brevitas
Enforce evidence-preserving structural output budgets on engineering prose. Apply automatically to chat answers and written drafts containing audit findings, security or diff review, gas analysis, `invariant` discussion, protocol analysis, or specification commentary, and on explicit $brevitas invocation. Govern volume, structure, and connective prose only. Do not apply to code comments, commit messages, or completeness-oriented specification documents.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-07
- Stars
- ★ 13
- Forks
- 6
01 前言 Front Matter
- name
- brevitas
- metadata
- { "version": "0.4.0" }
- description
- Enforce evidence-preserving structural output budgets on engineering prose. Apply automatically to chat answers and written drafts containing audit findings, security or diff review, gas analysis, `invariant` discussion, protocol analysis, or specification commentary, and on explicit $brevitas invocation. Govern volume, structure, and connective prose only. Do not apply to code comments, commit messages, or completeness-oriented specification documents.
02 资源文件 38
- 参考文档agents/openai.yaml207 B
- 参考文档DEMONSTRATION.md2.3 KB
- 参考文档evals/cases/fixtures/held-diff-review/output-gpt-5.6-sol.md1.8 KB
- 参考文档evals/cases/fixtures/held-diff-review/output-gpt-5.6-terra.md1.3 KB
- 参考文档evals/cases/fixtures/held-diff-review/prompt.md724 B
- 参考文档evals/cases/fixtures/held-diff-review/source.md2.6 KB
- 参考文档evals/cases/fixtures/held-gas/output-gpt-5.6-sol.md3.5 KB
- 参考文档evals/cases/fixtures/held-gas/output-gpt-5.6-terra.md3.8 KB
- 参考文档evals/cases/fixtures/held-gas/prompt.md786 B
- 参考文档evals/cases/fixtures/held-gas/source.md3.1 KB
- 参考文档evals/cases/fixtures/held-invariant/output-gpt-5.6-sol.md2.5 KB
- 参考文档evals/cases/fixtures/held-invariant/output-gpt-5.6-terra.md2.2 KB
- 参考文档evals/cases/fixtures/held-invariant/prompt.md720 B
- 参考文档evals/cases/fixtures/held-invariant/source.md3.0 KB
- 参考文档evals/cases/fixtures/held-solidity-auditor/output-gpt-5.6-sol.md2.0 KB
- 参考文档evals/cases/fixtures/held-solidity-auditor/output-gpt-5.6-terra.md1.6 KB
- 参考文档evals/cases/fixtures/held-solidity-auditor/prompt.md761 B
- 参考文档evals/cases/fixtures/held-solidity-auditor/source.md2.2 KB
- 参考文档evals/cases/fixtures/held-x-ray/output-gpt-5.6-sol.md4.0 KB
- 参考文档evals/cases/fixtures/held-x-ray/output-gpt-5.6-terra.md3.7 KB
- 参考文档evals/cases/fixtures/held-x-ray/prompt.md808 B
- 参考文档evals/cases/fixtures/held-x-ray/source.md4.9 KB
- 参考文档evals/cases/fund-safety-evidence-exception/case.json280 B
- 参考文档evals/cases/fund-safety-evidence-exception/original.md3.0 KB
- 参考文档evals/cases/fund-safety-evidence-exception/target.md3.1 KB
- 参考文档evals/cases/solidity-partial-recovery/case.json245 B
- 参考文档evals/cases/solidity-partial-recovery/original.md680 B
- 参考文档evals/cases/solidity-partial-recovery/target.md647 B
- 参考文档evals/cases/xray-replay-authority/case.json256 B
- 参考文档evals/cases/xray-replay-authority/original.md955 B
- 参考文档evals/cases/xray-replay-authority/target.md663 B
- 参考文档evals/corpus.json55.1 KB
- 参考文档evals/README.md5.1 KB
- 参考文档EVOLUTION.md4.4 KB
- 其他Makefile527 B
- 脚本scripts/brevitas.py17.7 KB
- 脚本scripts/held_corpus.py34.6 KB
- 脚本scripts/run_evals.py3.0 KB
03 SKILL.md 正文
<p align="center"> <img src="../../assets/characters/brevitas.png" width="1200"> </p>
Brevitas
Frontier
Brevitas owns the engineering-prose structure frontier, not Hexaemeron's delivery or Solidity frontier. EVOLUTION.md records its version, held target, next job and maturity state. Do not run or recommend another frontier pass once that ledger is mature.
<!-- marketplace-context:start -->
Where this sits
Brevitas enforces mechanical volume and structure limits on engineering prose while protecting every item of evidence that can change the decision.
Current frontier. The linter has not been forward-tested across a held cross-model corpus of engineering reviews, and preservation of counterexamples and reproduction steps remains agent-checked. <!-- marketplace-context:end -->
Apply this after Imprimatur has checked the vocabulary and Vulgate has settled the register. Sapheneia may separately shape the surrounding interaction or a bounded durable record. Brevitas changes volume, structure, and connective prose only; it does not alter word choice, voice, AuDHD presentation, or the facts another skill established.
Synkrisis renders a comparison report from fixed templates, and this skill may constrain its engineering-prose structure without validating the finding. Rendering refuses a finding carrying causal language, so neither skill can introduce a claim the findings do not hold.
Precedence
Preserve evidence before satisfying any budget. Never delete or weaken:
- addresses, transaction hashes, 4-byte selectors, bare SHA-256-shaped digests, or Git object ids;
file:linereferences or numeric claims;- concrete counterexamples or reproduction steps; or
- an explicit statement that a fact, property, or conclusion could not be established.
