README.md

Caption event memory

This directory contains the event memory generated by Experiment 007 for long videos. The primary artifact is events.jsonl event_details.jsonl keeps the complete validated memory and run lineage for later evaluation or development.

events.jsonl

Each line is one linked memory record. The level field identifies its role:

• event: one atomic action in the video
• phase: a group of adjacent events with one coherent subgoal
• video: the global summary of the video

Common fields:

• record_id: stable event, phase, or video record identifier

• video_id: source video identifier

• method: generation branch, normally b3_gebd_state

• time_span: [start_ms, end_ms] in source-video time

• text: canonical event caption, phase summary, or global summary

• parent_id: phase identifier for an event, or video summary identifier for a phase

Event records also contain:

• attributes: actor, action, affected and related objects, tools, location, source, destination, and observable before, after, and changed states
• support: directly observed facts, separate inference, uncertainty, evidence timestamps, and confidence

event_details.jsonl

There is one line per completed video. It contains the source path, run fingerprint, method status, and the original validated VideoEventMemory for each method. This preserves phase metadata, object observations, quality flags, processing statistics, and other provenance that is intentionally omitted from the compact event view.

Generation flow

The pipeline samples the video and runs EfficientGEBD to propose human-like event boundaries. It cleans duplicate or overly short intervals, then builds an evidence package with ordered frames before, during, and after each event, together with nearby OCR and object detections when available. A vision language model interprets each package without seeing any downstream question, separating observed actions and states from uncertain inference. The response is validated, linked to neighboring events, grouped into phases, and summarized at video level. The final validated memory is exported to the two JSONL files above; incomplete workspaces are removed only after both files are written successfully.