Language quality telemetry — blind spots (2026)
Language quality telemetry — blind spots (2026)
Section titled “Language quality telemetry — blind spots (2026)”Probe charter
Section titled “Probe charter”Map planned language-quality signals (Phase 4 runtime monitors: vox.idiom.*) against privacy guarantees (telemetry-trust-ssot.md).
Evidence (2026-05-11)
Section titled “Evidence (2026-05-11)”- Trust SSOT forbids shipping source by default; aggregates must be carefully bounded.
- Phase 4 describes idiom fingerprint export for corpus weighting — opt-in channel must be explicit in implementation.
Blind spots
Section titled “Blind spots”| Topic | Risk |
|---|---|
| Diagnostic content | Full messages may leak identifiers; need hashing/redaction tiering. |
| File paths | Paths are PII in many workspaces — strip or hash. |
| Cross-session joins | Even aggregates become identifiable when joined with mesh/orchestrator IDs. |
Next steps
Section titled “Next steps”- Specify payload schema for any
vox.idiom.*event incontracts/before enabling emission. - Cross-link evaluation harness docs when Mn-T12 eval gating lands (mesh MENS plan).
- Update mens-training-data-contract.md if corpus weighting consumes telemetry-derived tags.