halt-as-deliverable
When a pre-registration, pre-flight or sanity gate refuses to produce your deliverable, the catch is often worth more than the deliverable. Use when a discipline you built caught your own work.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 1
- Forks
- 1
01 前言 Front Matter
- name
- halt-as-deliverable
- description
- When a pre-registration, pre-flight or sanity gate refuses to produce your deliverable, the catch is often worth more than the deliverable. Use when a discipline you built caught your own work.
02 资源文件 4
- 参考文档case-study.md3.7 KB
- 参考文档evals/evals.json3.3 KB
- 参考文档EVIDENCE.md5.2 KB
- 参考文档gotchas.md1.3 KB
03 SKILL.md 正文
HALT as Deliverable
Problem
A quality discipline (pre-registration, pre-flight check, sanity gate) refuses to produce the intended result because it detected an inconsistency in the surrounding state. The engineer reflex is to fix the inconsistency and re-run, treating the HALT as a setback — which discards the most valuable artifact the work produced. A discipline-gate catching its own author's accidental falsification, before any contaminated result ships, is an uncopyable credibility signal: it proves the discipline works on real, unrehearsed material. Where the framework's credibility matters more than the specific result, the HALT is the higher-value deliverable.
Use when
- A pre-registered prediction is falsified by a known class of error before the experiment executes: registry drift, version mismatch, schema change, baseline/tag documentation inconsistency.
- A pre-flight verification refuses the run because operational state is inconsistent with the experiment's preconditions.
- A self-audit, linter or regression check catches a defect in the author's own shipped work — the defect the audit existed to demonstrate the absence of.
- Any "the discipline refused to produce X, and here is why" outcome where the why is structurally informative rather than transient infrastructure noise.
Deciding between the two paths:
| Apply HALT-as-deliverable | Apply fix-and-re-run |
|---|---|
| The HALT reveals a structural inconsistency the audience would care about | The HALT reveals transient infrastructure (rate limit, flaky network) |
| The discipline's credibility is itself the claim at stake | The experiment's specific result is the claim at stake |
| The catching is uncopyable / hard to fake | The catching is generic / commodity |
| The audience is field-skeptical (impressed by the discipline) | The audience is result-focused (would not engage with meta) |
| The author's own work was caught | Someone else's work was caught |
Solution
The worked record behind every step is in case-study.md.
1. Commit the audit trail publicly before deciding the path forward
The HALT findings, the failed pre-registration, the pre-flight log — commit them to public history immediately, before deciding whether to re-run. This locks the catching in place; it cannot be retroactively papered over.
2. Make the catching the headline, not a quiet fix
The reflex — fix, re-run, ship the corrected number, mention the re-baseline in passing — is correct for routine engineering and WRONG when the discipline-catching IS the demonstration. Name the HALT in the title, document the audit trail explicitly, narrate the discipline-catching-itself moment as the lead. The corrected result is a number; the HALT is a story, and a story is a character claim about the work. In fields where the dominant pattern is "produce a confident number even when you shouldn't," publicly halting before fabricating is itself the artifact — loudness is substance here, not marketing.
3. Reframe the deliverable around the catching
If the original deliverable was "experiment produces result X showing claim Y," the reframed deliverable is: "the discipline pre-registered claim Y, attempted the experiment, refused to proceed because Z was inconsistent, and here is the audit trail." Same evidence, sharper story — the catching of Z is the demonstration of claim Y.
4. More than one HALT compounds
A series of HALTs catching different classes of inconsistency (documentation drift, operational state, environment configuration) is dramatically more credible than one HALT plus more than one silent fix. The compounded trail triangulates what doing this honestly looks like.
Verification
The reframe is correctly applied when:
- The audit trail was committed to public history BEFORE the reframe. Order matters; backdating breaks audit integrity.
- The headline names the HALT explicitly — "we caught X" is the lead.
- A reader who came for the original deliverable would accept the substitute on its merits, not because they were told to.
- The discipline is recoverable from the trail: someone implementing it would have caught the same class of error.
The reframe FAILS when the HALT reads as marketing spin; when the original claim was specific (a number, a comparison) and the substitute is vague — the reframe does not work for narrow factual claims; or when the audience explicitly came for the result, where meta reads as evasion.
Notes
Open gotchas.md when a pre-flight gate refuses to produce your deliverable and you are about to treat the HALT as a setback. It records the origin incidents and a second, independent occurrence.
- The pattern requires the discipline to be REAL — falsifiable, recoverable from the audit trail, applied uniformly, catching real problems including the author's own mistakes. Decorative pre-registrations produce HALTs that do not generalize.
- Identity match amplifies it: "the framework I built refused to validate my own shipped work" is uncopyable; "the framework caught someone else's work" is much weaker.
- The whole pattern collapses if the framing is perceived as motivated reasoning. Mitigations: commit the trail first, make every input to the catching publicly verifiable, and report HALTs unselectively.
04 同仓库更多 Skills
subagent-handback
Use when prompting, briefing or dispatching a research subagent, or when one goes idle with no report. Name a return channel, check its tool grant, verify every claim, negatives first.
disposition-schema
Use when dispatching parallel agents to verify or adjudicate findings (adversarial verify, security disposition, council fires), or before you consolidate their outputs and risk an ID collision.
decision-rights
Use when writing a handoff, plan, ADR or subagent brief, or one says "do not re-litigate" or "final". Separates evidence access from decision rights and marks what is non-negotiable.
dead-predicate
Use when a hook didn't fire, a reminder never arrived, or before claiming a rule is hook-enforced: a live router rule can match nothing (missing word, JSON escape, shadowing, stale name).
vacuous-check
Fail-open success checks. Use when a script says OK but nothing happened, a probe batch returns a false negative or 0 results, or `String(got) === String(want)` passes on a type mismatch.
stale-deploy
Use when asking "is it live", a deploy looks stale, or a poll loop exits instantly. Verify a GitHub Pages or CDN deploy serves new content: poll a marker new to this deploy.