# Weights of Giants (WOG) — AI Coding Agent & Developer Guide > 100% free, open-source Hardware-in-the-Loop (HIL) micro-vision benchmark & lineage registry. > Models run single-tenant on a physical, air-gapped Libre Computer La Frite (ARM Cortex-A53, 1024 MB RAM), orchestrated by a Raspberry Pi 5 Referee. - Website: https://weightsofgiants.com - Scoring Guide: https://weightsofgiants.com/info - Standings & Hardware Diagnostics: https://weightsofgiants.com/ranks - Chronicle Lineage Tree: https://weightsofgiants.com/archive - Open-Source Licensing Policy: https://weightsofgiants.com/licensing - Starter Kit & Local Verifier: https://github.com/seth935/wog-starter-kit --- ## 1. Active Challenge: Season Zero — Low-Energy Edge Vision (`wog-cv-v2.1`) - **Domain:** Fixed perimeter security, parking lots, and local NVR surveillance across 3 domains (`indoor`, `urban`, `rural`). Fixed camera viewpoint (no ego-motion compensation needed). - **Target Hardware (DUT):** Libre Computer La Frite (`la-frite-1gb`, AML-S805X-AC, quad-core ARM Cortex-A53 @ 1.2 GHz, 1024 MB DDR4 RAM, CPU-only inference). - **Package Cap:** `<= 40 MB` total model weights + manifest artifact. - **Memory Envelope:** `<= 800 MB` Peak VmRSS hard safety cap (`0%` composite weight in Season Zero because the Python/NumPy/OpenCV/ORT shim baseline dominates process RSS at `~273 MB`). - **Speed Envelope:** `>= 2.4 FPS` sustained end-to-end referee throughput (`fps_e2e`) required to place; `>= 10.0 FPS` earns full 100/100 speed credit (measured over the final 60 s after `>= 300 s` continuous load per domain, `frame_policy.skip = 0`). - **4-Pillar Composite Score (`wog-cv-v2.1`, gates calibrated on label version `gt-r4`, 0–100 per domain, averaged across Indoor/Urban/Rural):** - `40%` Detection — Per-clip averaged OBB mAP@50:95 (`>= 25%` gate) - `30%` Tracking — Track HOTA (`>= 30%` placement gate, `>= 50%` Champion recognition) - `25%` Speed — Sustained end-to-end FPS (`2.4–10.0 FPS` scale) - `5%` Robustness — 11-condition seeded stride-3 perturbation retention + 24-frame blank/null-control gate (`> 1` dirty frame fails) - Peak RAM `<= 800 MB` enforced strictly as a hard pass/fail safety gate (`0%` weight in Season Zero). - Domain score `>= 35` required to place. --- ## 2. Fixed On-Board Inference Contract (`dut_shim.py`) Weights of Giants NEVER executes arbitrary submitter Python scripts on the server or test rig. Your repository is declarative-only: - **Runtime:** `onnxruntime` (preloaded on the La Frite — do not vendor runtimes in your repo). Single-file `.onnx` weights (ONNX IR `<= 10`, opset `<= 22`; FP32, FP16, INT8, or mixed). - **Input Tensor:** RGB `float32` scaled `0.0–1.0` (`frame_rgb / 255.0`), NCHW layout `[1, 3, H, W]`, NO ImageNet mean/std normalization. Practical resolution: `input_size <= 512` (320x320 recommended for Cortex-A53). - **Output Tensor:** Only the model's first ONNX output tensor (`sess.run(...)[0]`) is read. Rows are: - Mode A (Managed `bytetrack-lite`): `[cx, cy, w, h, theta, score, class_id]` - Mode B (Native Track IDs): `[cx, cy, w, h, theta, score, class_id, track_id]` (`track_id >= 1`) - Coordinates `cx, cy, w, h` are normalized in `[0, 1]`. - Angle `theta` is in **radians** in `[-pi/2, pi/2)` (`0.0` is accepted for axis-aligned boxes). - Class IDs: `0` = `person`, `1` = `vehicle` (cars, trucks, buses, motorcycles, bicycles). - **Board NMS & Tracker Rules:** - Candidates with `score < 0.10` are dropped prior to NMS; at most `100` post-NMS detections per frame are kept. - Only `bytetrack-lite` is supported for managed tracking on the board. ### Declarative Manifest (`model.json` at repo root; `wog.yaml` also accepted) ```json { "runtime": "onnxruntime", "threads": 4, "domains": ["indoor", "urban", "rural"], "frame_policy": { "skip": 0, "resize": "letterbox" }, "stages": [ { "type": "detector", "model": "models/detector_int8.onnx", "input_size": [320, 320], "conf_threshold": 0.35, "iou_threshold": 0.45 }, { "type": "tracker", "algo": "bytetrack-lite", "params": { "track_thresh": 0.4, "track_buffer": 6, "match_thresh": 0.8 } } ] } ``` ### Hard Security & Licensing Exclusions - **No unsafe serialization files:** Repos containing `.pkl`, `.pickle`, `.pt`, `.pth`, `.joblib`, or custom executable entrypoints are rejected. - **Permissive open-source SPDX license required:** Only permissive SPDX licenses are accepted (`Apache-2.0`, `MIT`, `BSD-2-Clause`, `BSD-3-Clause`, `0BSD`, `ISC`, `Unlicense`, `CC0-1.0`, `Zlib`, `BSL-1.0`, `CC-BY-4.0`). Copyleft and share-alike licenses (`AGPL-*`, `GPL-*`, `LGPL-*`, `MPL-2.0`, `EPL-2.0`, `CC-BY-SA-*`, `ODbL-1.0`, `OpenRAIL` — including Ultralytics YOLO11/YOLOv8 derivatives) as well as non-commercial / proprietary licenses are rejected. - **Repository ownership:** The submitted GitHub or Hugging Face repository MUST belong to an account linked by the human submitter on Weights of Giants. --- ## 3. Recommended Agent Workflow (Zero-to-Draft) 1. **Start from a working base:** Clone `https://github.com/seth935/wog-starter-kit` (note: its bundled `models/detector_int8.onnx` is an untrained `~0% mAP` schema stub — replace it with a trained permissive detector) OR inspect permissive Genesis baselines on the Chronicle (`GET /api/submissions/archive?arena=vision`) to fork and optimize. 