forming-cybernetic-commitments

name	forming-cybernetic-commitments
description	Use when a formed task needs durable human-approved meaning, completion obligations, evidence claims, non-goals, or live/destructive authority boundaries.

Forming Cybernetic Commitments

Purpose

Create the only human-approved artifact used by a v2 controlled run:

docs/cybernetics/runs/<slug>/commitment.control.json

This stage forms a commitment, not a solution, plan, file/command choice, runtime JSON, or target execution.

Read references/outcome-fidelity-frontier.md when the requested result has a material fidelity, cost, or user-inference tradeoff.

For post-draft feedback, run the ephemeral check in references/human-review-feedback.md before review.

Commitment Body

Capture only:

• atomic authoritative sources with an inline exact quote, an optional locator reference, and one disposition per binding source;
• one objective;
• blocking obligations, each with source links, required state, observable done conditions, result places, exact evidence claim, and rejected substitutes;
• non-goals;
• concrete action classes and target scopes that require fresh approval;
• forbidden actions;
• an output contract only when the user explicitly made it part of the result.

For a stateful user-visible or protocol result, add observable_paths only for the externally meaningful transitions that completion must prove. Do not map implementation internals or work packages.

Keep design, planning, file/command choice, delegation, review mechanics, and strategy details agent-owned.

Outcome Fidelity Frontier

Before drafting, use current facts to compare paths only when fidelity/cost has a meaningful tradeoff and no binding source already requires exact behavior. Compare user value, truthful visible claim and inference risk, implementation / verification / operations cost, evidence confidence, and upgrade path.

State a highest-fidelity reference; derive a balanced or minimum truthful path when useful; remove dominated paths. Do not invent numerical value/cost scores. Ask the user only when remaining paths materially differ in visible semantics. Otherwise recommend one. Candidate paths are temporary analysis, not commitment fields. Translate the selected path only into observable obligations, evidence claims, and rejected misleading substitutes.

Pre-Commitment Semantic Convergence

Read ../_shared/counterexample-gate-protocol.md before executing the Gate.

Before asking for approval:

1. preserve direct user requirements and baseline source coverage;
2. classify post-draft feedback and run its projection check;
3. evaluate the Outcome Fidelity Frontier when material, using safe facts;
4. resolve semantic debts in objective, obligations, non-goals, evidence, or authority while keeping implementation choices agent-owned;
5. record resolved debts and accepted assumptions in a hash-bound ledger under evidence/, including the resolution and its basis_ref rather than only a status label;
6. hash the canonical commitment body;
7. execute that falsification Gate against the exact hash and bind its independent review evidence and ledger.

Repair and rerun on an established counterexample. Do not ask the user to approve a draft with a pending or failed Gate.

After every draft mutation, run:

python3 .agents/skills/using-commitment-control-json/scripts/commitment_runtime.py \
  state docs/cybernetics/runs/<slug>

Follow next_action. RunPreCommitmentSemanticConvergence, RunPreCommitmentCounterexampleReview, and ReviseCommitmentDraft are agent-owned: continue in the same turn and do not emit a user prompt. RequestHumanSemanticDecision asks only for the named ledger debt or assumption. RequestHumanCommitmentApproval shows the exact commitment hash, obligations, non-goals, and authority boundaries.

Ask the user only for a real semantic decision, unavailable required source, or the final commitment approval. Safe reads, review execution, and draft repair are agent-owned actions.

Approval

Show the user exactly:

• objective;
• obligations and done conditions;
• non-goals;
• for post-draft feedback, preserved baseline, changed targets, explicit additions, and feedback not promoted into the commitment;
• fresh approval and forbidden action boundaries;
• remaining assumptions, if any;
• what approval authorizes: runtime execution against this commitment hash.

After explicit approval, set status and approval.status to approved, bind both approved_commitment_hash and approved_semantic_ledger_hash, and validate:

python3 .agents/skills/using-commitment-control-json/scripts/commitment_runtime.py \
  validate docs/cybernetics/runs/<slug>

Then print this literal handoff and stop:

/goal Use .agents/skills/using-commitment-control-json and execute docs/cybernetics/runs/<slug>. Read commitment.control.json as the only approved commitment. 允许使用 subagent 执行独立 counterexample review。Continue automatically for agent-owned strategy and implementation work; stop only for a typed human decision (then form a successor commitment), fresh authority grant, external dependency, proved impossibility, or user abort.

Do not execute the runtime in the same commitment-forming turn.

Reviewer-attestation limits are defined by the shared Counterexample protocol.