← Documentation home

Canonical Markdown source · Oct 20, 2018

AI agent machine-scorecard projection — technical recommendation

technical-recommendations/ai-agent-machine-scorecards-2026-08-25.md · 41 lines · SHA-256 54ddf67be5a4

Impact summary

`ai-agent-machine-evidence.ts` now owns fail-closed projection of verified public

trial receipts into the canonical AI-value audit. Calliope v10 and Clio evidence-

hook v6 are replayed against their exact fixtures and existing strict parsers

before the audit reports observations, mean model cost per system task, mean

latency, safety failures, and aggregate machine-gate state. Both scorecards move

from 0% to 60% completeness while quality lift and reviewer effort remain null;

their grade stays `unproven`. Runtime behavior, provider spend, deployment, and

publication authority are unchanged.

Ranked coupling and failure findings

  1. High — audit/readiness verification duplication: readiness and audit now

invoke the same strict parsers through separate orchestration paths; a future

receipt-version change could update one path but not the other.

  1. High — aggregate latency cannot be independently recomputed: privacy-safe

receipts retain min/mean/max but not individual timings, so the parser proves

ordering and ceilings rather than recalculating the mean from raw observations.

  1. Medium — 60% is measurement completeness, not value: machine gates can all

pass while blinded quality or reviewer effort fails; UI and docs must preserve

that distinction.

Actions

| Owner | Action and acceptance criteria | Validation |

|---|---|---|

| AI Reliability | Extract a versioned shared receipt registry used by readiness and audit; adding or superseding a receipt changes one registration only | Focused readiness, machine-evidence, and audit tests pass with no duplicated canonical path constants |

| AI Evaluation | Collect two genuine packet-bound responses for Calliope and Clio; populate quality and effort only through the strict aggregator | Run `pnpm audit:ai-agent-value:blind-review:aggregate` on each exact private packet/key and two distinct complete responses |

| Trust | Keep scorecard completeness visibly separate from proven benefit; no 60% machine row may change grade or authority | `node --import tsx --test --test-concurrency=1 tests/services/ai-agent-machine-evidence.test.ts tests/services/ai-agent-value-audit.test.ts` |

| Observability | If privacy review permits, add a bounded latency histogram or digest so aggregate mean can be replayed without retaining prompts, outputs, or reviewer data | Add a failing receipt-mutation test, update both trial parsers, and rerun provider-free replay tests |

Next-cycle hypothesis

A single versioned evidence registry can remove audit/readiness receipt-path drift

without weakening strict parser replay or changing any grade. Falsify by mutating

a fixture hash, cost, or registered path and running the focused readiness plus

machine-scorecard suites; every mutation must fail or remove evidence credit.