szl-governed-decision
Make decisions that carry their own evidence. Wrap any classifier, policy engine or System One model in a fail-closed record that names why it refused, cites a digest rather than the text, and never asserts a number without a receipt on disk. Use when a decision must survive audit, when a model's output should be advisory rather than authoritative, or when a claim needs a state - MEASURED, BLOCKED, UNVERIFIED, UNAVAILABLE, NOT_CLAIMED - instead of a confident sentence. Complements typed-primitive skills: they produce the judgement, this governs what the judgement is allowed to do.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-15
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- szl-governed-decision
- description
- Make decisions that carry their own evidence. Wrap any classifier, policy engine or System One model in a fail-closed record that names why it refused, cites a digest rather than the text, and never asserts a number without a receipt on disk. Use when a decision must survive audit, when a model's output should be advisory rather than authoritative, or when a claim needs a state - MEASURED, BLOCKED, UNVERIFIED, UNAVAILABLE, NOT_CLAIMED - instead of a confident sentence. Complements typed-primitive skills: they produce the judgement, this governs what the judgement is allowed to do.
02 资源文件 2
- 脚本scripts/check_claims.py688 B
- 脚本scripts/decide.py1.3 KB
03 SKILL.md 正文
Governed decision
A decision is not a label. It is a record: what was decided, by which path, on what evidence, and what remains unproven.
The four refusal paths
Every decision reports exactly one.
PRE_AGGREGATION— a guard fired before any score existed. Sub-reason required:INJECTION_GUARD(adversarial pattern matched a versioned policy list, the model was not consulted),META_CUE(an enumerated handling cue), orUNIDENTIFIED.ZERO_PINNED— scores were computed and one was exactly zero, so the weighted geometric aggregate is zero. Weakest-link, not average.BELOW_THRESHOLD— all scores positive, aggregate under the floor.NONE— a confident decision was issued.
UNIDENTIFIED is not a fallback for convenience. If a refusal happens and no known guard
explains it, record UNIDENTIFIED and go find the guard. Naming a mechanism you have not
located in the code is a guess wearing a label.
Never default a missing score
Asking for a score vector on a decision that computed no scores must raise, not return zeros. In a zero-pinned aggregate, a substituted zero is indistinguishable from a genuine veto — it manufactures the exact condition the system treats as fatal.
The model provider contract
A model attached to this kind of decision may:
- lower one axis, never raise it — take the minimum, never the maximum;
- justify a lowering with a span that occurs literally in the input, verified before the score is honoured;
- return nothing at all, in which case the axis is untouched and the reason is recorded.
It may not return a label. Under zero-pinning this makes a bad provider incapable of causing a confident error: the worst it achieves is over-refusal, which shows up immediately as churn against a control corpus.
Consult the provider only on decisions that reach the aggregator. A decision already refused by a guard must never be credited to the provider.
Claim states
Attach one to every assertion, and generate prose from the states rather than writing prose and hoping it matches.
| state | meaning |
|---|---|
| MEASURED | a receipt exists on disk and the number was read from it |
| UNRATIFIED | rests on labels no human has ratified |
| DEGRADED | works, below the standard you hold yourself to |
| BLOCKED | a negative result, recorded rather than omitted |
| UNVERIFIED | unmeasured; a probability is a claim until calibration measures it |
| UNAVAILABLE | no evidence exists; never rendered as health |
| NOT_CLAIMED | a deliberate refusal to assert |
Rules that hold without exception
- Carry
sha256(input), never the input, into any receipt, span or telemetry attribute. - Never fabricate a measurement you did not take. Energy is
null, not0. - An unsigned receipt says
signed: false. It renders as UNAVAILABLE, never as a pass. - Derive ids from content, not from clocks or randomness, so re-running produces no churn.
- Pin model versions. A moving
-latestsilently changes answers under a tuned threshold. - Fix acceptance criteria before running the experiment, so they cannot be rewritten to match the result.
- When a claim turns out wrong, add a correction with its receipt. Do not edit the old one away.
Scripts
scripts/decide.py— classify text, print the governed record and a receipt payload. It needs a szl-typesafe-triage checkout as its repository root.scripts/check_claims.py— verify every claim in a ledger cites a file that exists.
04 同仓库更多 Skills
szl-unit-invariants
Checks declared SI dimensions, allowlisted unit conversions, expected ranges and fixed pointwise invariants (equal, sum, difference, product, ratio) over selected measurement or simulation columns, flagging incompatible quantities, scale errors of 1000x and broken conservation checks. Use when merging instrument exports with different units, validating simulation output against conservation laws, or when a reviewer asks how unit errors were ruled out. Not an uncertainty-propagation or symbolic tool.
szl-uncertainty-lineage
Propagate declared measurement uncertainty through a small arithmetic dependency graph using a first-order delta approximation. Use when a researcher needs to see how corrected inputs or shared-source correlation change a reported standard uncertainty; require unit checks separately.
szl-typesafe-ai
TypeSafe/Jev: Choice/Noul/Score in code, not TypeScript types. System One models turn natural language and application state into probabilities that software can compose. Use when building routing, ranking, extraction, verification, or replacing an LLM prompt-and-parse step with structured decisions; also when brainstorming what TypeSafe could make possible. Do not use for TypeScript/Java/Rust static types, Zod/Pydantic/mypy/JSON Schema/protobuf codegen, or asking whether a pipeline is type-safe. At SZL, Jev is an optional second reader only.
demo
暂无简介
demo
暂无简介
szl-skill-update-review
Compare pinned local skill packages using inventories and a retained lock; report changed skills, files, declarations and reruns without approving safety or validity.