TRACT — Translating Requirements Across CRE Trees

TRACT assigns security framework controls to positions in the OpenCRE hierarchy using a fine-tuned bi-encoder, creating transitive crosswalks between any pair of frameworks automatically.

The Problem

Security frameworks define overlapping requirements independently. NIST 800-53, MITRE ATLAS, OWASP, CSA, and the EU AI Act each describe AI security controls in their own terminology. Practitioners manually crosswalk between them — a process that is slow, error-prone, and breaks every time a framework updates.

The Solution

TRACT treats crosswalk construction as a hub assignment problem: each control is independently mapped to a CRE hub — a node in OpenCRE's universal security taxonomy. Controls from different frameworks that map to the same hub are crosswalked transitively.

g(control_text) → CRE_hub_position     # NOT pairwise f(A, B) → similarity

This scales linearly (not quadratically) with the number of controls, and adding a new framework automatically crosswalks it with every existing framework.

flowchart LR
    A["Raw framework<br/>(PDF, CSV, JSON)"] --> B["tract prepare"]
    B --> C["Standardized JSON"]
    C --> D["tract validate"]
    D --> E["tract ingest"]
    E --> F["CRE hub<br/>assignment"]
    F --> G["Crosswalk DB"]
    G --> H["tract export"]

    style A fill:#f9f,stroke:#333
    style G fill:#9f9,stroke:#333

Loading

What is OpenCRE? The Open Common Requirement Enumeration is a community-maintained taxonomy that organizes security requirements into a hierarchy of 400+ hubs. It links controls from NIST, OWASP, CWE, ISO 27001, and dozens of other frameworks. TRACT uses it as the universal coordinate system for security controls.

Key Results

Model accuracy figures are withdrawn pending re-measurement (2026-08-15). The previously published hit@1 and its interval did not survive audit: the interval was arithmetic on the point estimate rather than the bootstrap it claimed, the per-fold deltas do not reproduce, and on a larger held-out roster of 1,265 controls fine-tuning showed no improvement over the zero-shot base model (macro delta -0.0004, p = 0.98). See the erratum on the model card. No replacement figure is quoted until a run clears a pre-registered gate.

Metric	Value
Assignment accuracy (hit@1)	withdrawn, see above
Improvement over zero-shot	withdrawn, see above
Crosswalk assignments	5,238 across 31 frameworks
Assignment breakdown	4,390 ground truth · 528 expert-reviewed · 320 model predictions
AI↔traditional bridges	46 accepted (of 63 candidates)
Evaluation	LOFO cross-validation with hub firewall

The counts above are inventory, not model accuracy. They are not a quality claim either: the large majority of published assignments are OpenCRE's own ground-truth links, imported rather than reviewed here, and the reviewed portion is the 878 model predictions judged by a single reviewer whose agreement against 20 hidden calibration items was 13/20. Inter-rater reliability is not measured. An earlier version of this paragraph, written the same day as the withdrawal above, claimed every assignment was expert-reviewed. That was wrong and is withdrawn. Full experiment narrative in tract_experimental_narrative.ipynb.

Quick Start

Explore without model artifacts (works immediately after install):

git clone https://github.com/rocklambros/TRACT.git
cd TRACT
pip install -e ".[dev]"
tract prepare --file examples/sample_framework.csv --framework-id demo --name "Demo Framework"
tract validate --file examples/sample_framework_prepared.json

Full assignment workflow — tract assign downloads the model on first use, no manual setup:

pip install -e ".[phase0]"
tract assign "Implement input validation for AI model training data"

The first tract assign downloads the pinned model (~1.3 GB) from HuggingFace into the shared HuggingFace cache, verifies its checksums, then runs the prediction. Later runs reuse the cache. tract --version prints the package version and the pinned model revision in use.

To pre-fetch the model (for CI, an airgapped host, or to also pull the crosswalk database), run tract download first; tract assign then uses what is already there.

No TensorFlow needed. TRACT runs on PyTorch only. It sets USE_TF=0 at import so transformers never loads a TensorFlow that happens to be installed alongside it (a TF import deadlocks sentence-transformers on some macOS setups). Set USE_TF=1 to override.

Framework Coverage

TRACT processes 31 frameworks with 2,802 controls total.

AI Security Frameworks (12 frameworks, 977 controls)

Framework	ID	Controls
CSA AI Controls Matrix	csa_aicm	243
MITRE ATLAS	mitre_atlas	202
AIUC-1 Standard	aiuc_1	132
EU AI Act	eu_ai_act	126
NIST AI Risk Management Framework	nist_ai_rmf	72
CoSAI AI Security Risk Map	cosai	55
OWASP AI Exchange	owasp_ai_exchange	54
EU GPAI Code of Practice	eu_gpai_cop	40
OWASP GenAI Data Security	owasp_dsgai	21
NIST AI 600-1 GenAI Profile	nist_ai_600_1	12
OWASP Top 10 for LLM	owasp_llm_top10	10
OWASP Top 10 for Agentic Apps	owasp_agentic_top10	10

Traditional Security Frameworks (19 frameworks, 1,825 controls)

