Methodology
A support assertion answers one narrow question for a capability revision × exact product target × environment profile. A rendered cell may be yes, partial, no, unknown, or na.
yesmeans current public documentation demonstrates the row’s capability without a material limit.partialmeans the capability exists with a documented plan, platform, transport, rollout, environment, or interaction limit.norequires an explicit current source saying the capability is unsupported. Silence is not evidence of no.unknownis the default whenever evidence has not been reviewed or remains ambiguous.nameans the row does not apply to that kind of harness and still needs public evidence.
Capability families and atomic rows
Broad labels such as MCP, plugins, Agent Skills, subagents, model and context management, file inputs, data security, operational reliability, collaboration, persistent instructions, and WebMCP are capability families, not compatibility cells. A family has independently testable atomic children. For example, MCP tool calling does not imply MCP prompts, resources, sampling, authorization, transport, or Apps support. Likewise, basic subagent delegation does not imply concurrency, nesting, inherited tools, isolated context, independent model choice, or structured handoff.
Family progress is derived from child rows for each exact product. The catalog reports both reviewed children and supported-or-partial children, plus the full status breakdown. It never authors or infers one umbrella family status.
Atomic rows may also declare typed requires, extends, alternative, or related edges. Siblings are derived from their shared family parent.
Evidence contract
Every non-unknown cell is authored as a version entry with one or more note IDs, an explicit target, an environment profile, structured qualifiers, and typed evidence references. Each evidence reference resolves to a stable resource ID and records when it was reviewed. The resource list links to public vendor documentation, release notes, maintained first-party repositories, or the relevant open-standard documentation.
Targets are versioned releases when a stable release is public. Continuously deployed web products use a dated hosted observation. When only current local-product documentation is available, the record says dated-documentation rather than pretending a particular binary was reproduced.
Evidence classes stay separate: documented, vendor-attested, listed, tested, reported, inferred, and not found do not carry the same authority. Catalog v1 publishes documentation and listing evidence. The public test registry contains proposed definitions only; Can My Agent Use has not executed runtime conformance tests against these harnesses.
Plan, policy, region, authorization, transport, protocol revision, role, feature flag, runtime, preview, and vendor-extension limits remain structured qualifiers. They are not folded into prose and lost when the matrix is rendered.
Compatibility status and product lifecycle are separate. An exact cell can be currently unsupported while its lifecycle is planned, or partially supported while it is experimental or preview. Lifecycle values are untracked, requested, planned, experimental, preview, stable, and deprecated; they require the same evidence discipline as the associated compatibility assertion.
Limits and measured values
Some of the most useful harness facts are numeric: advertised and effective context tokens, maximum output, file bytes and count, document pages, audio or video duration, concurrent subagents, requests or tokens per reset window, run duration, storage, retention, and price. A number is not portable without its unit and scope.
Every measured value should identify the exact harness, model or mode, plan, environment, region or policy when relevant, what the value limits, whether it is a documented maximum or a tested observation, the reset or billing period, the review date, and the boundary behavior. A model API limit is not copied into a chat, desktop, editor, or CLI surface without product evidence. When an advertised maximum and an effective usable limit differ, preserve both and explain the reservation, truncation, sampling, or compaction behavior.
Prompt caching is split into three claims: reuse, operator controls, and telemetry. File upload is split from model-visible understanding and from preprocessing disclosure. These distinctions keep a paperclip icon, a provider model card, or a generic caching statement from becoming a broader compatibility claim than the evidence supports.
Do not source a claim from rumors, search snippets without a stable page, community recollection, private beta screens, or screenshots of interfaces that are not public. When a document describes a model or API rather than the named harness, it does not automatically prove harness support.
Community reports can surface an investigation or contradiction, but do not change a definitive support state without authoritative or reproducible corroboration. When sources conflict, the contradiction stays public instead of being smoothed into a color.
Tracks and coverage
Tracks (current, preview) are catalog version stacks. They are not a claim that a vendor publishes channels with those names, and evidence for current never fills preview automatically.
The feature-page percentage is the share of published product tracks that have moved off unknown. It measures catalog coverage. It does not measure usage, adoption, quality, or market share. The first review covers MCP tools, MCP Apps, workspace files, terminal access, computer use, and skills for a small set of products. All other cells stay unknown until reviewed.
Corrections
Public documentation changes. A correction should supersede the affected statement, update the review date, preserve a clear explanation of the documented limit, and replace stale resource links without erasing the reason for the change. Contributors should follow CONTRIBUTING.md; the shared schema rejects non-unknown cells without targets, environment profiles, notes, and typed evidence.
The product deliberately does not publish a synthetic harness score or fabricated market-share figure. Useful operational metrics are evidence freshness, sourced-cell coverage, contradiction age, revision pinning, source diversity, review latency, and future conformance-test coverage.