If evidence does not fit, the budget yields. Cut prose, never evidence. Mark an irreducible over-budget finding with the evidence exception below.
Compose
-
Answer immediately. Do not restate the request or introduce the answer.
-
Apply these physical-line budgets:
- Direct answer: at most 6 nonblank lines before the first list or code fence.
- Finding: at most 5 prose lines: claim, location, mechanism, impact, fix.
- Code: at most 40 content lines per fence.
-
Give each finding this checkable shape:
[High] Claim. Location: `src/Contract.sol:42` Mechanism: Concrete causal path. Impact: Concrete consequence. Fix: Smallest correction.Use only the severity word; do not add a severity-justification paragraph.
-
Put findings adjacent. Add no transition, preamble, or summary between them.
-
Use a table only with at least 3 data rows and 3 real-data columns.
-
Use headings only when the draft has at least 3 sections. A document title does not count as a section.
-
Use at most one qualifier per claim.
For an irreducible finding, place <!-- brevitas: evidence-exception reason="counterexample requires ordered steps" --> immediately before it.
Use the exception only when compression would remove protected evidence. Keep any extra lines as evidence, counterexample, reproduction, or establishment-limit statements; do not use the exception to retain connective prose.
Delete
Delete request restatements, list preambles, post-list summaries, process narration, bold-label-colon items, trailing offers, unrequested next-step menus and stacked hedges. Delete confidence theatre, including:
importantly; notably; it's worth noting
Do not reprint visible code. Quote only the lines carrying the defect.
Lint
Run the checker before sending substantive chat prose or saving a report. Pipe chat drafts through stdin; pass written work as a file:
python3 scripts/brevitas.py - < draft.md
python3 scripts/brevitas.py report.md
python3 scripts/brevitas.py report.md --source uncompressed.md
make -C /path/to/brevitas lint FILE=/path/to/report.md
Use --mode answer for direct answers and --mode report for reports. auto
infers a report from findings or at least 3 sections. For existing material, always
pass --source; the checker fails if an address, transaction hash, selector,
bare digest, Git object id, file:line reference or numeric token disappears.
Abbreviated Git ids require Markdown code quoting, an explicit Git label, or an
owner/repository@oid form. Fix every diagnostic and rerun for exit 0.
Fiat's canonical audit ledger may use --mode fiat-audit-record only after the
same tree passes audit_synopsis.py --check .. This explicit mode starts from
report behaviour and suppresses only B010 for the schema-owned single H2 and
B011 for its findings table. Every other rule remains active. Brevitas does not
parse or establish the Fiat schema, count findings, cover risks, or establish a
controller receipt; the synopsis checker and Fiat controller retain those jobs.
Host-required status commentary is outside the draft lint boundary. Do not suppress status messages required by the execution environment.
Evals
Run make test. evals/cases/ contains preserved audit, x-ray and security-review
excerpts with source paths, ranges and pinned fixture digests. scripts/run_evals.py
checks fixture integrity, target structure, evidence-token survival, positive-case
compression and exact retention for the evidence-exception case.
Exclusions
Do not apply this skill to code comments, commit messages or specifications where completeness is the point. Do not change lexicon, tone, register or accessibility.
Promise Machine contract
brevitas-structure-check
- Promise: A successful answer-, report-, or explicit Fiat-audit-record-mode lint establishes that the named draft satisfies the applicable Brevitas mechanical line, finding, code-fence, table, heading and qualifier budgets.
- Evidence: The exact draft bytes, selected lint mode, emitted diagnostics and exit status from
scripts/brevitas.py. - Evidence classes: checked
- Boundary: The lint does not establish factual accuracy, evidence completeness, severity correctness, voice, accessibility or fitness for a completeness-oriented specification.
fiat-audit-recordsuppresses only B010 and B011 and is admissible only after the same tree passes Fiat's synopsis check; Brevitas does not parse or establish the audit schema, finding count, risk coverage or controller receipt. - Authorises: Presentation of the checked draft as structurally conformant to the selected Brevitas mode.
- Consequence: 0
- Refuses: Calling prose Brevitas-clean when the checker did not run, reported a diagnostic, inferred the wrong mode or received an excluded document class.
- Recovery: Select the right mode, preserve protected evidence, remove the named structural defect and rerun the checker.
- Exceptions: none
brevitas-evidence-preservation
- Promise: A successful lint with
--sourceestablishes that protected transaction hashes, addresses, 4-byte selectors, bare SHA-256-shaped digests, Git object ids, file-line references and numeric tokens from the named source survive literally in the compressed draft. - Evidence: The exact source and draft bytes, source-token comparison, any explicit evidence exception and a zero-exit lint result.
- Evidence classes: checked, recomputed
- Boundary: Token survival does not establish semantic equivalence, complete reasoning, correct conclusions, that a protected digest, selector or Git object exists or is valid, or preservation of evidence classes that the token checker cannot represent; unlabelled abbreviated hexadecimal text is outside the Git-object promise.
- Authorises: Saving or handing off the compressed draft as a derived artefact whose mechanically protected tokens remain present.
- Consequence: 1
- Refuses: Deleting or weakening protected evidence to meet a volume budget, or claiming preservation without supplying the source comparison.
- Recovery: Restore the missing evidence, use a scoped evidence exception when ordered proof requires it and rerun with the original source.
- Exceptions: none
04 同仓库更多 Skills
example
A deliberately unsupported evidence class fixture.
example
A router fixture target.
promise-machine
A deliberately unresolved router.
example
A deliberately unclassified skill fixture.
example
A deliberately unattributed exception fixture.
example
A deliberately conflated version fixture.