|
|
--- |
|
|
pretty_name: Loko Release Gate Records |
|
|
task_categories: |
|
|
- object-detection |
|
|
language: |
|
|
- en |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
# Loko Release Gate Records |
|
|
|
|
|
Public, third-party anchored **Release Gate** decision artifacts for pinned vision model revisions. |
|
|
|
|
|
This repository publishes two runs that demonstrate the difference between: |
|
|
- **Run A (Baseline Repro):** reproducible benchmark surface on COCO |
|
|
- **Run B (Production Gate):** class-scoped operational gating under low-light conditions |
|
|
|
|
|
This is not a benchmark leaderboard claim. It is an audit-friendly release record pattern. |
|
|
|
|
|
> Note: Hugging Face’s Dataset Viewer may show a warning because this repo is an artifact record store (reports, manifests, receipts), not a dataset split intended for browsing. |
|
|
|
|
|
--- |
|
|
|
|
|
## What a “Release Gate Record” contains |
|
|
|
|
|
Each run includes: |
|
|
- Pinned **model repo + exact commit revision** |
|
|
- Pinned **dataset repo + exact commit revision** |
|
|
- Declared **release predicate** (the gate) |
|
|
- Deterministic **PASS/FAIL** container (certificate) |
|
|
- Metrics JSON (filled when the run is executed) |
|
|
- Offline HTML report + receipt bundle |
|
|
- Checkout manifest + signature |
|
|
- Full folder manifest (hash index) |
|
|
|
|
|
--- |
|
|
|
|
|
## Runs |
|
|
|
|
|
### Run A — Baseline Repro (COCO) |
|
|
|
|
|
**Goal:** Bind a release decision to pinned, reproducible inputs on a benchmark surface. |
|
|
|
|
|
- Model: `Ultralytics/YOLOv5` @ `553ca28641b8c2129596b0ecc50fc0902f7bd7ac` (`yolov5s.pt`) |
|
|
- Dataset: `rafaelpadilla/coco2017` @ `7d5136d6952da2df12d99106e2cce2a65324617d` (split: `validation`) |
|
|
- Gate: `coco_baseline_map50_floor_v1` (mAP@0.50 >= 0.36) |
|
|
|
|
|
Artifacts: |
|
|
- Report: [run-a-baseline-repro/report.html](run-a-baseline-repro/report.html) |
|
|
- Certificate: [run-a-baseline-repro/release_certificate.json](run-a-baseline-repro/release_certificate.json) |
|
|
- Metrics: [run-a-baseline-repro/evaluation_metrics.json](run-a-baseline-repro/evaluation_metrics.json) |
|
|
- Checkout manifest: [run-a-baseline-repro/checkout_manifest.json](run-a-baseline-repro/checkout_manifest.json) |
|
|
- Signature: [run-a-baseline-repro/checkout_manifest.sig](run-a-baseline-repro/checkout_manifest.sig) |
|
|
- Full hash index: [run-a-baseline-repro/loko_manifest.json](run-a-baseline-repro/loko_manifest.json) |
|
|
|
|
|
--- |
|
|
|
|
|
### Run B — Production Gate (Low-Light, class-scoped) |
|
|
|
|
|
**Goal:** Demonstrate production-style gating: PASS/FAIL is defined by operational predicates, not benchmark bragging. |
|
|
|
|
|
- Model: `Ultralytics/YOLOv5` @ `553ca28641b8c2129596b0ecc50fc0902f7bd7ac` (`yolov5s.pt`) |
|
|
- Dataset: `SatwikKambham/ex-dark` @ `f2a15374dae8c90395f461c5ba0546c034751665` |
|
|
- Scope: class-scoped to `person` |
|
|
- Gate: `exdark_person_operational_gate_v2_strict` at confidence 0.5 |
|
|
- precision >= 0.85 |
|
|
- recall >= 0.80 |
|
|
|
|
|
