code/README.md

# Annotation pipeline — code backup (EgoProactive synthetic data)

Code that produced `egoconv_dense_*` and `egolongqa_*` in this repo. Backed up 2026-08-17 because
the CPU VM that ran it (`217.18.55.12`) was destroyed and `/workspace` is not a persistent volume.

## ⚠️ Known loss
`annotate_full.py` — the resume-capable driver loop that ran the 235-clip dense pass — existed
**only on that VM and is gone**. It is reconstructible: it iterated `egoconv235_manifest.json`,
called the ambient_agent per clip with `annotation_prompt.txt`, wrote one JSON per clip, and
skipped clips whose output already existed (that skip-on-existing behaviour is how the run was
resumed after the OpenRouter 402). `launchers/run_dense.sh` and `resume_dense.sh` show exactly how
it was invoked. The closest surviving equivalents are `scripts/annotate_v2.py` / `annotate_videos.py`.

## Contents
- `ambient_agent/annotation_prompt.txt` — **the tuned t3 prompt** (the valuable artefact).
  Policy: setup-phase coverage from t=0, one cue per distinct step (no target count), collapse
  repetition, fire at onset, −0.5 s shift. Was **untracked** in the ambient_agent working tree.
- `ambient_agent/ANNOTATION_GUIDE.md` — operating notes.
- `ambient_agent/agent.py`, `llm.py` + `UNCOMMITTED.diff`, `BASE_COMMIT.txt` — local modifications
  on top of `ambient-intelligence-hq/ambient` (see BASE_COMMIT.txt); never committed upstream.
- `scripts/annotate_*.py` — annotation variants from `/workspace/ambient` (a repo with **no remote**).
- `launchers/*.sh` — how the dense + egolongqa runs were launched, resumed and transferred.

## Prompt-tuning history (do not re-learn this the hard way)
t2 overfit the tuning videos: 0.643 on tuning → **0.494 held-out, worse than baseline**. t3 was
chosen because setup-coverage + repetition-collapse generalised (~+0.02). Always validate a prompt
change on ≥5 held-out videos, never on the videos used to tune it.

## Orchestrator
`deepseek/deepseek-v4-flash-0731` via OpenRouter (API key from env, never hardcoded) + a
self-hosted NVFP4 Qwen3.6-27B vision model behind Docker vLLM.