Skip to content

Production Harnesses

Production harnesses are the components G6 has promoted to its production allowlist — the set held to the strictest reliability contract. They are not a separate kind of component: in G6 every component is a harness (see P1: Every component is a harness). "Production harness" simply means a component that has met the full set of contract gates and been added to the curated allowlist that release tooling treats as load-bearing.

What makes a harness a production harness

The allowlist lives in policies/production_harnesses.yaml and grows one component per reviewed change, only once its contract tests are green. To qualify, a component declares a complete BlockContract:

  • Purpose & I/O — what it does and the inputs/outputs it promises.
  • Mitigated failure modes — at least one failure mode it exists to catch.
  • Known failure modes — the new failure modes the harness itself introduces (a verifier can hallucinate too), so the next layer can catch them.
  • State surface — what it persists, where it is queryable, and for how long.
  • Verification — how the harness checks its own output, and whether human review is required.
  • Fallback — the recommended next step when it cannot complete.

These map to the gates C1–C5 documented under Principles → Contract Gates.

Why harnesses exist

A raw LLM call is fast but unaccountable. A harness wraps that call with a contract, verification, evidence, and a recovery path so the result can be routed, audited, and learned from. The design follows the reliability guards G6 holds itself to:

  1. First-principles framing — model the problem and its domain before acting.
  2. Reliable algorithm before execution — name the verification method up front.
  3. Redundancy and grounding — independent checks rather than a single self-assessment.
  4. Failure modes surfaced before execution — a pre-mortem, not a post-mortem.
  5. Reasoning audit — inspectable artifacts instead of hidden chain-of-thought.
  6. Happy paths and long tails — handle the edge cases and the economics, not just the demo path.
  7. Bottleneck elimination — turn a recurring failure into a new harness.

No single harness is sufficient on its own. G6 composes them in a Swiss-cheese arrangement: multiple independent layers, each with different holes, so a failure that slips past one is caught by the next.

Allowlist status is a contract bar, not a safety certification

Promotion to the production allowlist means a component meets G6's contract gates and reliability tests. It is appropriate for MVP and pilot use. It is not, by itself, regulatory compliance or a proof of correctness in a regulated domain — those still require domain-expert review and the launch-readiness checks described in Deployment.

The current production harnesses

Harness What it does
reliability_pipeline Stage-gated pipeline that carries a run from assessment through to a verified, qualified-draft, or blocked-escalated completion.
goal_engine Decomposes high-level goals into guardrailed search trees with resource bounds — the entry point to the problem-solving pipeline.
core Foundation types every other component builds on: the AIBlock base, the railway-oriented Result, guardrailed SearchTree, and resource bounds.
csf Cross-Linked Safety Framework — verifies a bounded agent against a safety-epsilon bound and returns a formal SafetyDecisionReport before LLM calls, file writes, or code execution.
agent_claude Beta, review-pending Anthropic Claude agent with tool-calling loops, retry, circuit breaker, and lifecycle management.
llm_router Routes completions across OpenRouter and Ollama with resource enforcement, model discovery, and hardware-aware recommendations.
context_engine Stateful context tracker over an 8-dimensional context vector; classifies each snapshot as stable, unstable, novel, or fatigued.
autofix Propose-only repair harness: diagnoses reliability incidents and proposes fixes (mitigating false success claims and over-large patches) without applying them.
reliability The reliability machinery's own meta-canary — it declares a contract over itself and persists incidents and scorecards to the ledger.
job_framework Foundation layer for G6's economic job agents: shared base block, per-sector safety profiles, audit trail, and rate limiting.

Each page is generated from the component's live BlockContract, so it reflects the contract in the running code.