Update card: Sep-8 gate results (RICH IoU 0.407), pending full 620v run, Space 402 status

README.md

@@ -67,3 +67,12 @@ A 23-dim prosody rule predicts those labels at 0.994 on **unseen videos** — ab
 - Pipeline notebook (execution-verified): [Colab v19](https://colab.research.google.com/gist/Das-rebel/188a3bc5d4346c8189372f00c8bc2d39/Colab_Complete_Pipeline_v19.ipynb)
 - Recovered July-16 dataset (for the historical analysis): `wavlm_training_data_expanded.npz` — archived in the ChuckleNet project backups.
 - All claims in this card trace to JSON artifacts in the [ChuckleNet repo](https://github.com/Das-rebel/ChuckleNet/tree/main/docs).
+
+---
+## Update 2026-09-08 (v20 gate complete, full 620v in progress)
+- **New validated numbers (Tier-1, 20 curated videos, Colab v20.2d):** FULL IoU-F1@0.2=0.197 (PR-AUC 0.208, 7.91% pos); RICH subset (9 videos, 14.2% pos) = **IoU-F1@0.2=0.407**, F1=0.338, PR-AUC=0.376.
+- The RICH number is the first weak-label result above the naive 0.29 line and the project's 0.310 (pre-sweep 118v best was later corrected to 0.3302 with IoU@0.2 merge gap 0.8).
+- **Provenance:** results saved locally + Drive (gdrive:chuckle_net_results/). Notebook: `ChuckleNet_Kaggle_v21.ipynb` (Kaggle P100 run) + `ChuckleNet_Final_Colab_v20.ipynb` (Colab, GATE_N=0 pre-set).
+- **Pending:** full 620-video run (Colab pending GPU reset; Kaggle v3 queued/pushed; resume from 20-video checkpoint; expected ~6-9h per mode, results saved to `results_v21.json`).
+- **Label tier reminder:** these are Tier-1 (caption-marker weak labels). Before any gold claims must complete Tier-2 (StandUp4AI-118v anchor vs 0.3302) and Tier-3 (Gillick-162v anchor vs 0.559) — see docs/LABEL_HIERARCHY.md.
+- **Space status:** `Hayasuki/chucklenet` space (404/empty) — previous 402 prevented full deployment; deployment deferred until full-run results confirm a publishable scale-up.