Three different questions.
| Question | What is recorded |
|---|---|
| Implementation | Exploring, planned, in development, experimental, internally implemented or included in a reviewed release |
| Evidence/readiness | Source review, observed development proof or exact-package qualification, with passed/partial/blocked/failed/inconclusive outcome |
| Usability | Observed ability to find, complete, inspect, revise and recover in a defined task |
A test pass percentage cannot become a usability rating or a product completion percentage.
Separate profiles for three workflows.
| Human | MCP agent | Fusion |
|---|---|---|
| Discoverability | Discovery | Appearance |
| Creation and editing | Task completion | Animation |
| Feedback | Inspection | Editability |
| Recovery | Revision | Transfer reliability |
| — | Recovery | — |
For a feature that does not author an element, Fusion can be not applicable with a reason. Missing evidence for an expected route stays Not assessed.
A proposed scale, not assigned scores.
| Anchor | Observable meaning |
|---|---|
| Not assessed | No adequate observed walkthrough; no number |
| 0 — Blocked | An attempted task cannot complete correctly or critical recovery/result fails |
| 1 — Fragile | Substantial expert intervention/workaround; unreliable result or recovery |
| 2 — Constrained | Simple completion with repeated assistance or material interruption |
| 3 — Usable with caveats | Normal case completes with documented setup and manageable limits |
| 4 — Dependable in tested cases | Representative revision/recovery work without substantial assistance |
| 5 — Refined in tested scope | Clear first/repeat use and demonstrated recovery, independently reviewed |
Every judgement needs an environment.
- App stage/version and exact candidate/package.
- OS, architecture, relevant hardware/GPU and resources.
- Resolve edition/version and transport, or AI client/version as applicable.
- A defined task, starting state, intended result and limits.
- Reviewer/method/date and inspected evidence, with appropriate public clearance.
Expose the bottleneck, never a product score.
If a numeric track summary is later adopted, the proposal is to use the lowest applicable dimension and explain what limits it. Any missing expected dimension makes the track Not fully assessed. Do not average Human, MCP and Fusion into one score.
Assessment counts describe the set reviewed, not product completion. The current explorer shows unassessed profiles without fabricating evidence.
Changes can invalidate old proof.
Recheck affected scenarios when the feature, renderer, generator, bridge, package, resources, OS/GPU, Resolve route or AI client changes. Credible regressions require visible revalidation status. Keep old assessments dated rather than silently copying their score to a new candidate.
Source/compiler structure can support investigation. Human UI behaviour and actual destination appearance/editability require observation. Public evidence summaries must be cleared of private material.