Skip to content

runtime-forensics.md

Latest commit

 

History

History
93 lines (64 loc) · 2.48 KB

runtime-forensics.md

File metadata and controls

93 lines (64 loc) · 2.48 KB

MiroFish Runtime Forensics

Use this file when a simulation technically "ran" but you need to decide whether it actually produced meaningful behavior.

Primary Evidence Sources

Check these before trusting any final report:

  • state.json
  • simulation_config.json
  • run_state.json
  • twitter/actions.jsonl
  • reddit/actions.jsonl
  • twitter_simulation.db
  • reddit_simulation.db

Core Questions

Ask these in order:

  1. Did both enabled platforms actually start?
  2. Did rounds advance beyond initialization?
  3. Did multiple agents act, or only a tiny subset?
  4. Did the action mix evolve over time?
  5. Do files keep changing even if the UI looks stuck?

Signs Of A Healthy Run

  • both platforms produce actions if both were enabled;
  • action counts keep increasing over meaningful rounds;
  • more than one or two agents appear repeatedly;
  • the action mix is not just startup posts or idle behavior;
  • database files exist and are readable.

Signs Of A Trivial Or Misleading Run

  • only a few startup actions exist;
  • one platform is effectively silent;
  • the run ends quickly with little or no round progression;
  • the report exists, but raw artifacts are thin;
  • the UI says "running" while logs stop changing.

File-First Checklist

run_state.json

Check:

  • configured rounds;
  • current rounds;
  • stopped or completed status;
  • any mismatch between expected and actual progress.

Per-platform actions.jsonl

Check:

  • whether actions continue after initialization;
  • which agents appear most often;
  • whether the last actions are recent and varied;
  • whether one platform never really joined the run.

SQLite databases

First introspect the database instead of assuming table names:

SELECT name
FROM sqlite_master
WHERE type = 'table'
ORDER BY name;

Then inspect whichever action, post, comment, or trace tables actually exist in your run.

Stop-Early Conditions

Stop the run and fix inputs or route quality if:

  • action growth stalls early;
  • both platforms look repetitive or incoherent;
  • only initialization artifacts are present;
  • the model route keeps failing structured output or status updates.

Practical Classification

Classify the runtime before you move on to report analysis:

  • healthy: sustained multi-agent behavior, usable evidence
  • partial: some evidence exists, but one platform or one stage is weak
  • trivial: technically completed, but too little happened to trust the report

If the run is partial or trivial, document that explicitly.