SELF/XProof-driven caseWritten brief
Case 08 · Modality routing

Evidence-Led Language Tutor

Turn mixed speech and writing evidence into one focused learning target without double counting, vanity scores, or premature mastery claims.

Open the practice foundry
Evidence stays separate
SpeechWriting
One active anchorspeech-thought-group
Produced output · focused work taskpause-placement
Validation boundaryHeld-out validation pending · outcome unverified
Primary decisionRoute one evidence pin
Primary resultGuided practice target

LimitationNo real learner, audio, transcript, writing sample, Telegram message, provider, runtime state, personal metric, proficiency, mastery, retention, or learning improvement is shown.

Split-Signal Practice Foundry · synthetic

One work task. Two evidence modes. One active learning target.

2 fictional artifacts · 2 tracks · 1 anchor · 0 double counts · 0 external calls

Independent evidence

Synthetic spoken attempt · no recorded voiceActive anchor
The main constraint / is partner approval, so I recommend a one-week delay.

full coverage · synthetic-spoken-attempt

Original writing fixture · not personal writingObserved · waiting
This is still unresolved, so it should move by one week.

full coverage · explicit-fictional-writing

2 artifacts · 2 tracks · 1 active anchor · 0 double counts

Active practice plane

speech

delivery-cueKeep the complete subject and predicate inside one thought group.

Allowed
pause placement · thought-group boundary
Excluded
writing score · mastery · pronunciation quality

Choose one minimal correction · speech · anchor speech-thought-group · outcome unverified

Transfer field

  1. 1dnear-transferdue
  2. 3dnew-audiencedue
  3. 7dheld-out-transferunverified
  4. 30dretention-recheckunverified

Immediate action can be guided. Transfer and retention remain unverified until eligible held-out evidence exists.

lt-ef294f72
Inspect evidence boundary
MaturityVerified private prototype · current source local-only
EvidenceClean current source · fictional public reconstruction
The evidence problem

A transcript is not writing, and immediate correction is not learning.

The foundry preserves provenance, chooses one high-value target, changes the exercise by modality, and keeps delayed review honest until held-out evidence exists.

Recorded source proof

Verified modality contracts, fictional public evidence

Private tutor substrate

Verified private prototype · current source local-only

Fresh local verification · 2026-08-29Neutral label · no source media

The pinned source separates speech, explicit writing, reference and unknown evidence; keeps acoustic diagnostics speech-only; carries coverage; schedules held-out review; and implements bounded private-data retention.

Current source HEAD is newer than the documented installed revision. No service continuity, learning result, current private data, provider behavior or production effectiveness is claimed.
Written-first handoff

Need learning automation that respects evidence instead of inflating a score?

Start with modality provenance, one active target, minimal correction, held-out transfer, review timing, privacy and a truthful insufficient-data state.

Open the written brief