Skip to content

Evidence Policy

Decision support requires evidence. Qualitative claims are allowed, but the reader should be able to see how much confidence to place in them.

Evidence Levels

Level Meaning Use for Required context
Measured Backed by a benchmark with workload, environment, and date Latency, throughput, utility, cost, leakage tests Workload, data shape, hardware/software, date, limits
Deployment-backed Backed by a named deployment, case study, incident, or postmortem Operational lessons, integration failures, governance issues Organization/project, maturity, source, caveats
Literature-backed Backed by papers, standards, specifications, or technical reports Security properties, known attacks, formal guarantees Source, scope, assumptions, date where relevant
Expert judgment Editorial judgment from maintainers or contributors Early guidance where public evidence is thin Why the judgment is plausible and what would change it
Needs evidence Useful claim that should not yet be treated as decision-grade Open tradeoff assertions and suspected pain points What evidence would make the claim usable

Source Quality Labels

Evidence level says what kind of support a claim has. Source quality says how much independence and context the source provides.

Label Meaning Good use Caveat
Primary / official Documentation, report, paper, or data from the deploying organization, standards body, regulator, or project team Establish what was shipped, what parameters were claimed, and what the operator says the system does May omit failures, user impact, or independent validation
Peer-reviewed / academic Journal, conference, or technical paper with methods and results Support measured results, attack claims, and limitations with methods attached May still be a research prototype rather than a production deployment
Independent analysis Evaluation by a party not operating or selling the system Check utility impact, risk, criticism, and external validity Methods, incentives, and data access still need review
Vendor case study Marketing, customer story, or product documentation from a vendor or platform operator Identify candidate deployments and product controls Do not treat as independent proof of maturity, utility, or privacy
Press / secondary News, blog coverage, analyst write-up, or conference report Background and leads to primary sources Use only when primary evidence is unavailable, and label the uncertainty
Unsourced / illustrative Maintainer-created example or value with no supporting measurement Teaching examples and benchmark templates Never use as decision evidence

Use both labels together when needed:

(Evidence: Deployment-backed. Source quality: Vendor case study. Reviewed 2026-06-17 — product documentation confirms the control exists, but no independent production evaluation was found.)

Deployment Maturity Labels

Deployment maturity is separate from source quality. A production system can still be supported only by vendor documentation, and a peer-reviewed paper can still describe only a prototype.

Label Meaning
Production Used in a live public, commercial, clinical, civic, or institutional workflow
Production, batch / periodic Used in a real workflow, but run as a periodic release or recurring study rather than an always-on service
Pilot Tested with real organizations or real data, but not clearly sustained as a production workflow
Demonstration experiment Tested with real partners or realistic data to prove feasibility; operational adoption is not established
Research prototype Published experiment, demo, or study; may use real data but is primarily research evidence
Vendor case study Deployment described mainly by a vendor, platform operator, or customer story; maturity may be real but independently thin
Proposed use case Plausible application with no named deployment evidence on the page
Unclear Public sources do not provide enough detail to classify maturity

Claim States

Use these labels when reviewing pages:

State Meaning Action
Decision-grade Claim is scoped and supported enough for practical use Keep dated and linked
Usable with caveat Claim is useful but has workload or maturity limits State caveat next to claim
Expert judgment Claim is maintainer judgment, not sourced evidence Mark explicitly and add to claim register if important
Needs evidence Claim is plausible but not yet reliable Soften wording and add needed evidence
Too broad Claim is true only in some settings Narrow the claim before sourcing
Remove Claim is generic, unsupported, or not decision-relevant Delete it

Resolved and Unresolved Claim Status

Each entry in the Claim Register carries a Resolved or Unresolved status. A claim is Resolved when sufficient evidence has been identified or the claim has been narrowed and labelled so readers can calibrate their confidence (inline notes are in place on the relevant pages). A claim is Unresolved when it remains too broad, too vague, or unsourced for decision use. The register records a dated note explaining what evidence was found or what is still missing. Status is re-evaluated on each editorial pass; inline evidence notes on content pages use the form *(Evidence: <level>, <date> — <one-line note>.)*.

Claims That Need Evidence

Add evidence or mark uncertainty for claims about:

  • cost;
  • latency;
  • scalability;
  • maturity;
  • production readiness;
  • security strength;
  • privacy strength;
  • common failure rates;
  • tooling quality;
  • regulatory acceptance.

Dated Claims

Fast-moving claims should include:

  • Last reviewed;
  • evidence date or source date;
  • workload or environment if measured;
  • scope of the claim;
  • what may have changed.

Use this metadata block for fast-moving decision pages:

!!! info "Review status"
    Last reviewed: 2026-06-05
    Evidence level: Expert judgment
    Snapshot scope: Public guidance and examples available at review time.

Wording Rules

Prefer:

As of 2026, public benchmarks for this workload suggest...

Avoid:

This PET is slow.

Prefer:

This is an expert-judgment claim until benchmarked for the target workload.

Avoid:

This is production-ready.

Prefer:

This protects inputs from the model service under the stated key-management assumptions.

Avoid:

This protects user privacy.

Citation Standard

A citation should support the specific claim near it. A page-level reading list is not enough for decision support.

For benchmark claims, include workload, dataset or data shape, hardware, software versions, date, and what the benchmark does not cover.

For deployment claims, include maturity: production, pilot, research prototype, or unclear.

Review Workflow

  1. Find claims that affect a PET choice.
  2. Narrow claims until the protected asset, adversary, workload, and output are clear.
  3. Assign an evidence level and claim state.
  4. Add a source, benchmark, deployment entry, or caveat.
  5. Add unresolved high-impact claims to the Claim Register.