Research A+ Readiness

Purpose

`pnpm research:a-plus:readiness` turns the A+ ambition into a strict evidence contract. It writes JSON and Markdown operator artifacts under `artifacts/research-a-plus/`. The ordinary command reports gaps without failing; `pnpm research:a-plus:readiness:check` exits non-zero until every requirement passes.

Score boundary

There is no interpolated vanity score. The evaluator returns `A+` only when all requirements pass and otherwise returns `ungraded`. Technical completeness cannot compensate for missing independent review, production evaluation, external impact, attributable acquisition, or provider-confirmed positive net income.

The gate requires:

1. One cohesive, machine-readable, desktop/mobile-accessible research journey.

1. Public methodology, schema, contribution, correction, attribution, expert-review, JSON Feed, RSS, and research-software discovery contracts.

1. At least 50 non-synthetic AI evaluation cases with precision and recall at or above 90%, citation and refusal accuracy at or above 95%, p95 latency at or below 10 seconds, and observed cost.

1. One independently reviewed version-2-or-later investigation with a material correction, explicit novelty decision, and published limitations.

1. Three qualified independent researchers and one independent subject expert.

1. At least three matched baseline cases with 20% median time improvement and 80% reproduction.

1. At least one attributable external citation, reuse, or accepted correction.

1. Fresh hashed SEO, funnel, and approved-campaign receipts with a qualified review action.

1. Provider-confirmed live settlements producing positive net income after refunds, fees, and attributable costs; sandbox rows remain excluded.

1. Automated freshness alerts, exact next actions, approval locks, and 30/60/90-day reporting.

1. Fresh tests, lint, build, security, documentation, and release-digest evidence.

Evidence rules

Fresh operational receipts expire after 35 days. Quality evidence expires after seven days; external research validation and impact evidence expire after one year. Fixture, demo, placeholder, sample, sandbox, synthetic, self-reviewed, click-only, replay-only, and non-attributable evidence cannot pass.

The configuration file intentionally contains no external claims and accepts

only version and boundary metadata. Scoring projections supplied through

`--input` are rejected, listed in the report, and never evaluated. External

evidence enters only through dedicated signed artifacts whose integrity

verifiers run in `collectLocalEvidence`. Never copy projection fields into

configuration or relax thresholds to make the report green.

The resulting scorecard is itself a signed, replayable artifact. It retains the

verified evidence projection and rejected configuration fields, reruns all eleven

requirements at the recorded evaluation time, and requires canonical equality for

grade, summary, evidence descriptions, next actions, anti-substitution boundary, and

false authority flags. `research:evidence-operations` verifies this artifact before

reading it and falls back to no passed requirements when it is absent or modified.

A rehashed A+ label or hand-edited requirement cannot alter the operator queue.

Commands

pnpm research:a-plus:readiness
pnpm research:a-plus:readiness -- --input=path/to/boundary-config.json --now=2026-08-14T12:00:00Z
pnpm research:a-plus:readiness:check

No command publishes, contacts researchers, sends email, changes production configuration, enables payments, or establishes novelty.

Investigation, reviewer, and measured-improvement projections in the persisted

scorecard retain their complete verified provenance. Verification replays the

external-validation artifact and same-study baseline/validation join before

recomputing requirements; plausible codes or aggregates without sources are invalid.

The same rule applies to every external lane. The private scorecard retains the

complete open-release, production AI evaluation, external impact, attributable

acquisition, and net-income artifacts. Verification reruns each domain evaluator and

requires canonical equality with the scorecard projection. Missing, unused, invalid,

or rehashed-but-semantically-modified provenance invalidates the whole scorecard.

Technical requirements use the same boundary. Workflow proof retains and replays the

complete desktop/mobile browser report. Automation proof retains exact `package.json`

bytes whose SHA-256 must match the package entry in the verified release manifest.

Quality proof retains the complete release-gate artifact. A scorecard therefore cannot

claim A+ by adding invented workflow, automation, or green-gate booleans.

`pnpm research:evidence-operations` includes the open-release requirement alongside

the seven external lanes. When production proof is absent, its first workstream is

the non-network `research:open-release-handoff` preparation packet. It binds the

exact release and fifteen targets while retaining a blank human decision and false

authority; only the packet's post-approval sequence reaches capture. The queue never treats

local route existence as deployment evidence and never performs a network probe by

itself. After a human-approved exact-release production deployment, the operator may

run the capture command: it reads only fifteen fixed `www.metamuseum.org` Commons

resources, including JSON Feed, RSS, and JSON-LD software metadata; refuses redirects,

non-200 responses, disallowed content types, and bodies

over 2 MB, and writes a new receipt file without overwrite. The separate

`research:open-release-evidence` importer then verifies deployment chronology, exact

release identity, response hashes, and every required route. Neither command deploys

or publishes anything, and capture output alone is not A+ evidence. On each operations

run, a configured probe path is replayed through the same evaluator; the queue changes

its exact command from capture to import only when the bundle is integrity-valid for

the current release. A missing, malformed, stale-at-capture, or cross-release bundle

cannot advance the workstream.

Captured probe rows retain the exact public body in canonical base64. Both capture and

import enforce the 2 MB per-resource bound; import decodes every body, recomputes its

SHA-256, and reparses its declared JSON, JSON Feed, RSS, HTML, or Markdown structure.

This makes the signed artifact independently replayable without trusting copied hash

fields. These are public allowlisted responses, never participant or customer data.

Technical recommendation packet

Impact: the project now has one authoritative A+ claim boundary owned by Product Governance. It changes no research conclusion or external system; it converts eleven previously narrative expectations into deterministic requirements and next actions.

Top risks:

1. External-validation substitution: feature breadth or synthetic panels could be mistaken for scholarly review. Owner: Research Partnerships. Acceptance: three independent researchers, one subject expert, and a materially corrected versioned investigation pass the evaluator.

1. Evaluation-set leakage: internal fixtures could inflate AI precision, recall, citation, or refusal results. Owner: AI Evaluation. Acceptance: at least 50 attributable non-synthetic cases with a retained source and digest meet every threshold.

1. Outcome conflation: traffic, clicks, sandbox payments, or gross receipts could be presented as impact or profit. Owner: Revenue Operations. Acceptance: external reuse is separately evidenced and net income is computed only from live settlement less refunds, fees, and attributable costs.

Actions:

• Product Governance: build and browser-prove the cohesive question-to-review Research Commons journey. Validate with focused page/service tests and desktop/mobile Playwright evidence.
• AI Evaluation: define a frozen production-quality evaluation set, run the dedicated production-evaluation importer, and let the A+ command verify its signed artifact.
• Research Partnerships: use the approval-gated expert pathway to collect pseudonymous, consented, revision-bound reviews; never treat recruitment or replies as validation.
• Revenue Operations: import fresh aggregate acquisition and provider settlement evidence through the existing Organic Income Operator pipeline.

Next-cycle hypothesis: a unified Research Commons surface can make one existing research packet reproducible without exposing direct identifiers or granting publication authority. Falsify with desktop/mobile browser proof, accessibility inspection, packet-schema validation, and a test asserting every consequential transition remains human-gated.

Documentation contract

This runbook is canonical under the repository agent-facing documentation contract. `/docs`, `/api/docs/manifest`, and `/api/docs/content` expose it automatically.