Framework	ID	Controls
CAPEC	capec	349
NIST 800-53	nist_800_53	300
ASVS	asvs	277
CWE	cwe	246
DSOMM	dsomm	183
ISO 27001	iso_27001	93
WSTG	wstg	59
OWASP Cheat Sheets	owasp_cheat_sheets	50
NIST SSDF	nist_ssdf	44
ENISA	enisa	38
SAMM	samm	30
CSA Cloud Controls Matrix	csa_ccm	29
NIST AI 100-2	nist_ai_100_2	28
ETSI	etsi	27
NIST 800-63	nist_800_63	25
BIML	biml	20
OWASP Top 10 2021	owasp_top10_2021	10
OWASP Proactive Controls	owasp_proactive_controls	10
OWASP Top 10 for ML	owasp_ml_top10	7

CLI Overview

All 20 subcommands grouped by workflow stage:

Stage	Commands	Description
Setup	download	Optional pre-fetch of model + crosswalk database (assign auto-downloads the model on first use)
Explore	tutorial hierarchy compare	Learn TRACT, inspect hubs, compare frameworks
Prepare	prepare validate	Extract and validate framework controls
Assign	assign ingest accept	Map controls to CRE hubs
Review	review-export review-validate review-import review-proposals	Expert review workflow
Analyze	bridge propose-hubs import-ground-truth	Discover connections, suggest new hubs
Export	export export-canonical	Output assignments (CSV, JSON, OpenCRE, canonical snapshot)
Publish	publish-hf publish-dataset	Release model and dataset to HuggingFace

See docs/cli-reference.md for full options and examples.

Project Structure

flowchart TD
    subgraph INPUT["Input"]
        RAW["data/raw/<br/>Immutable source files"]
        API["OpenCRE API"]
    end

    subgraph PARSE["Parse"]
        P["parsers/<br/>12 framework parsers"]
        PREP["tract prepare<br/>LLM-assisted extraction"]
    end

    subgraph CORE["Core"]
        PROC["data/processed/<br/>Standardized JSON"]
        TRAIN["data/training/<br/>Hub links"]
        T["tract/<br/>Core library"]
    end

    subgraph OUTPUT["Output"]
        RES["results/<br/>Metrics, reviews"]
        BUILD["build/<br/>HF staging"]
        DB["crosswalk.db"]
    end

    RAW --> P
    RAW --> PREP
    API --> TRAIN
    P --> PROC
    PREP --> PROC
    PROC --> T
    TRAIN --> T
    T --> RES
    T --> BUILD
    T --> DB

Loading

Where to Go Next

I want to...	Go to...
Try the examples	Examples
Add a new framework	Framework Guide
Understand the model and methodology	Architecture
Look up a command	CLI Reference
Look up a term	Glossary
Understand canonical export schema	Architecture § Canonical Export
Review hub descriptions for quality	Hub Description Review Guide
Contribute code	Contributing
Report a security issue	Security Policy

Published Artifacts

• Model: rockCO78/tract-cre-assignment on HuggingFace
• Dataset: rockCO78/tract-crosswalk-dataset on HuggingFace
• Experimental narrative: tract_experimental_narrative.ipynb — 14-section Jupyter notebook covering the complete research journey

Canonical Export (OpenCRE RFC)

TRACT produces per-framework canonical JSON snapshots designed for OpenCRE's incremental import RFC. Each framework export contains:

• snapshot.json — A StandardSnapshot with all controls, CRE mappings, filter policy, and a SHA-256 content hash for integrity verification
• changeset.json — A keyed diff against the prior export with 6 operation types (ADD/UPDATE/DELETE for controls and mappings), impact analysis, and scope classification
• embeddings.npz (optional) — Per-framework embedding slice from the deployment model

StandardSnapshot Schema

Field	Type	Description
schema_version	string	Always "1.0"
framework_id	string	Framework identifier (e.g., csa_aicm)
framework_name	string	Human-readable name
export_date	string	ISO 8601 UTC timestamp
content_hash	string	SHA-256 of all non-volatile fields
tract_version	string	Git SHA of TRACT at export time
model_adapter_hash	string	SHA-256 of the LoRA adapter weights
filter_policy	object	Confidence floor, OOD/ground-truth exclusion rules
controls	array	CanonicalControl objects (id, title, description, hyperlink)
mappings	array	CREMapping objects (control→hub with confidence, rank, provenance)

Changeset Operations

Operation	Description
ADD_CONTROL	New control added to framework export
UPDATE_CONTROL	Control title, description, or hyperlink changed
DELETE_CONTROL	Control removed from export
ADD_MAPPING	New control→hub mapping added
UPDATE_MAPPING	Mapping confidence, rank, or provenance changed
DELETE_MAPPING	Mapping removed

# Preview canonical export
tract export-canonical --dry-run

# Export a single framework with embeddings
tract export-canonical --framework csa_aicm --with-embeddings

# Export all eligible frameworks
tract export-canonical

See docs/cli-reference.md for full options.

License

CC0 1.0 Universal — dedicated to the public domain. That covers TRACT's own code, schemas, parsers and derived mappings.

Third-party framework content keeps its publisher's terms. NOTICE lists every framework with its licence and upstream source, names the two sources whose text is not redistributed here at all, and records the ones whose terms are still undetermined.