PaperClipAI/doc/evals.md

Paperclip evaluation guide

The Slack connector probe catalog organizes eleven manual model acceptance probes and a selector for existing deterministic regressions (pnpm test:slack-connector). It is not a registered model campaign; transport fixtures do not prove that an agent chooses a tool or that a real Slack interaction completes.

The explicit-only live provider connection suite is a Product E2E workflow for fresh subscription/API-key/gateway connections, with attended login and independent artifact checks against local or staging targets.

Paperclip has two live eval families with different questions, owners, and evidence. Choose the family before selecting a model, profile, or case.

The explicit-only native instruction consolidation comparison uses six Product E2E cells per source variant. It measures the completion constraint reduction separately from the earlier native tool-description trial. Provider-free start/resume payload capture is a byte measurement; behavioral qualification requires the original paired live outcomes and retained content. Neither source admission nor a scripted pass proves model behavior.

• Runner Evals: real Runner/provider behavior against a seeded mock control plane. Definitions live in paperclip-evals/evals/paperclip-runner; see the direct live protocol evals.
• Product E2E Evals: real browser, Paperclip server, database, Runner, provider, and (where selected) Daytona, using an isolated instance and grading oracle. See tests/runner-e2e and Everyday Workflows.

Runner Evals answer whether a real runner/provider can perform a bounded protocol operation against the expected control-plane contract. Product E2E Evals answer whether a person can complete a product workflow through the real Paperclip surfaces and whether the resulting artifact and state are usable. The names describe the system under test; “headless” is an execution option, not an eval category.

The explicit Product E2E completion-updates suite compares onboarding and idle, busy, multiple-task, and restart Agent Chat handoffs on native Claude/Codex. It separates mechanical completion delivery/result access from semantic review of the retained answer; see the probe contract.

The explicit-only task-titles suite checks that production guidance causes a real native agent to name prompt-only standard/Ask tasks early, while preserving user-supplied titles. Its oracle correlates browser creation, native tool receipts, durable titles, audit ownership, and the reloaded task UI; fixture prompts contain no naming instructions.

The explicit-only native connection guidance suite adds neutral decline prompts, same-task run-attributed explanations, and measured no-use controls across three native local profiles. Its fifteen configured cells are preparation for future matched instruction comparisons, not a live result. Historical Everyday cases and production prompts are preserved.

Selecting a family

Pi's explicit-only pi-controls Product suite tests Stop while native permission is unanswered and browser-originated same-turn steering, locally and on Daytona. Its fixture contract separates control acknowledgment, actual message consumption and owned process retirement. The current Pi matrix has 26 explicit cells (13 local and 13 Daytona), including these four controls, pending-native-question controller restart, and a Daytona-only exact-Pi-child provider-death journey. Provider death must expire the original unanswered card and reject stale replies without a replacement run; any durable fallback is distinct from restoration. File editing also requires native edit/validation events and a registered artifact download. All remain unqualified until measured on the exact candidate runtime and harness.

Use Runner Evals for a runner protocol, adapter, transport, native session, tool grant, or one-turn provider qualification question. The workflow checks out an exact paperclip-evals revision, builds the Runner and viewer, runs a live roster, and renders the canonical Evalbook report. The control plane is a seeded test authority, so a passing result does not prove browser UX, production server behavior, database persistence, Daytona behavior, or a real third-party mutation.

Use Product E2E Evals for browser interaction, issue/task lifecycle, approval and clarification UI, project/repository selection, persistence over a controller restart, artifact delivery, billing/evidence behavior, or runner continuity in local or Daytona environments. The harness creates a fresh Paperclip instance per cell and uses public APIs and the production browser surface. The suite's Everyday Workflows are Product E2E even when their results are imported into Evalbook.

Do not combine a partial Runner campaign and a partial Product E2E campaign into one score. A campaign is comparable when its definition/grader, model/profile, environment, and contract match. The evaluated Paperclip revision may intentionally differ for a before/after fix comparison; record it as a comparison axis.

