shin ש-vs-ס: low-band decisive-gap guard (block noise that squeaked past floor)

shin ש-vs-ס: lower presence floor 0.02->0.003 (fix isolated-sound false negatives)

Hybrid routing: STT for sound/syllable, math for words

Add STT 'pocket' for MVP ש/ס & ק/ת detection (toggle, math preserved)

ש scoring: replace brittle s_band heuristic with model-driven ש-vs-ס

k_sound: raise burst-centroid threshold 2500->3400 for child voices

Fix sh_sound crash (dict.score) + lenient s_band for isolated stage-1 ש

evaluate: never leak Python exceptions to the client; log traceback instead

Add WavLM wetness/lateral-lisp detector — ש-syllable pipeline only

SH syllable: log temporal/envelope features for wet-detection probe

SH syllable: add CTC-based CH-substitution gate

SH syllable: extend extraction window from 120ms to 200ms

SH syllable: log full feature set for wetness calibration

K syllable: top-choice CTC with clinical-pedagogy accept rules

K syllables: switch to multi-way CTC competitive alignment

K syllable: remove broken decay gate (was rejecting all correct K)

K-syllable pipeline: 4-gate acoustic validation (per-vowel + decay + voicing)

Generalize ק incorrect feedback — no longer assumes ת substitution

Add silence / no-signal gate to all scoring pipelines

Drop sub3k gate from ש pipeline; raise s_band threshold to 0.45

Add CV-syllable scoring pipelines (sh_syllable_*, k_syllable_*)

Unify shin scoring: apply 4-gate system to all ש words

Add s_band + sub3k gates for sh_sound (15/15 calibration accuracy)

Log full feature set for sh_sound calibration

Widen centroid ceiling for sh_sound + skip noise reduction

Remove s_band gate for isolated SH — not device-invariant

Add s_band gate for isolated sh_sound — catches wet CH and S substitution

Bypass AI gate for isolated sh_sound — centroid only

Fix plosive detection: onset detection + plosive-band centroid

Fix isolated plosive detection: skip denoise, fix trim, fix burst detection

Replace AI-based isolated K/T detection with burst spectral centroid

Improve isolated K/T: top-5 frame average + minimum 0.5s audio guarantee

Add non-stationary spectral gating noise reduction to preprocessing pipeline

Add Competitive Alignment for K (ק) vs T (ת) substitution detection

🎯 Lower centroid floor from 3000 to 2700 Hz

🔧 Remove s_band gate — it measured S-like energy, not fricative presence

🔧 Add leading silence trimming before Wav2Vec2 alignment

🔧 Remove Slush Filter gate — unreliable on mobile microphones

opus fixing

🚀 Replace static slush filter with Dynamic Peak detection to avoid vowel transitions

🎯 Deploy universal data-driven calibration: 40ms targeted slush filter, AI>=0.70, Energy>=0.025

🐛 Fix JSON serialization error by casting numpy floats to native python floats

🧠 Strip AI of grading authority to prevent LM autocorrect bias. Rely 100% on physical acoustics.

⚖️ Enforce ruthless clinical AI threshold (0.80) to ban all substitutions and lisps

🔧 Revert quantization to restore 100% precision and adjust energy threshold to 3%

🧹 Simplify to binary CORRECT/INCORRECT using science-based acoustic boundaries

🎯 Fine-tune S-Sub threshold to 10% for mobile rejection

🧠 Use Wav2Vec2 rejection + acoustic energy to perfectly detect S-Substitution

🛡️ Make S-Substitution robust against mic variations using pure Centroid frequency

🚀 Fix extraction bug: Extract FORWARD from Wav2Vec2 onset