Intent-to-Evidence Work Slice Extension — Minimum Contract

Purpose

Allow a relevant Work Slice to separate the human-owned desired outcome from verifiable expectations before planning or execution. The extension reduces situations where an agent silently decides what “correct” means.

Boundary

This is an optional extension inside the existing Work Slice and Spec Bundle. It is not a lifecycle, PRD, SDD, execution plan, runtime policy or replacement for readiness gates.

Use it for Standard or High-Assurance work when ambiguity could change scope, risk, architecture or validation. Lite work does not use it by default.

Ownership

• The human owner confirms Intent and Expectations.
• AletheIA governs proportionality, readiness, evidence and closure.
• Adaptive Skills may clarify or identify gaps but cannot confirm intent.
• The harness executes within the confirmed boundaries.
• Evidence proves expectations; Reconcile evaluates whether intent was preserved.

Minimum references in a Work Slice

intent_to_evidence:
  intent_ref: <record ref or not_needed>
  expectations_ref: <record ref or not_needed>
  guessing_risk_ref: <review ref or not_needed>
  evidence_map_ref: <map ref or pending>
  proportionality: required | optional | not_needed

The references point to records; they do not copy large specifications into the Work Slice.

Intent Record

An Intent Record states the outcome, not its implementation:

intent_record:
  intent_id:
  human_owner:
  desired_outcome:
  constraints: []
  success_scenarios: []
  failure_scenarios: []
  connected_intents: []
  non_goals: []
  ambiguity_markers: []
  decisions_required_before_execution: []
  owner_confirmation:
    status: draft | confirmed | needs_review
    confirmed_by:
    evidence_ref:

An agent may draft questions and candidate wording. It may not set confirmed on behalf of the human owner.

Expectations Contract

Expectations define observable success, failure and boundaries without prescribing architecture unless architecture is an explicit constraint:

expectations_contract:
  expectations_id:
  linked_intent:
  success_conditions: []
  failure_conditions: []
  boundaries:
    must: []
    must_not: []
  acceptance_examples: []
  rejection_examples: []
  validation_methods: []
  stop_conditions: []
  human_review:
    required:
    review_focus: []

The execution plan may reference expectations but may not silently redefine them.

Agent Guessing Risk

agent_guessing_risk:
  level: low | medium | high
  holes:
    - field:
      reason:
      required_action: clarify | accept_assumption | stop
  verdict: ready | needs_clarification | stop

• low: gaps do not affect scope, risk or success.
• medium: approach-affecting assumptions must be explicit and reviewable.
• high: intent, success, failure, scope or risk is unresolved; execution stops.

A ready verdict is invalid when risk is high, intent is unconfirmed or a stop-condition decision remains open.

Evidence and Reconcile

Each relevant expectation maps to source-backed evidence with one status: proven, partially_proven, not_proven or not_applicable.

Closure records:

• overall verdict: pass, pass_with_gaps, fail or review_required;
• whether intent was preserved: yes, partially or no;
• deviations, missing evidence and accepted limitations;
• human review reference when required;
• learning or documentation follow-up.

Technical tests alone cannot prove human intent. Missing evidence remains missing.

Presence in the loop

Presence checkpoints are proportional. Use them before a first write, structural change, large diff or closure only when the risk and review burden justify interruption. Late approval of an unreadable diff is not a substitute for earlier clarification.

Privacy

Records store concise outcomes, constraints, classifications and references. Do not store private prompts, secrets, personal data or restricted source content.