docs/READING_GUIDE.md

Start here

1. Read the repository README for scope and the actual outcome.
2. Open GAME_LOG.md for the benchmark's game events.
3. Open MODEL_RESPONSES.md for readable agent responses.
4. Check basic_counts_and_invariants.log for the measured instruction count and Trace invariant checks passed.
5. Download the dataset to inspect individual CPU trace records.

Read individual CPU instructions and memory accesses

The converted .trace.zip is a ZIP of binary DynamoRIO records, not a text file. Unzipping it does not produce readable assembly. Use matching DynamoRIO 11.91.20708 on Linux x86-64. GPU inference is not in this trace.

After downloading the repository, run from its root:

export DYNAMORIO_HOME=/path/to/DynamoRIO-Linux-11.91.20708
SAMPLE=traces/sample_0000_instr_500000000_510000000/drmemtrace.python3.11.00361.2764.dir
"$DYNAMORIO_HOME/bin64/drrun" -t drmemtrace \
  -indir "$SAMPLE/trace/window.0000" -tool view -sim_refs 100 \
  > trace-preview.txt 2>&1

This requests a text view of the first 100 records, not 100 agent actions or necessarily 100 instructions. The view tool can still scan the entire input. The view command is provided for readers; a text instruction dump is not included in this release. The shipped counts/invariant results were actually run.

To reproduce validation of the already-converted window:

"$DYNAMORIO_HOME/bin64/drrun" -t drmemtrace \
  -indir "$SAMPLE/trace/window.0000" \
  -tool basic_counts:invariant_checker -jobs 1

The original validation converted from raw/window.0000. Passing only the process directory selects window 0 in this version; for multiple samples, validate every window explicitly. Keep shared raw module/encoding files with their process directory. Raw reconversion may need binaries from the original environment; the already-converted trace includes instruction encodings.

Reproduce collection

The exact executed scripts are under code/. run_v3.sh records the original Phoebe paths; edit paths and output names for your environment. Install the pinned AgentBench commit and dependencies, start a separate compatible model server, and set MODEL_URL. Do not point tracing at the GPU server.

The portable collector accepts the command after --; see python code/run_agentbench_samples.py --help. Native tracing options are:

-offline -split_windows -trace_after_instrs 500000000
-trace_for_instrs 10000000 -retrace_every_instrs 990000000

The 990M value is the untraced gap, so trace starts are 1B apart. Sample labels use requested half-open ranges. Actual endpoints are approximate.