Ankatta Capture V0

Ankatta Capture V0

Ankatta fits the N'Ko/Malinke speech stack as a lexicon-assisted intent capture surface. It is not the recognizer and it is not the translator. Its job is to make human evidence collection easy when the speaker does not know exact N'Ko spelling.

The page records short audio, lets the speaker enter what they meant in English or rough Malinke/Bambara/Jula, and uses the scraped Ankataa dictionary as a quiet vocabulary aid. Dictionary matches appear automatically from the meaning or rough phrase field. Checked matches are saved only as candidate lexicon context. They do not prove that the audio contained those words.

The default UI is now a family training surface with three modes:

Adult validation
  -> record natural speech
  -> type or select the intended meaning
  -> save governed evidence

Kid multiple choice
  -> see a known prompt
  -> pick the phrase
  -> say it aloud
  -> save the take

Baby tap cards
  -> tap a large word card
  -> repeat with an adult
  -> save the take

The original simple capture flow is still preserved in code as
`html_app_simple_capture()`, but the served app is the family trainer.

The packet flow remains intentionally small:

Record
  -> Meaning
  -> optional Ankataa candidates
  -> Save Evidence

Advanced fields such as phrase identity, optional N'Ko, notes, local export, and raw status remain available behind the advanced drawer. They are useful for review, but they are not the main user journey.

Prompted kid and baby captures do not require typed meaning. The prompt itself
becomes the expected meaning and rough Latin/Malinke phrase. This is stronger
than open-ended capture because the intended label is known before recording,
but it is still not training truth until governed review admits it.

This keeps the architecture aligned with AGP and the acoustic world model:

audio evidence
  -> speaker intent
  -> optional lexicon candidates
  -> governed review
  -> replayable acoustic memory
  -> later ASR / N'Ko rendering / translation

The important boundary is that Ankatta reduces human friction without weakening truth. It gives the system richer evidence, but every accepted transcript or translation still has to pass governed review.

Run

cd [home]/Desktop/nko-brain-scanner/experiments/acoustic_gate
python3 serve_ankatta_capture_v0.py --host [ip] --port 8791 --open

Current local service:

http://[ip]:8791/

Current tailnet HTTPS service for iPhone:

https://mohameds-macbook-air.tail226fc2.ts.net:8791/

Data

The app loads:

[home]/Desktop/nko-brain-scanner/pipeline/data/dictionary/ankataa_dictionary_20260319.json

Current dictionary count: 2,101 rows.

Captured packets are written to:

/tmp/nko_speech_calibration_review_v0_current/ankatta_capture_packets/

Manifest:

/tmp/nko_speech_calibration_review_v0_current/ankatta_capture_packets/ankatta_capture_manifest.jsonl

Proof Boundary

Ankatta packets are saved with:

labelStrength=human_intent_weak_label
trainingAdmissible=false
translationClaim=none

That means they are useful evidence, but not training truth yet.