TRACE Open Source
Trajectory Analysis for Conversational Evidence
TRACE is the operational implementation path for AI Behavioral Trajectory Forensics. Where ABTF defines the methodology, TRACE provides the software workflow: how to move from raw conversational artifacts to reviewable outputs with documented chain of custody, explicit classifier configuration, and reproducible evidence packages.
Why TRACE exists
Conversational AI incidents are increasingly reviewed in settings that require more than screenshots and narrative summaries. Investigators need a toolchain that preserves source integrity, records analytic decisions, and makes the resulting package inspectable by another reviewer.
TRACE was built for that problem. The priorities are:
- repeatability
- auditability
- bounded hosted-model risk
- human-review visibility
- exportable evidence packages rather than opaque scores
Core capabilities
Transcript ingest and provenance
TRACE accepts multiple conversational evidence formats, normalizes them into a shared internal schema, and records source hashes and custody information before transformation. Current parser coverage includes JSON, CSV, plain text, court-style transcripts, AXIOM-style JSON exports, and UFED-style XML exports.
Classification workflow
TRACE classifies:
- system messages against the behavioral harm taxonomy
- user messages against the TRACE vulnerability scale derived from C-SSRS
- rolling interaction windows for trajectory-level interpretation
Classification can run through deterministic heuristics, mock hosted mode, hosted-provider inference, and manual review pathways with explicit override rationale capture.
Correlation and report export
From the classified transcript, TRACE computes repeatable findings such as inappropriate response rate, pattern distribution, and crisis failure rate. It then exports an evidence package containing machine-readable artifacts, verification output, manifests, audit logs, Markdown reporting, and PDF reporting.
Reliability and validation support
TRACE includes dual-coder import, inter-rater reliability computation, validation fixtures, and benchmark workflows. That makes it usable not only for case analysis but also for method evaluation and operational hardening.
Operational stance
TRACE is designed as a forensic workflow candidate, not as an autonomous judge.
- TRACE does not make final forensic determinations.
- TRACE does not replace expert review.
- TRACE is built to preserve provenance, overrides, and reviewability.
That stance is deliberate. In high-stakes harm analysis, hidden automation is a liability. TRACE keeps the analytical chain visible.
Open-source access
- GitHub: github.com/empathyethicist/trace
- Package name:
trace-forensics - Current documented version in the local project metadata:
0.1.3
Fast start
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
trace init --root ./trace-workspace
trace validate --reference ./validation/companion_incident.json --root ./trace-workspace/validation_runs
Relationship to the paper
The AI Behavioral Trajectory Forensics paper defines the method and evidentiary boundaries. TRACE carries that method into a usable software workflow for investigators, researchers, and expert reviewers who need more than an abstract description.
The paper gives the forensic logic. TRACE gives the operational toolchain.