Harness Contract and Lifecycle

Key jargon

Term	Plain-language meaning
Contract	An explicit agreement about inputs, outputs, errors, permissions, and invariants.
Lifecycle	The stages from request admission through completion, cancellation, or recovery.
Invariant	A condition that must remain true at every relevant stage.
Version pin	A recorded dependency, prompt, model, schema, or policy version used for a run.

Key concepts

The lifecycle should make every state transition and terminal outcome explicit.
Contracts let evaluation target the complete system instead of treating a fluent response as success.

Concept map

flowchart LR
    A["Admit validated request"] --> B["Execute versioned lifecycle"]
    B --> C["Enforce invariants"]
    C --> D["Close with result and evidence"]

Lifecycle

Scope: user, task, harm, data, action, and stop boundaries.
Baseline: representative examples and current non-AI performance.
Prototype: smallest prompt/model/workflow that can test value.
Evaluate: quality, safety, latency, cost, and subgroup behavior.
Harden: permissions, isolation, validation, recovery, and logging.
Pilot: limited users/data/actions with rollback.
Operate: monitoring, incidents, versioning, and human escalation.
Change: regression gates for model, prompt, tool, retrieval, or policy.
Retire: revoke access, handle memory/data, preserve required evidence.

Harness contract

Write down purpose, non-goals, owner, users, data classes, models, tools, memory, action inventory, evaluation suite, budgets, approval gates, incident owner, retention, and retirement trigger.

Definition of done

A demo that produces a good answer is not production readiness. Done means the system meets a documented contract on representative tests and fails within known boundaries.

Dumpster-derived lesson

The local research repeatedly supports artifact-backed, stage-level evaluation: fluent final output must not substitute for trace, state, or effect evidence.