Versioned, bounded and conservative
Inspection methodology
The launch policy is deterministic: the same admitted scalar sequence, input profile and versioned policy produce the same findings and proposed edits. Results name their engine, Unicode data, provenance probe and policy versions so a receipt can be interpreted later.
Pinned reference data
Classification uses Core Text’s Unicode 15.1.0 and Emoji 15.1 data artifact. Provenance preflight uses the versioned Core C2PA text probe against the C2PA 2.4 structural forms covered by the property profile. The browser asset identity and server declaration are hash-bound during startup.
Unicode names are presentation derived from pinned data, not machine outcome codes and never labels supplied by input. Valid emoji joiner/tag context is checked before individual invisible sequence components are classified.
Closed evidence and result shapes
The processor emits one finding per occurrence, with a stable ID, code, severity, one bounded evidence span, rationale, action and closed detail fields. Counts distinguish total and returned items. This property does not truncate a successful result: if all editable occurrences cannot be returned and receipted within limits, the run refuses.
A completed transformation is a child of one complete inspection. Selected edit spans are recomputed, checked for overlap, applied in source order and recorded with exact before scalars and empty after scalars. Output differences must map one-to-one to receipt operations.
Coverage and limits
Coverage explicitly lists Unicode controls, three C2PA structural layers, statistical watermarks, AI authorship and file metadata. Unsupported rows stay visible rather than making a narrow completed inspection sound universal. Unknown future format controls produce partial coverage and disable transformation.
Admission caps bytes, UTF-16 code units and Unicode scalars. Separate ceilings cover findings, changes, evidence, result metadata, local export bytes and deadline. The C2PA probe has its own raw-byte, candidate, nesting and elapsed-time budgets. Exceeding a ceiling produces a refusal, never partial output presented as complete.
Change record
The policy changelog begins with version 1: browser-local UTF-8 text and Markdown, conservative per-occurrence deletions, no normalization, protected provenance and no workbench telemetry. A future change that broadens supported mechanisms or edit permissions requires a new policy version, reviewed corpus and revised public claim.
Primary references
These sources describe the external standards or mechanisms discussed above. The property’s cleanup decisions remain narrower than those specifications.