This is the semi-automated manual tier of a WCAG assessment: the checks that need interaction (keyboard, focus, state changes, reflow) or human review (focus visibility, error recovery, reading order), which a static rule engine can't decide. Work against the running page; use source only to map a finding to `file:line`. Locate and assess — don't fix (that's `accesslint:accessibility-fix`). The automated tier is `accesslint:accessibility-scan`; `accesslint:accessibility-audit` runs both under WCAG-EM.

The shared rules — severity, the no-proxy boundary, high-risk patterns, conformance, grounding — are in [`../shared/methodology.md`](../shared/methodology.md). Read it when a call needs judgment. The rules that always apply are below.

## Grading

Tag each finding with a severity (user impact) and an evidence basis (what you can support). Keep the two separate.

- ● Verified — deterministic and reproducible. Cite the selector, the interaction, and the observed a11y-tree or DOM fact. Without that proof it is not ●.
- ◐ Flagged — you have evidence, but the decision needs a person. Attach the evidence and your opinion; don't decide it yourself.
- ○ Human-required — needs assistive technology or lived experience. Hand it off; don't emulate it.

When unsure between two evidence grades, use the lower one. The icons show evidence basis by fill, not color.

A finding that stacks a deterministic fact on an interpretive call takes the lower grade. The machine-checkable half (e.g. no programmatic status in the a11y tree — a 1.3.1 fact) is ● evidence *inside* a ◐ finding; the interpretive conclusion (e.g. that color is the sole carrier of the meaning — 1.4.1) stays ◐. Citing the ● half does not upgrade the whole.

Severity (user impact, separate from evidence basis):
- Critical — blocks a core task, with no workaround.
- Serious — a major barrier; the task is possible but difficult.
- Moderate — noticeable friction; the task still completes.
- Minor — a small inefficiency or polish issue.

## The ledger — a denominator, not a script

The checkpoint areas below are the run's denominator: every criterion in them ends the run in exactly one state, and the report says which.

- **● verified** or **◐/○ flagged** — you drove the check and have the evidence.
- **N/A** — the triggering feature isn't on the page (no form, no media, no drag UI). Decided from the snapshot; free.
- **Not exercised** — the feature is present but you didn't drive it. Reported as **undetermined**, never silently dropped, and never as a pass.

Not-exercised is a legitimate, honest outcome and it costs nothing. Drive only what the page's features and the engine's gaps demand; don't work through checkpoints to make the report look thorough. An "undetermined — not exercised" line and a driven-and-verified line differ in evidence, not in honesty.

Dedup **before** driving, not after: the engine (`accessibility-scan`) owns the statically-detectable criteria — name/label presence, `lang`, `autocomplete`, semantic markup, contrast of text on flat backgrounds. If a scan result set or engine-owned SC list was provided (as `accessibility-audit` does), never re-check those criteria; ledger them as engine-owned. Without one, read what you can from the initial snapshot rather than re-deriving what the engine would catch.

| Area | SCs | Trigger to drive |
|------|-----|------------------|
| Keyboard & focus | 2.1.1, 2.1.2, 2.4.3, 2.4.7, 2.4.11 | Always — one batched walk (below) |
| Structure & semantics | 1.3.1, 1.3.2, 2.4.1, 2.4.6 | Always — read from the initial snapshot; no extra driving |
| Names, roles, states | 4.1.2, 2.5.3, 4.1.3 | Custom widgets / stateful controls present (static name presence is engine-owned) |
| Visual adaptation | 1.4.1, 1.4.4, 1.4.10, 1.4.12, 1.3.4, 2.3.3, 1.4.11 | Reflow/zoom on every page (one resize each); motion checks only if motion is present |
| Forms & errors | 3.3.1–3.3.3, 1.3.5, 3.3.7, 3.3.8 | A form is present (label/`autocomplete` presence is engine-owned) |
| Media & timing | 1.2.x, 2.2.1, 2.2.2 | `<video>`/`<audio>`, autoplay, or timeouts present |
| Pointer & target | 2.5.1, 2.5.7, 2.5.8 | Target size comes free from the batched walk; drag checks only if drag UI |
| Content & navigation | 3.1.1/2, 2.4.4, 3.2.3/4, 3.2.6 | `lang` and link purpose from the snapshot; cross-page consistency is `accessibility-audit`'s (not exercised here) |

Per-checkpoint procedure and default grades: [`references/checkpoints.md`](references/checkpoints.md). Read it when a triggered area needs its detailed steps, or when the user asks for a deep pass (`--deep`: drive every triggered area through its full procedure).

## Evidence budget — cap spend by grade

Calibrated uncertainty must be cheaper than false certainty. The grade a finding can reach bounds the evidence worth gathering:

- **●** — full proof, gathered once: selector, interaction, observed DOM or a11y-tree fact. Don't repeat an interaction you already recorded.
- **◐** — hard cap:

…