Artifacts: |
|
|
- Report: [run-b-production-gate/report.html](run-b-production-gate/report.html) |
|
|
- Certificate: [run-b-production-gate/release_certificate.json](run-b-production-gate/release_certificate.json) |
|
|
- Metrics: [run-b-production-gate/evaluation_metrics.json](run-b-production-gate/evaluation_metrics.json) |
|
|
- Checkout manifest: [run-b-production-gate/checkout_manifest.json](run-b-production-gate/checkout_manifest.json) |
|
|
- Signature: [run-b-production-gate/checkout_manifest.sig](run-b-production-gate/checkout_manifest.sig) |
|
|
- Full hash index: [run-b-production-gate/loko_manifest.json](run-b-production-gate/loko_manifest.json) |
|
|
|
|
|
--- |
|
|
|
|
|
## Quick verify in 30 seconds |
|
|
|
|
|
1) Open the **certificate** and confirm the decision: |
|
|
- Run A: PASS |
|
|
- Run B: FAIL |
|
|
|
|
|
2) Confirm the **receipt bundle** is fully indexed: |
|
|
- `checkout_manifest.json` lists `receipt_bundle/**` with sha256 + sizes |
|
|
|
|
|
3) Confirm the folder is fully indexed: |
|
|
- `loko_manifest.json` lists the full run folder (including the manifest + signature files) with sha256 + sizes |
|
|
|
|
|
4) Confirm the signer identity is pinned: |
|
|
- `PUBKEY_PIN.txt` contains the pinned pubkey hash |
|
|
- `operator_ed25519_public.pem` provides the demo public key used for signing |
|
|
|
|
|
--- |
|
|
|
|
|
## How to fetch this repo by revision |
|
|
|
|
|
```python |
|
|
from huggingface_hub import snapshot_download |
|
|
|
|
|
snapshot_download( |
|
|
repo_id="LokoAI/release-gate-records", |
|
|
repo_type="dataset", |
|
|
revision="main" |
|
|
) |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## Offline package (ZIP + .loko payload) |
|
|
|
|
|
This repo also publishes a single Windows zip that contains: |
|
|
- the two `.loko` payloads (Run A + Run B) |
|
|
- their `.loko.sha256` receipts |
|
|
- a lightweight verifier script |
|
|
- `sha256.txt` for the zip contents |
|
|
- `RUNME.txt` |
|
|
|
|
|
Download: |
|
|
- `ReleaseGate_win.zip` |
|
|
- `ReleaseGate_win.zip.sha256` |
|
|
|
|
|
### Windows (PowerShell) |
|
|
|
|
|
```powershell |
|
|
Expand-Archive .\ReleaseGate_win.zip -DestinationPath .\ReleaseGate -Force |
|
|
|
|
|
# Verify Run A |
|
|
python .\ReleaseGate\bin\verify_release_gate_loko.py --loko .\ReleaseGate\incoming\run-a-baseline-repro.loko --out .\REVIEW_run_a --pubkey-pin 1f61014bc8bf5bedc9389c5950d9a46c703c8d91701d762b892fd1a924dcd367 |
|
|
|
|
|
# Verify Run B |
|
|
python .\ReleaseGate\bin\verify_release_gate_loko.py --loko .\ReleaseGate\incoming\run-b-production-gate.loko --out .\REVIEW_run_b --pubkey-pin 1f61014bc8bf5bedc9389c5950d9a46c703c8d91701d762b892fd1a924dcd367 |
|
|
``` |
|
|
|
|
|
Verification performs: |
|
|
- `.loko.sha256` validation |
|
|
- Safe extraction |
|
|
- Pinned signer identity check (pubkey pin) |
|
|
- `checkout_manifest.sig` signature validation |
|
|
- `loko_manifest.json` file index verification (sha256 + size) |
|
|
|
|
|
Expected result: |
|
|
- Run A → PASS |
|
|
- Run B → FAIL (strict operational gate not satisfied) |
|
|
|