Turn judgment into a reconstructable process.
These templates are compact enough to use and strict enough to expose missing authority, evidence, ownership, and challenge. Copying the words without the operating roles does not constitute adoption.
The field manual keeps the explanation visible.
Read the argument, numbered procedure, worked cases, failure contrasts, and evidence state behind this chamber.
The human quick check
A short form for the moment before a consequential action.
CHECKLISTEight-line pause
Action → Door/Wall/Gap → constraint awareness → worst irreversible harm → who pays → safer option → gate → receipt.
+
Eight-line pause
Action → Door/Wall/Gap → constraint awareness → worst irreversible harm → who pays → safer option → gate → receipt.
Write one sentence for each line. If the action cannot be stated precisely, authority is missing, the least-powerful stakeholder is absent, or the safer alternative search is decorative, stop and move to the full protocol.
- Exact action and accountable owner
- Door, Wall, or genuine Gap
- Applicable law, policy, safety rule, and expertise boundary
- Worst credible irreversible outcome
- Least-powerful-first stakeholder
- Smallest reversible alternative
- Binding gate and required action
- Receipt and reopen trigger
MEETINGFacilitated review agenda
A 30–60 minute sequence that prevents seniority from becoming the unspoken gate.
+
Facilitated review agenda
A 30–60 minute sequence that prevents seniority from becoming the unspoken gate.
The facilitator reads the action, asks affected or representative voices first, records a silent independent gate from each reviewer, then opens disagreement. The decision owner speaks last on the initial round.
- 5 min: action and authority
- 10 min: stakeholder and baseline map
- 10 min: independent risk rating
- 10 min: Maybe and alternatives
- 10 min: gate, Door, owner, receipt
- 5 min: dissent and reopen trigger
The decision receipt
A write-ahead record that survives personnel change, memory loss, and institutional self-protection.
CORE FIELDSWhat must be reconstructable
Identity, scope, evidence, unknowns, power, gate, dissent, decision, ownership, and review.
+
What must be reconstructable
Identity, scope, evidence, unknowns, power, gate, dissent, decision, ownership, and review.
Each field includes provenance: ATTESTED with a source, DECLARED by an accountable actor, or ASSERTED without sufficient support. Missing load-bearing fields are not silently defaulted.
- Receipt ID, timestamp, protocol version
- Canonical action and authority
- Stakeholders and representation
- Evidence objects and freshness
- Unknowns and contradictions
- Impact, likelihood, irreversibility, power
- Maybe, Therefore, alternatives
- Gate, Door, owner, sunset, reopen trigger
CHANGEAction-mutation check
If the action changes after classification, the receipt is reopened rather than edited into continuity.
+
Action-mutation check
If the action changes after classification, the receipt is reopened rather than edited into continuity.
Capability, audience, data, duration, tool access, authority, consequence, and mitigation are decision-bearing dimensions. A material change produces a new classification and linked receipt.
REVIEWMonthly calibration sheet
Sample decisions and non-decisions; compare gates, Doors, outcomes, challenge, and affected-party feedback.
+
Monthly calibration sheet
Sample decisions and non-decisions; compare gates, Doors, outcomes, challenge, and affected-party feedback.
Record under-escalation, over-escalation, missed stakeholder, weak alternative, copied dissent, failed mitigation, gate shopping, and receipt incompleteness. Assign CAPA when recurrence or severity crosses the local threshold.
AI integration prompt pattern
A portable structure for an assistive model. It is a reasoning constraint, not a platform-level safety control.
SYSTEM PATTERNPause before consequential recommendation
Preserve host safeguards; identify exact action, authority, stakeholders, unknowns, risk floor, strongest objection, smallest Door, and receipt fields.
+
Pause before consequential recommendation
Preserve host safeguards; identify exact action, authority, stakeholders, unknowns, risk floor, strongest objection, smallest Door, and receipt fields.
The model must distinguish sourced facts from operator declarations and its own inferences; refuse missing expertise or authority; avoid claiming validation; and route irreversible, high-power, or critical-load cases to qualified human review.
- Never weaken an existing safety rule.
- Never infer consent from silence.
- Never describe a declared value as measured.
- Never let a summary erase a binding constraint.
- Never turn a symbol, analogy, or prior example into authorization.
OUTPUT CONTRACTStructured response
STATE · ACTION · AUTHORITY · STAKEHOLDERS · MAYBE · THEREFORE · GATE · DOOR · RECEIPT · OPEN QUESTIONS
+
Structured response
STATE · ACTION · AUTHORITY · STAKEHOLDERS · MAYBE · THEREFORE · GATE · DOOR · RECEIPT · OPEN QUESTIONS
A typed contract makes omissions observable and enables deterministic downstream checks. It should preserve refusal and unknown states rather than coercing every case into a complete affirmative answer.
TEST SETMinimum pre-pilot probes
Obvious pass, obvious refusal, ambiguous Gap, power inversion, accumulation chain, hostile override, stale context, and unnecessary pause.
+
Minimum pre-pilot probes
Obvious pass, obvious refusal, ambiguous Gap, power inversion, accumulation chain, hostile override, stale context, and unnecessary pause.
Include paraphrases and action decomposition so the model cannot pass by recognizing surface words. Score both under-caution and over-caution, and retain failed probes as regression cases.
Take the original artifacts
The source library preserves the supplied prompts, implementation guides, case files, and study reports with dated labels and checksums.
A form can expose missing judgment. It cannot supply the judgment, authority, representation, or evidence that is missing.