Ownership and codepaths

Runner Evals are owned by the Runner/evals maintainers. Definitions, rosters, case prompts, and the report program live in the sibling private repository paperclipai/paperclip-evals; Runner integration, viewer, aggregation, and publication code live under packages/paperclip-runner and the runner-protocol-live-evals.yml workflow. The public-facing report uses the same Evalbook renderer and Runner Lab viewer as the trusted report after sanitization.

Product E2E Evals are owned by the runner E2E maintainers. The catalog and harness are under tests/runner-e2e; the package scripts are test:e2e:runner, test:e2e:runner:unit, test:e2e:runner:typecheck, and test:e2e:runner:report. README.md, FIXTURES.md, SECURITY.md, and EVERYDAY-WORKFLOWS.md are the detailed sources of truth. The harness starts the server and embedded database, creates the company/agent/task through the real APIs, drives Chromium, and invokes the selected local or Daytona runner.

The explicit-only agent-chat-hardening Product E2E suite covers native chat recovery, hiring, status evidence, and review handoff on local and selected warm Daytona paths. Its fixture contract distinguishes startup cancellation from active response cancellation and HTTP send replay from ambiguous provider action recovery. Select it explicitly; --all excludes it.

The explicit-only production hiring templates suite adds two local native Codex/Claude cells. It exercises API-created production CEO defaults, an explicitly requested hiring skill/reference read, a permanent coder hire, independently computed saved JSON fixtures and worker reuse. Each cell requires five work turns and admits at most two strictly attributed server task-completion turns. Every actual run remains counted; unknown or extra-work turns fail. Source/read coverage and workflow outcome are separate: missing read provenance leaves the candidate/baseline pair uncomparable even if work succeeds. Baseline bundles and coder examples derive from their own source revision, without requiring candidate wording or length.

The explicit-only context-integrity Product E2E suite covers ordered public comment continuation and explicit invocation of an assigned pinned skill across the seven selected legacy/native local profiles. Select it by suite or exact execution ID because --all excludes explicit-only suites. Each cell applies a 1,000-cent company and agent budget hard stop before task creation and records both limits in its evidence.

The explicit-only stock-harness suite reuses skill, ordered-continuation, and chat-restart journeys across eight local legacy/native profiles with production-default hires. It closes the custom QA manual coverage gap. Its required credential-free prerequisite maps vendor instruction layering, the tiny hire bundle, and shared startup/resume reductions to executable checks. The 24 live cells are configured; no live qualification is claimed from their setup or unit calibration.

The explicit-only agent-chat-stories suite covers the experimental settings lifecycle for a configured native agent and follow-ups during active work. Its fixture-driven file wait and persisted-plan oracle are documented in the Product E2E guide. It does not qualify the native onboarding wizard or change the native API-tool rollout defaults.

The explicit-only grok-qualification and grok-subscription-qualification Product suites exercise Grok Build with API and company subscription authentication respectively. Keep their results separate; the subscription fixture seeds an explicitly supplied login and does not qualify interactive login. See the Grok fixture contract.

The explicit Direct blocker guidance suite checks the legacy coordination skill against human authority, missing hiring permission, and requester scope decisions through saved browser interactions.

Validation ladder

The explicit-only public MCP suite evaluates paid assistant delegation, later retrieval, feedback, review, uncertain retries and permission boundaries. It uses the Product E2E fixtures, launcher, evidence packaging and dashboard, with separate external-assistant and team-worker billing. The 2026-10-01 results retain two complete model matrices, provenance, costs and the earlier failure history.

Start with credential-free checks and a catalog listing. For Product E2E:

pnpm test:e2e:runner:typecheck
pnpm test:e2e:runner:unit
pnpm test:e2e:runner -- --list

For one explicitly selected local cell, configure only the credentials named by that cell in .env.runner-e2e.local, then run a narrow ID:

pnpm test:e2e:runner -- --id core-compatibility.runner-codex.local.message-marker

Use the selectors documented in the runner E2E README for a suite, profile, case, group, or environment. Daytona needs the immutable image digest and DAYTONA_API_KEY; follow the README and fixture security guide. --all excludes manual suites such as everyday-workflows. Select that suite explicitly; use a narrow selector while developing a fixture.