2. **Optimize for ARM Cortex-A53:** Prioritize small `input_size` (e.g. `320x320`), INT8/static quantization or lightweight backbones, Peak VmRSS `<= 800 MB` hard cap, and clean OBB NMS + `bytetrack-lite` thresholds (`track_buffer: 6` for `~2.5-6 FPS` clips). 3. **Score & verify locally before pushing:** ```bash pip install -r requirements-train.txt python score_local.py python verify_local.py ``` Remember: local desktop/laptop FPS does NOT equal Cortex-A53 La Frite FPS; official speed, RAM, and accuracy come only from the physical La Frite rig. 4. **Push to the human's GitHub or Hugging Face repo** with a valid `LICENSE` file and `model.json` at the repo root. 5. **Stage the candidate for human approval** using either method below (DO NOT attempt to bypass human approval or spend Run Tokens autonomously): ### Option A — Upload directly to the user's 10-Slot Profile Staging Area (Recommended when the user provides `WOG_STAGING_KEY`) Users can mint a scoped `stage_only` API key (`wog_stage_...`) on their profile (`/profile/#staging-area`). Stage-only keys can upload up to **10 active candidates** into the user's private staging queue, list slots, or discard slots — they can **never** approve runs or spend Run Tokens. ```bash curl -X POST https://weightsofgiants.com/api/staging/drafts \ -H "Authorization: Bearer $WOG_STAGING_KEY" \ -H "Content-Type: application/json" \ -d '{ "model_name": "EdgeOBB-INT8-v1", "model_description": "Quantized 320x320 OBB detector tuned for ARM Cortex-A53", "repo_url": "https://github.com/OWNER/REPO", "repo_commit": "OPTIONAL_7_TO_40_HEX_SHA", "arena": "vision", "source_type": "original", "parent_lineage": ["original"], "tags": ["int8", "obb"], "built_with": "claude-code", "agent_notes": "verify_local.py passed · 320x320 · 14.2 MB ONNX" }' ``` - `GET https://weightsofgiants.com/api/staging/drafts` — inspect current staged slots (`slots_used` / `max_slots: 10`). - `DELETE https://weightsofgiants.com/api/staging/drafts/{draft_id}` — discard a staged draft to free a slot. - Hard guards enforced on staging: 10-slot cap (`409` when full), linked GitHub/HF repo ownership, optional key-level `repo_lock`, duplicate staged commit check, SPDX license & exclusion checks, and optional inline `manifest` validation. - Once staged, tell the human to review and manually click **Approve & Submit** at `https://weightsofgiants.com/profile/#staging-area`. ### Option B — Zero-Key Pre-Filled Draft Review Link ```text https://weightsofgiants.com/?submit=1&repo=&name=&desc=&tags=&built_with= ``` Valid `built_with` slugs: `hand-tuned`, `claude-code`, `codex`, `cursor`, `antigravity`, `gemini`, `copilot`, `automl`, `other` When the human opens that link, the submission form pre-fills the safe draft fields, runs a live SPDX license check, and lets the human review lineage, accept the Open-Source Submission Agreement, and click once to spend 1 of their 3 free regenerating Run Tokens (+1 every 24h). --- ## 4. Public Read-Only API Endpoints - `GET https://weightsofgiants.com/api/submissions/leaderboard?arena=vision&category=composite` Categories: `composite`, `detection`, `tracking`, `footprint`, `main`, `stability`, `unique`. - `GET https://weightsofgiants.com/api/submissions/archive?arena=vision` Full verified Chronicle history with parent lineage and public diagnostics. - `GET https://weightsofgiants.com/api/platform/info` Active hardware profile and platform status. - `GET https://weightsofgiants.com/api/platform/check-repo-license?url=` Pre-check whether a public GitHub or Hugging Face repo passes SPDX license screening. --- ## 5. Season Zero Pilot Limitations (Transparency · Labels `gt-r4`) - Pilot ground truth (`gt-r4`) is machine-generated (Mask R-CNN ResNet50-FPN v2 masks fitted with `cv2.minAreaRect`, plus a Dual-Teacher `RT-DETR-R101` + Mask R-CNN OBB + temporal-median static-background filter on the 8 dense traffic & mixed CCTV clips) across 12 scored clips (`>= 200` GT boxes each) + 3 decoy canaries from 6 open-licensed surveillance video sources. - COCO-teacher agreement caveat: Because the Dual-Teacher labelers are COCO-trained, off-the-shelf COCO baselines agree partially with the teacher's person/vehicle definitions, whereas models fine-tuned on real CCTV footage may score lower than they should until human-verified labels replace `gt-r4` (keep this in mind for the `50%` Champion HOTA bar). - Peak RAM (`<= 800 MB`) is enforced strictly as a hard safety gate (`0%` composite weight) in Season Zero while the Python/NumPy/OpenCV/ORT shim dominates resident RSS (`~273 MB`). - Scores measure performance strictly under published Season Zero test conditions on the physical Libre Computer La Frite board.