# Farka changelog

Completed product changes and production evidence. Dates refer to finished work,
not roadmap intentions.

## 2026-09-01, continual fine-tuning

Dataset lineages pin a frozen content-hash holdout, so every version of a corpus
is scored on the same terms and adding rows never moves an existing row across
the train/holdout boundary. A continuation run warm-starts from a delivered
adapter and is measured against that predecessor rather than the base model.
Proven end to end on Azure: iteration two of a Qwen2.5 1.5B specialist trained
from its predecessor's adapter, passed the quality gate, settled, and released
its machine.

## 2026-08-13, verified model evidence

Nine catalog models were promoted from catalog-only to live-verified using
production PostgreSQL evidence. Public model claims now separate catalog
availability from exact workflow proof.

## 2026-08-12, blind promotion and full tuning

SmolLM2 360M full tuning improved blind accuracy from 21/60 to 58/60. Qwen2.5
Coder 7B QLoRA improved from 15/30 to 30/30.

## 2026-08-05, export contract verification

Independent end-to-end checks passed target-specific weights, deployment-bundle,
and SDK artifact contracts, plus the missing-artifact failure paths.

## 2026-08-02, versioned legal acceptance

Terms, Privacy, Acceptable Use, and model-specific license evidence became
versioned, hashed, and enforced before paid quotes.

## 2026-07-23, production security hardening

Tenant scoping, authentication, billing, delivery locks, rate limits, upload
handling, and production readiness received a full-stack security pass.

## Contact

edi.hasaj@farka.ai