j j"UdZddlmZddlZddlZddlmZmZddlm Z ddl m Z eje ZedGd d ZedGd d Zd dddddZdaded<daded<ddZ ddZddZy)uECG-FM numerical verifier (Phase 4). The interpretation LLM produces measurements (heart rate, intervals, axis, ST deviation) along with its qualitative read. Multimodal LLMs are demonstrably weak at reading numbers off rendered EKG strips — they confabulate plausible-looking values. This module runs a numerical model on the underlying WFDB signal to produce ground-truth measurements, then the caller reconciles: if the LLM and verifier disagree by more than a configurable threshold, the verifier wins and the corrected number is flagged on the case. This module is intentionally lazy — the model loads on first call so the service can boot even on a host that hasn't fetched the weights yet. If torch or the model files are missing, every call returns measurements_skipped with a clear reason rather than failing hard. Setup (one-time, on the droplet): cd apps/ekg-tutor/interpret .venv/bin/pip install torch torchvision wfdb .venv/bin/python -c "from huggingface_hub import snapshot_download; snapshot_download('bowang-lab/ECG-FM', local_dir='./models/ecg-fm')" Then set EKG_TUTOR_ECG_FM_PATH=./models/ecg-fm in /root/secrets.env. ) annotationsN) dataclassasdict)Path)AnyT)frozenc^eZdZUded<ded<ded<ded<ded<ded<d ed <d ed <y ) MeasurementVerificationstrfieldfloat llm_valueverifier_valueunitdelta thresholdbool agreementcorrection_appliedN)__name__ __module__ __qualname____annotations__O/var/www/adampowell.pro/apps/ekg-tutor/interpret/src/pipelines/ecg_fm_verify.pyr r &s. J I LOrr c>eZdZUded<ded<ded<ded<d dZy ) VerifierResultr available str | Noneskipped_reason model_versionzlist[MeasurementVerification] verificationsc |j|j|j|jDcgc] }t |c}dScc}w)Nrr!r"r#)rr!r"r#r)selfvs rto_dictzVerifierResult.to_dict9sI"11!//151C1CD1CAfQi1CD   EsA N)returnzdict[str, Any])rrrrr(rrrrr2sO00 rrg$@)ratez intervals.prz intervals.qrsz intervals.qtcaxisr_modelr _load_errorctt ttfStjj ddj }|s dadtfSt |js d|adtfS ddl} tjdd adtfS#t$r dadtfcYSwxYw#t$r}d |adtfcYd}~Sd}~wwxYw) zReturns (model, error_reason). Caches the result so the load only runs once per process. On missing torch / weights / etc., returns (None, reason).NEKG_TUTOR_ECG_FM_PATHzEKG_TUTOR_ECG_FM_PATH not setz$ECG-FM weights directory not found: rz6torch is not installed; run: uv pip install torch wfdbuECG-FM model loader not implemented — Phase 4 scaffolding only. Wire up bowang-lab/ECG-FM in pipelines/ecg_fm_verify.py to enable verification.z8ECG-FM loader not implemented (Phase 4 scaffolding only)zECG-FM load failed: ) r.r/osenvirongetstriprexiststorch ImportErrorloggerwarning Exception) ecg_fm_pathr8es r_load_model_lazyr?Ps[4{""**..!8"=CCEK 5 [    # # %<[MJ [  ! !  ^ Q [  ! !N [  !" !,QC0 [  !s0:B?B7B43B47 C C CCc\t\}}|td|xsddgStd)zRuns ECG-FM on the WFDB signal and reconciles with LLM measurements. `llm_measurements` is a flat dict of the LLM's numeric reads, keyed by the same field strings used in THRESHOLDS. Missing fields are ignored. NF unavailabler%uuECG-FM forward pass not implemented; Phase 4 scaffolding only. Loader returned non-None unexpectedly — investigate.)r?rNotImplementedError) signal_pathllm_measurementsmodelerrs rverify_measurementsrGwsH"#JE3 }/-   "  A rcdtjjddjdk7S)zACheap check for /health endpoints. Doesn't trigger the lazy load.r1r2)r3r4r5r6rrr is_availablerIs& ::>>12 6 < < >" DDr)r)ztuple[Any, str | None])rCz str | PathrDzdict[str, float]r)r)r)r)__doc__ __future__rloggingr3 dataclassesrrpathlibrtypingr getLoggerrr:r r THRESHOLDSr.rr/r?rGrIrrrrRs4# )   8 $ $ $     "     Z$!N&DEr