@metaharness/workspace-probe

Evaluation + Darwin-Mode bridge for @metaharness/workspace-lens. Turn Jacobian-Lens interpretability receipts into (1) a flywheel-consumable workspace_probe score and (2) Darwin-Mode mutation evidence that rejects structurally-brittle prompt/policy mutations.

Pure · dependency-light (only the workspace-lens types) · deterministic · $0.

Why

@metaharness/workspace-lens reads the model's internal verbalizable workspace into a WorkspaceLensReceipt per decision. This package projects a set of those receipts into evaluation signal:

• workspaceProbeScore(receipts) → { score, meanDrift, flagRate, criticalRate, cleanFraction }. score (= cleanFraction) is the fraction of decisions whose workspace was clean — no critical safety trigger AND drift below threshold. A harness/policy that makes the model hold steadier, safer intermediate concepts before answering scores higher. Use it as a workspace_probe Score dimension.

• gradeMutationByWorkspace(baseline, mutant) → { keep, reasons, baseline, mutant }. A Darwin-Mode veto: reject a mutation that improves the final answer at the cost of the internal process — it raises the critical-trigger rate, destabilizes the workspace (drift up), or lowers the clean fraction. "Final answer up, workspace grip down" = structurally brittle. Pair with the usual gold/final-answer gate: keep a mutation only if both pass. It never weakens the answer gate.

Quickstart

import { buildReceipt } from '@metaharness/workspace-lens';
import { workspaceProbeScore, gradeMutationByWorkspace } from '@metaharness/workspace-probe';

const receipts = decisions.map(d => buildReceipt(lens, d.prompt, d.states, { createdAt: d.ts, concepts }));

const probe = workspaceProbeScore(receipts, { driftThreshold: 0.25 });
// → { n, score, meanDrift, flagRate, criticalRate, cleanFraction }

const verdict = gradeMutationByWorkspace(baselineReceipts, mutantReceipts);
if (!verdict.keep) rejectMutation(verdict.reasons);   // structurally brittle

CLI

For audit workflows that don't want to write TypeScript. All inputs are JSON, all output is JSON on stdout (composes with jq):

npx @metaharness/workspace-probe diag           lens.json
npx @metaharness/workspace-probe readout        lens.json activations.json [--top-k N]
npx @metaharness/workspace-probe probe          receipts.json [--drift-threshold F]
npx @metaharness/workspace-probe grade-mutation baseline-receipts.json mutant-receipts.json

• diag → the fitted lens's metadata (model/lens id, dModel, fitted layers, vocab size).
• readout → the workspace tokens per activation ({layer, position, h}[]).
• probe → the workspaceProbeScore over a receipt set.
• grade-mutation → the keep/veto verdict comparing two receipt sets.

Exit codes: 0 ok, 2 usage error, 1 runtime error. A fitted lens artifact is produced out-of-band (open-weight model + backward pass — see @metaharness/workspace-lens).

Runnable demo

examples/demo.mjs is a $0, no-model end-to-end walk-through (synthetic lens → readout → receipts → probe score → mutation veto):

node packages/workspace-probe/examples/demo.mjs

Links

• workspace-lens (the primitive): https://www.npmjs.com/package/@metaharness/workspace-lens
• MetaHarness: https://github.com/ruvnet/metaharness

License

MIT.