paperclip/packages/paperclip-runner/docs/capability-eval-conformance.md

4.2 KiB

Capability Eval-Derived Conformance

Capability turns the Paperclip Evals corpus into an executable, offline conformance suite. The suite derives its cases from the checked-in Capability traceability derivative (spec/capability/eval-traceability.yaml) — it does not clone or read an external eval repository, and it starts only the in-process mock control plane. It hard-fails unless the derivative declares schema version 2, exactly 106 rows, exactly 16 groups, and unique case IDs.

Sources: src/conformance/capability-eval-suite.ts and its test src/conformance/capability-eval-suite.test.ts; the reporter scripts/run-capability-eval-suite.mjs.

The corpus

  • 106 cases across 16 groups. Per-group counts: hb 5, co 6, st 8, cm 6, se 4, su 4, bl 5, dp 3, ix 9, ap 6, ar 4, er 9, rf 22, mh 4, rs 3, wk 8.
  • Each case declares an actor role, a task mode, an input scenario, and the expected outcome, all bound to a row in the capability contract.

Assertion classes

Every case belongs to one assertion class, and each class checks a different kind of invariant:

  • agent_tool_contract — a semantic tool call produces the expected typed effect and state change.
  • authorization_policy — a capability is exposed, denied, or unlocked by a grant exactly as its disposition requires.
  • control_plane_invariant — a control-plane-owned action happens without any agent tool, and no tool can perform it.
  • combined_multi_hop — a sequence of operations across turns produces the expected cumulative state and respects forbidden-operation rules.
  • restraint_no_call — the correct behavior is to make no call; the case passes only if the agent deliberately does nothing further.

Fake-agent matrix and bounded Codex sample

The suite executes each case with a deterministic fake agent whose plan is fixed by a fixture seed, so repeat runs are byte-identical. Optional-tool rows are run twice — once with the unlocking grants (must be allowed and must mutate state) and once ungranted (must be absent or denied with no state change); control-plane-owned rows remain absent in every configuration; and restraint_no_call rows must produce an empty state diff.

The fake-agent surface is 14 always-agent tools plus 4 optional tools unlocked by four seed grants (discovery:tasks:read, discovery:agents:read, delegation:tasks:create, governance:approvals:request) — 18 operations. The suite binds that surface through both the fake-agent and Codex bindings and asserts the two operation lists are byte-identical (18/18).

A bounded Codex binding sample picks one representative case from nine groups (hb, dp, bl, ap, ar, ix, mh, rs, wk) and checks that checkout_task is absent from the Codex surface while every other sampled operation is present. It is an offline parity check; no real Codex or network is contacted, and the browser explorer holds no credential.

Running it

# Run the 106-case suite in-process (one vitest file drives all cases).
pnpm --filter @paperclipai/paperclip-runner test:capability-evals

# Build the public surface and write the parity report with per-group counts,
# assertion classes, the fake-agent matrix, the bounded Codex sample, and the
# semantic-operation execution counts.
pnpm --filter @paperclipai/paperclip-runner report:capability-evals

The reporter writes .paperclip-local/evidence/capability/eval-parity-report.{json,md}.

Run the bounded provider conformance matrix separately. It creates exactly one real Codex turn for each of the 16 checked-in eval groups while retaining the in-process mock control plane:

pnpm --filter @paperclipai/paperclip-runner report:capability-live-evals

Each failure carries its case ID, assertion class, semantic operation, authorization decision, and final state diff. The report is generated on demand and is not committed; delete it before running docs:validate (it carries no OKF frontmatter). See the verification commands reference.