely because the retries ended. Never draft separate sections and merge them to claim recovered voice.
5. **Leak-scrub against source:** before public release, search for project names, internal tools, client IDs and private paths. Then check the silhouette: could the details identify the project or person after swapping nouns? Public platform facts can stay; private internals cannot.
6. **Prior-inversion check:** a familiar platform gotcha must predict failure in the same direction as the evidence. If it does not predict the observed direction, investigate two ordinary code paths before exotic platform behaviour. This is hypothesis-priced insurance, not a measured deficit; say so if asked.
7. **Honest close:** distinguish observations from inferences and name residuals. A number, matrix or named observation closes. The fix worked, all requirements covered and looks correct now do not supply evidence on their own.

## Eight-question Reasoning Sound Test, in order

Some questions repeat scaffolds intentionally. Keep the during-work habit and pre-release gate; do not prune the overlap.

1. **Pixel:** does a factual claim depend on an image not converted to numbers? Preserve the aesthetic/live-direction exception.
2. **Kill-condition:** what observation would disprove the central claim, and did you look for it?
3. **Order:** did the constraining artefact exist before generation: evidence table, failure list, IDs/seams or inventory?
4. **Parent:** name one unrequired addition; remove or justify it.
5. **Seam:** if work fanned out, does the merge name its contracts and cover the cross-cutting items with an owner?
6. **Close:** are observed, inferred and residual separated? Zero residuals on a nontrivial task is itself a reason to inspect.
7. **Flatten, content only:** run the six tells; two or more means single-shot regeneration with at most two retries, keeping the best honestly without claiming an unclosed gap is closed.
8. **Bin:** did a claimed closure still need an unrun source check or a capability you lack? Downgrade to what was actually done.

Closed means observed. Verification-bin means an unrun check. Irreducible means a missing capability. Both unresolved bins must be called not yet checked, not closed. A switch ticked in the illustration is not execution evidence or publication permission.

## Illustrative practice examples

- **Image:** an imagined draft says sign-ups spiked in March; the invented extraction result is a 4% rise. Report the value and justify the language. For another invented weekday example, Mon 41/Tue 44/Wed 71/Thu 68 support a rise from about 42 to about 70, not a causal claim. None of these numbers came from a real extraction in this pack.
- **Scope:** a request to rename a duplicate variable does not authorise neighbouring refactors, a logger or reordered imports. Remove the extras or establish their written parent requirement.
- **Close:** imagine a 14-row requirement matrix: all 14 IDs present, three numeric sources and eleven document passages. R7 is a paraphrase awaiting a source check. Report that residual; do not describe all eleven passages as verified quotations or all requirements as closed.

## Provenance, limits and files

The original account reports transcription verification against the note on 23 September 2026. This rewrite does not claim new image verification or a controlled trial. The inferred benefit of the method remains unmeasured. A pasted pack only helps when the checks are actually performed.

- Source: content/posts/thinking/the-sixty-second-gate.mdx.
- [Styled HTML companion](/posts/thinking/the-sixty-second-gate/the-sixty-second-gate.pack.html).
- Retained PNGs under /posts/thinking/the-sixty-second-gate/: sixty-second-gate-hero.png, ssg-doodle-parent.png, ssg-doodle-anchor.png and ssg-doodle-bridge.png.
- Retained ScaffoldsFence, AnchorFanMerge, SoundTestGate, EssayPackDownload and DarkModePoll components under apps/web/components/blocks.