For Runner Evals, the narrowest useful local validation is the report program's help/validation path and the deterministic Runner checks documented in runner-workflow-evals.md. Hosted direct live runs must use the default-branch workflow, an exact 40 character evals_sha, an explicitly selected roster (or the maintained enabled all campaign), and the protected paid environment. The complete hosted command is intentionally kept in the workflow and direct live protocol guide. Live provider runs can spend money; use the existing workflow authorization and the user's stated scope when selecting them.

Failure taxonomy

Record the primary failure class and preserve the evidence that supports it.

• Product failure: evidence shows Paperclip or Runner behavior violates the authored case or a hard invariant, such as wrong task state, missing approval gate, lost persistence, bad artifact, or incorrect protocol operation.
• Model/provider behavior failure: the provider turn completed with usable evidence but the model gave the wrong answer, ignored an interaction, failed to complete the authored operation, or violated a semantic assertion. It is scored as behavior, not silently retried as infrastructure.
• Grading/evidence failure: the case or matcher cannot establish its claim, a required recording/screenshot/result is malformed, or the report contract is invalid. Fix the harness or grader before interpreting the score.
• Infrastructure failure: the evidence points to provider/profile unavailability, transport admission failure, service startup failure, a missing credential/image, or inability to produce usable evidence. Startup, transport, and timeout symptoms can instead be product defects when evidence implicates Paperclip or Runner; classify from the observed failure and supported cause, rather than the symptom name alone. Preserve the artifact.

Missing usage or price data means unknown, not free. Keep provider-reported costs separate from estimates, and include retry costs when available. Latency, cleanup, billing coverage, and unpriced usage are dimensions of the result and should remain visible alongside the primary class. A timeout after successful product state reads can be a product behavior failure; a failed server-health read may be infrastructure, but inspect its cause. Use the family-specific classifier and read the attempt evidence before changing an analytical label.

Evidence, provenance, and history

Retained result snapshots and dated measurement reports belong in paperclip-evals; application tests, Product E2E fixtures/graders, and executable scenario inventories remain in this repository. Keep a compact results index with immutable archive links and public report links, as in the lifecycle baseline. The private archive is not a dependency of app test execution. Keep large logs, traces, and videos in the existing campaign artifact storage.

An Evalbook report is a presentation of immutable attempt records, not the source of truth. Keep the campaign ID, Paperclip commit, paperclip-evals commit, catalog/roster or definition fingerprint, model/profile, environment, grader version, selected cells, retries, and provider/runtime usage with the report. Public projections follow each family's reviewed allowlist and may include sanitized fixture conversation, named tool outcomes, screenshots, and structured evidence intended for public history. Credentials, secrets, private data, raw unredacted records, and hidden reasoning stay out of public projections.

Distinguish a complete campaign from a partial campaign. A narrow selector, manual diagnostic, missing cell, or infrastructure retry can be useful evidence without being a qualification run. History should retain both, with explicit coverage and completeness, while trend and latest-green views compare only compatible complete campaigns. Refreshing an existing report from retained evidence has zero provider calls and is a new presentation of the old measurement, not a new model run.

Existing public histories are available at Runner protocol history and Runner Product E2E history. The consolidated eval hub is at pages.paperclip.ing/evals.

For a repeatable workflow, use the matching skill: paperclip-evals, add-runner-eval, or add-product-e2e-eval.

Diagnose failures before buying another campaign

Use this loop to turn eval failures into product improvements. The unit of work is a broken user outcome or invariant, not an individual red cell.

