Patch 1.2 · Section 6
Confidence never creates authority.
The minimum operating contract for all twelve AI employees: five authority levels, five confidence bands, and routing decided outside the model. A 99% score that conflicts with an expired credential loses to the credential.
Section 6.4 · Routing simulator
Pick an authority level and a confidence band. The rule decides, not the score.
Requested authority level
Reported confidence band
Deterministic routing outcome
C3 may be presented as a recommendation with evidence. Material workflow effects remain a human decision and hard gates still apply.
L2 permitted behavior
Recommend an option with evidence and confidence; cannot commit state.
C3 permitted handling
May be presented as a recommendation or accepted for low-risk reversible work within authority; hard gates still apply.
Confidence may narrow permitted handling. It may never expand authority.
Section 6.2 · Authority levels
What an AI employee may ever be asked to do
Observe
Read only within authorized scope; no workflow output.
e.g. Summarize history, identify missing context.
Allowed with logging and tenant controls.Draft / Extract
Produce structured extraction, classification or draft; no material recommendation.
e.g. Extract VIN field, draft status message.
Human or deterministic validation as the task requires.Recommend
Recommend an option with evidence and confidence; cannot commit state.
e.g. Rank eligible vendors, flag document inconsistency.
Human decision required for material workflow effects.Bounded assist
Invoke explicitly allowlisted reversible tools that do not alter protected state, only after deterministic validation.
e.g. Create a draft task, request missing information.
Disabled unless individually approved and tested.Protected execution
Dispatch, release, eligibility override, permission grant, financial posting, legal stop removal.
e.g. Any protected state or action.
Prohibited for AI in Trial 1.Section 6.3 · Confidence bands
Percentages are not assumed to be calibrated probabilities
C0
Invalid / unavailable
Default interpretation
No usable output: timeout, policy refusal, malformed schema, missing evidence or model outage.
Permitted handling
No reliance. The deterministic path continues safely; retry, queue or human handling; alert where operationally material.
C1
Low / conflicting
Default interpretation
Material uncertainty, conflicting sources, poor input quality, out-of-distribution case or unsupported claim.
Permitted handling
Abstain from recommendation; human review mandatory; preserve the uncertainty reason and evidence.
C2
Moderate
Default interpretation
Plausible output but the task threshold is not met or evidence is incomplete.
Permitted handling
May assist triage or drafting; human verification mandatory before material use.
C3
High
Default interpretation
Task-specific threshold met and required evidence and validators agree.
Permitted handling
May be presented as a recommendation or accepted for low-risk reversible work within authority; hard gates still apply.
C4
Very high / verified
Default interpretation
Calibrated threshold met, deterministic validations pass, evidence complete, no conflict, task within the approved envelope.
Permitted handling
May reduce review only where the task contract explicitly allows. Never grants protected authority.
Until calibration is demonstrated, the surface displays qualitative status and routes conservatively.
Section 6.4 · Routing rules
Eight rules that keep the score honest
Define thresholds by employee and task. A ZIP-code extraction threshold cannot be reused for custody, bankruptcy, release, eligibility or legal interpretation.
Confidence must fall or abstain when required evidence is missing, sources conflict, input is degraded, policy checks fail or the case is outside the evaluation distribution.
Use deterministic validation where possible: schema, checksum, VIN format, date range, required fields, allowlists, policy lookup and cross-source agreement.
For high-impact recommendations, require source evidence, uncertainty reasons, alternatives considered and the exact rule and credential snapshot.
Never convert a model's self-reported certainty into authority. Calibration compares historical outputs with verified ground truth.
Human reviewers must see the recommendation, evidence and uncertainty before deciding; hard-gate results are displayed visually separate to avoid automation bias.
Disagreement is a learning signal, not automatically a human or AI error. Capture the reason taxonomy and later adjudication.
If confidence telemetry, model identity, policy version or evidence linkage is missing, treat the output as C0 for material use.
Section 6.1 · Operating contract
Fourteen fields, versioned, before an employee touches a Trial path
Required per AI employee
- Employee name, business purpose and accountable human owner
- Allowed task types and explicit prohibited tasks
- Permitted data classes, tenants and sources
- Approved tools and tool-level permissions
- Permitted outputs: extract, classify, summarize, recommend, draft or act
- Protected actions that always require a human or are never allowed
- Input/output schema and deterministic validation
- Evidence and citation requirements and conflict handling
- Task-specific confidence bands and calibration dataset
- Human escalation conditions and service-level expectation
- Outage, timeout, malformed-output and kill-switch behavior
- Model, provider, prompt/policy and evaluation versions
- Monitoring: accuracy, abstention, override, drift, latency, cost and safety events
- Retention, privacy, vendor processing and incident obligations
Section 6.5 · Mandatory AI decision record
Every material AI output leaves a reconstructable record
| Field | Requirement |
|---|---|
| Inputs / evidence | Source record and document versions, hashes or durable references; input-quality flags. |
| Output | Structured recommendation, extraction or draft, and alternatives where applicable. |
| Confidence | Raw score if meaningful, qualitative band, threshold version and uncertainty or abstention reason. |
| Authority | Employee authority level, requested tool or action, and whether human approval is mandatory. |
| Controls | Deterministic validators and hard gates evaluated, with outcomes. |
| Human outcome | Reviewer, decision, time, override or disagreement reason and subsequent adjudication. |
| Operational result | Committed business effect, or explicit statement that no protected state was changed. |
If confidence telemetry, model identity, policy version or evidence linkage is missing, the output is treated as C0 for any material use — regardless of how certain the model reported itself to be.