# FinanceGPT ML Service — Public Technical Report

## Purpose
FinanceGPT separates deterministic financial calculation from learned-model inference. The ML Service is a private FastAPI runtime used for approved model training, inference and evaluation. Laravel remains the control plane for consent, datasets, model/version governance, activation, inference audit, drift review and evidence.

## Operating boundary
- The ML runtime does not receive FinanceGPT database credentials.
- Approved workflows send bounded numeric feature payloads rather than raw documents, provider credentials or transaction descriptions.
- Learned-model output cannot place trades, send payments, approve financial actions or bypass FinanceGPT execution policies.
- Installation of CUDA runtime packages does not prove that physical GPU hardware is present.
- Runtime readiness is published only after numerical imports, service startup and deterministic inference smoke tests pass.

## Model families
FinanceGPT currently governs model families for financial anomaly detection, conditional financial scenarios and time-series financial path generation. Candidate models require evaluation evidence and the relevant activation approval before they can become active.

## Runtime verification
The public runtime manifest records only sanitized posture fields:
- verification status and timestamp;
- Python, NumPy and PyTorch versions;
- numerical import status;
- service import/health status;
- model inference smoke-test status;
- verified execution device (`cpu`, `cuda` or `unknown`);
- approved model families available after a successful inference verification;
- non-sensitive reason codes when readiness is incomplete.

Internal host names, filesystem paths, credentials, package-install commands and private network details are excluded.

## CPU compatibility
The deployment environment previously reported a NumPy CPU-instruction compatibility failure. FinanceGPT therefore treats package installation and runtime verification as separate states. The supplied repair tooling replaces an incompatible NumPy wheel with an explicitly configured compatibility pin, then reruns the full numerical/service/inference verification before the public status can change to operational.

## Public endpoints
- `/developers/ml-service` — human-readable ML runtime and architecture page.
- `/developers/ml-service/status.json` — sanitized runtime status manifest.
- `/developers/ml-service/report.md` — this public technical report.
- `/developers/openapi.yaml` — machine-readable FinanceGPT API v2 contract.

## Private service endpoints
The FastAPI ML runtime is an internal service and is not exposed as a public customer API. Its internal API contract is authenticated independently and is consumed by governed FinanceGPT services only.