1. Freeze the evidence. Record the inspected application revision and each campaign's evaluated revision, definition/grader fingerprint, model/profile, environment, exact selected IDs, attempts, and usage coverage. Read the history feed and retained attempt records before launching models. A report refresh, skipped workflow, passing unit suite, or old-definition green cell is not a new live measurement. Inventory explicit-only suites separately from --all.
2. Reconstruct the failed boundary. Read durable task state, interactions, event chronology, source/worker identity, delivered output, and the failing assertion. State whether the test reached the boundary it claims to test. Separate observed failure, machine class, analytical cause, and confidence. A cancelled queued wake does not by itself prove lost work; a failed decline assertion does not prove unauthorized execution. A saved artifact does not prove the user received a correct completion update.
3. Group by cause and product contract. Join cells only when their evidence supports the same mechanism. Check for already-merged fixes and definition corrections before proposing new work. Keep product defects, provider/model behavior, grading defects, infrastructure, and unexercised boundaries distinct. Preserve the original grades when attribution changes.
4. Design the smallest general correction. Name the desired user behavior, the authoritative state/transaction or provider boundary that owns it, the affected callers, and the existing guarantees that must survive. Check it against PRODUCT.md and SPEC-implementation.md. A change to an intentional product rule is a contract change, not an excuse to delete its guard. Avoid case-name branches, phrase-specific prompts, unconditional retries, or weakening approval, ownership, cancellation, and budget gates to get green.
5. Prove the mechanism cheaply. Calibrate a grader against correct and plausible wrong retained evidence. Reproduce a product race with scripted providers or service tests. Pair every proposed fix with a regression that exercises the opposite boundary (for example, eligible delivery versus revoked access, benign obsolete wake versus interrupted active work). Presentation-only changes use the existing report refresh path and make no provider calls; do not claim replay can prove changed runtime behavior.
6. Select a bounded live confirmation. Write the exact failed representative IDs and only the passing controls affected by the change. Pin the source, definitions, models, and remote image. Record an attempt cap, provider and compute budget, wall-clock deadline, and concurrency before dispatch under the existing live-run authorization. Missing pricing means unknown, not free. Do not rerun unaffected green cells during diagnosis or retry usable behavior failures until they happen to pass. Preserve every attempt. Expand to a compatible qualification campaign only after the causal fix passes.
7. Close with evidence and remaining scope. Report original versus new measurements, exact selected coverage, regression results, cost coverage, and unresolved boundaries. A partial verification may close one defect; it cannot turn the full catalog green or establish reliability from one attempt.

Keep one triage record per cause with: affected cell IDs; evidence links; observed failure; supported cause and confidence; existing fix/revision; proposed product contract; invariant regressions; next exact live selection; budget and stop condition; owner; and disposition. Useful dispositions include confirmed product defect, model behavior, grader correction, infrastructure repair, fixed-but-not-remeasured, historical pass, and unqualified coverage.

Parallelize independent artifact inventory, deterministic checks, and isolated cells. Keep a cell's dependent turns ordered. When delegating, use inexpensive agents for bounded extraction, catalog reconciliation, and test execution; keep causal attribution, product design, and final review with the lead. Begin local browser campaigns at the documented conservative concurrency and increase only with measured host headroom. Hosted fanout must respect the workflow's provider and fleet caps; more simultaneous timeouts do not improve wall-clock efficiency.

Current tools support exact-ID selection and retained-evidence report refresh, but not an automatic cause-aware "rerun unresolved failures" planner. Build an explicit selection manifest rather than treating --all as that planner. Review the launcher's automatic retry policy when budgeting; Product E2E may create one fresh attempt for a retryable failure.

Install the authoring skills

The reviewable sources live in this repository's .agents/skills. For a multi-repository workspace, install the three skills at ~/paperclipai/.agents/skills (not ~/paperclipai/skills). From the Paperclip checkout, run:

for skill in paperclip-evals add-runner-eval add-product-e2e-eval; do
  install -d "$HOME/paperclipai/.agents/skills/$skill"
  install -m 644 ".agents/skills/$skill/SKILL.md" \
    "$HOME/paperclipai/.agents/skills/$skill/SKILL.md"
done

This replaces only the three named skill entrypoints. Run it again after updating their tracked sources. Each skill locates the repository independently of its installation directory.

Maintain the public hub

The hub is a static directory with two links to the existing history systems. It displays a dated snapshot, not a live scoreboard. It does not run models, create another result archive, or change the existing campaign URLs.

Build from the public history feeds and check its summary logic:

python3 -m unittest discover -s scripts/evals-hub -p 'test_*.py'
python3 scripts/evals-hub/build.py --output .paperclip/evals-hub

The hub checks need Python 3 and do not call model providers.

