model-quality-release-gate

AI Model Release Control Center

This repository is the methodology and artifact-index surface for a research-engineering project focused on the post-training → evaluation → production interface.

It is not a foundation model. It documents how training interventions, baseline/candidate experiments, deterministic evaluation, CI release policy, live-provider evidence, production traces and model promotion fit into one lifecycle.

Train → Evaluate → Compare → Investigate → Gate → Ship → Monitor → Learn

Post-Training Experiment Lab

The Lab represents each intervention as an explicit experiment with:

• baseline and candidate model identities
• training intervention (for example SFT, preference optimization or RFT/RLVR)
• dataset and evaluation-set provenance
• multi-objective metrics across code quality, correctness, instruction following, hallucination, safety, latency, cost and token efficiency
• release decision: SHIP, INVESTIGATE or HOLD
• ranked causal hypotheses
• follow-up experiments intended to falsify those hypotheses
• machine-readable experiment artifact

The default CodeModel-v1 → CodeModel-v2-sft results are illustrative. They demonstrate the investigation workflow without making unsupported benchmark claims.

Research principle

A regression number is an observation, not a root cause. The project therefore separates:

measurement → release consequence → hypotheses → follow-up experiments

For example, a safety regression after SFT may motivate investigation of training-data distribution shift, conflicting supervision, objective overspecialization, output-length changes or serving configuration. Reward-model bias is considered only when a preference/reward stage actually exists.

Production release methodology

Critical deterministic safety/correctness failures remain authoritative. Performance, cost or ambiguous quality trade-offs trigger investigation. Optional LLM judging contributes subjective evidence but cannot overrule deterministic critical failures.

The system also provides:

• tamper-evident Phase 3 release bundles
• CI enforcement
• versioned benchmark datasets
• live Hugging Face provider evaluation
• production-trace ingestion
• evidence-linked model lifecycle transitions

Public evidence chain

• GitHub engineering source: https://github.com/h00w/model-quality-release-gate
• Interactive Space: https://huggingface.co/spaces/h0000w/model-quality-release-gate
• Evaluation dataset: https://huggingface.co/datasets/h0000w/model-quality-release-gate
• Portfolio case study: https://hendarmawan.se/projects/model-quality-release-gate/
• Agentic AI Academy: https://hendarmawan.se/agentic-ai/

Five-level production-AI proof

This project uses the Production AI Five-Level Proof Model v1:

L1 Runnable → L2 Reproducible → L3 Capability-Validated → L4 Production-Candidate → L5 Production-Validated.

The canonical engineering source computes the level with make proof and records the result in a machine-readable proof.json. The configured ceiling for this release-control project is L4. L5 is not claimed without target-production observation, SLO and recovery evidence.

Canonical specification: https://github.com/h00w/model-quality-release-gate/blob/main/PROOF_MODEL.md