Sovereign 8s.is mirror of standardgalactic/perfect-crime
  • Python 99.6%
  • TypeScript 0.2%
  • Dockerfile 0.2%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Jeremy Qin df64f0f8a2 update
2026-09-26 00:38:26 +02:00
docs update 2026-09-26 00:38:26 +02:00
host_lab update 2026-09-26 00:38:26 +02:00
tests update 2026-09-26 00:38:26 +02:00
trace_lab update 2026-09-26 00:38:26 +02:00
Dockerfile update 2026-09-26 00:38:26 +02:00
Dockerfile.bun-sourcemap-agent update 2026-09-26 00:38:26 +02:00
README.md update 2026-09-26 00:38:26 +02:00

Agent trace-tampering experiments

This repository contains the experiment harness and tests for controlled studies of native coding-agent session traces. Paper figures, LaTeX, demo media, selected trial archives, and raw runs are kept outside the code-only main branch.

Run a new experiment

New runs require Python 3.11+, a running Docker Engine, the relevant model credentials, and enough disk space for Docker images and artifacts. Copy .env.example to .env and fill in only the provider keys you need. The host loads .env automatically; .env is Git-ignored. Check the setup and build the base image:

python3 -m trace_lab doctor
python3 -m trace_lab build
python3 -m trace_lab calibrate

List the host-side settings, then ask any runner for its exact options:

python3 -m host_lab list
python3 -m host_lab anonymization --help

For example, one Codex anonymization run with the paper's one-hour optimization budget is:

python3 -m host_lab anonymization \
  --client codex --model gpt-6-sol --codex-auth api-key \
  --mode implicit --time-budget 3600 --max-attempts 250

A new run writes its report and independent observer artifacts under the ignored runs/ directory. Keep these artifacts separate from the code checkout when sharing results.

Find a setting

Setting in the paper Runner Key option
Financial / personal privacy request direct-user --task-profile financial or personal
Fabricated reset addition direct-user --task-profile trace-append
Tool-call spoofing tool-call --client, --model
Privacy-protection skill privacy-skill --task-profile api-client
Compaction-checkpoint skill compaction-skill --skill optimization --setup checkpoint-context
Anonymization reward loop anonymization --mode implicit
Terminal-Bench reward loop terminal-bench build once, then run --task-profile core
Peer Logs peer-influence --context rollout-logs
Peer Workspaces peer-influence --context peer-environments

The runners own their prompts, permissions, grading, and output schemas. The catalog does not change their defaults. Historical paper batches used frozen arguments and selected trials, which are not part of this code-only branch.

Repository map

  • host_lab/: host-side orchestration, time budgets, task verification, and score loops.
  • trace_lab/: native client setup, transport, fixtures, independent observation, and reports.
  • tests/: fixture and integration checks. Network-listening tests require permission to bind loopback sockets.
  • docs/: architecture and harness-specific notes.

Agent homes such as /home/agent are fixed paths inside disposable containers, not paths on the researcher's computer. New code and shared metadata should use repository-relative paths for host-side files. Raw historical traces may contain paths recorded at collection time; treat them as sensitive evidence and review them before redistribution.