For offline checks, pass --history-dir <directory> containing runner-protocol-evals-history.json and runner-e2e-history.json. For a pre-merge preview, pass --docs-ref <branch-or-sha> to link the guide at that revision. The default guide link uses master.

Publish with the Paperclip page helper and the configured page-uploader credentials. Use Bash 4 or newer; macOS's system Bash 3 cannot run this helper. On macOS with Homebrew Bash installed, put $(brew --prefix bash)/bin first in PATH before these commands:

export PAPERCLIP_PAGE_BUCKET=pages.paperclip.ing
export PAPERCLIP_PAGE_BASE_URL=https://pages.paperclip.ing
export AWS_REGION=us-east-1
bash .agents/skills/paperclip-page/scripts/publish.sh .paperclip/evals-hub --slug evals --dry-run
bash .agents/skills/paperclip-page/scripts/publish.sh .paperclip/evals-hub --slug evals

For later refreshes, rebuild in the same output directory and publish with --update. Keep its ignored .paperclip-page/state.json ownership record; without that record, the helper will refuse to overwrite an existing prefix. Verify the public page and its links after publication. This manual refresh does not add a scheduled workflow. Preserve the measurement date when choosing a newer rendering of the same campaign.

Remaining native chat boundaries are in the explicit-only agent-chat-qualification suite: active task reassignment, user Retry after verified worker process loss, and multi-turn answers grounded in actual task records. See the workflow and qualification limits. The 26 native first-task cells exercise onboarding before native selection becomes the UI default. Live results and semantic answer reviews must accompany any qualification claim; catalog presence alone is not a pass.

The explicit-only native question/resume qualification separates a completed two-answer user journey from semantic-tool documentation qualification. It verifies the exact provider-pause or semantic-response-wake binding for each answer and preserves prior grades when the definition changes.

Lifecycle behavior baseline

The credential-free lifecycle baseline joins unit, scripted-runner, and database integration assertions to a scenario inventory before changing narrative-based lifecycle policy. Run pnpm test:lifecycle-baseline to retain current passes and failures. Its Product E2E matcher calibration is separate from live execution; unrun live coverage remains explicitly unmeasured.

The separate live lifecycle baseline defines 46 real-provider Product E2E cells, including paired narrative probes and named existing controls on legacy and native Codex. Discover it with pnpm test:e2e:runner -- --list --suite lifecycle-baseline. Historical execution results and follow-up coverage are recorded in that suite's guide.

Continuation accounting has an explicit-only eight-cell Product E2E baseline suite, complementing the deterministic lifecycle inventory.

The explicit Product E2E instruction-persistence suite verifies private file edits, nested and binary agent files, stopped-provider directory saves, server restart, and a fresh task's downloaded proof on local native/legacy Codex and native Daytona. See the Product E2E runbook.

The explicit-only Product E2E api-response-reading suite verifies retrieval of large saved API responses on local and Daytona native Codex runs. See the Runner E2E guide.

The explicit-only Product E2E extended-harnesses suite covers pending Cursor, Copilot and Pi ACP profiles on local and Daytona. See the fixture admission, credentials and budget contract. The private Runner Evals campaign of the same name provides complementary semantic protocol cases; catalog membership is not live qualification.

The explicit local Copilot protection fixtures exercise native permission denial with operator cancellation and bounded attached command settlement. Their registration remains separate from paid qualification; source/pack provenance and complete persisted evidence are required. The explicit-only Product E2E confirmation-replies suite tests conversational approval and rejection, persisted message provenance, approval before execution, ambiguous proposals, and the existing card-click path with native Claude/Codex. See the suite contract.

Hiring notification accounting now also requires exact completed action attribution. Missing native/provider ID mapping is uncomparable evidence; it must not be reported as a model task regression or waived through name/order matching. The fixture waits for both known completion callbacks and settled bracketed observations, including the gap before pending outbox work becomes a wake. Strict action replay and original machine verdicts are retained separately.

The explicit-only planning guidance utility comparison measures task decomposition and handoffs with current, short, and disabled skills.