# Sample LLM traces — Sharing-is-Caring badge

These JSONL files are unedited captures of every LLM call from real
playthroughs of *The Apprentice*. One JSON object per line, schema
documented in [`../README.md`](../README.md).

## Files

| File | Records | Calls | Notes |
|---|---|---|---|
| `three-playthroughs-fantasy-en.jsonl` | 117 | All `oracle-wizard-lora` (the fine-tune) | Three back-to-back playthroughs in one Python process. Default `Tobin / the Hollow` hero+village, English, fantasy theme. ~422k prompt tokens / ~43k output tokens total. 90 JSON-mode calls (resolutions + fork picks) + 27 text-mode calls (interludes + epilogue + theme expansion). |

## Quick exploration

Inspect with `jq`:

```bash
# Distinct models served
jq -r '.model_returned' three-playthroughs-fantasy-en.jsonl | sort -u

# Just the responses
jq -r '.response' three-playthroughs-fantasy-en.jsonl | head -3

# Calls by mode
jq -r '.mode' three-playthroughs-fantasy-en.jsonl | sort | uniq -c

# Latency distribution (ms)
jq -r '.latency_ms' three-playthroughs-fantasy-en.jsonl | sort -n | awk '
  {a[NR]=$1; sum+=$1} END {
    print "min:",a[1],"p50:",a[int(NR/2)],"p95:",a[int(NR*0.95)],"max:",a[NR],"mean:",sum/NR
  }'
```

## How these were generated

LLM-call tracing is default-on in this app. Set
`ORACLES_TRACE_DISABLE=1` to opt out. Each session writes to
`traces/oracles-trace-<12-char-hex>.jsonl`. The
[`curate_trace`](../../scripts/curate_trace.py) script copies one of
those session files into this directory under a friendlier label.