# Submission Generation Notes

Final variant: `memory_agent_top6_clean_light_plus`.

The pipeline uses a V36/MMDuet2-style reader for timing proposals, then regenerates concise candidate answers per candidate time using causal frames only. An external rule-based Memory Agent keeps only question-relevant textual states and filters redundant or contaminated turns. The final Clean-Light-Plus step removes only obvious generation artifacts.

## Main stages

1. Prepare official test videos and questions.
2. Clone public MMDuet2 and patch `proactive_eval/inference.py`.
3. Generate/reuse V36 raw timing proposals.
4. Build up to 6 candidate turns per video, with qtype-aware caps.
5. Regenerate answers per candidate time using V4_FINAL_K1 cleaner.
6. Apply Memory Agent novelty/relevance filtering.
7. Apply Clean-Light-Plus artifact filtering.
8. Run final JSONL sanity check.

## Clean-Light-Plus removed patterns

The filter removes global non-answer artifacts such as:

- `NO REPLY`
- `NS REPLY`
- `NSWYR`
- `REPLYING:`
- `no description available`
- `not available from this image alone`
- `from the provided audio description alone`
- `no image displayed`
- `no active movement occurs`
- `no actions take place`
- `no reaction shown`
- `no response available`
- `no activity related`
- `no additional context`
- `no additional elements`

It never deletes all turns for a video; at least one turn is preserved.

## Final sanity check

```text
records: 500
unique ids: 500
dup ids: 0
empty: 0
avg turns: 3.03
turn dist: Counter({3: 171, 2: 171, 4: 85, 5: 33, 6: 23, 1: 17})
time min/max: 4.0 / 30.0
chronological: True
bad field: True
```

No hidden labels, GT reply spans, leaderboard feedback, or video-id-specific manual edits are used.