behavior-transform.io

behavior-transform.io is a pre/post inference compliance intermediary. It sits between an authorized operator and an AI model provider, transforming sensitive operational context into language that falls within the provider's terms of service before inference, and stripping formatting artifacts from the response afterward.

Who this is for

Law enforcement, government, military, and authorized security operations teams that use AI models as part of their operational workflows. These entities conduct work that is legal, internally authorized, and conducted under formal rules of engagement, but that uses terminology and describes scenarios a general-purpose model provider will flag or refuse under its default content policy.

Enterprise agreements with providers exist for this purpose. Anthropic, OpenAI, and Google each offer government and enterprise tiers with scoped policy adjustments. Those agreements work, but they create a separate relationship per provider, each with its own onboarding, compliance surface, and failure modes. An organization using three providers maintains three policy boundaries, three sets of credentials, and three integration surfaces. When a provider tightens its content policy, every downstream consumer of that agreement absorbs the friction individually.

behavior-transform.io is a single compliance boundary. One integration point handles the vocabulary translation, authorization framing, and policy-gate logic for every provider behind it. One boundary means one failure point to diagnose, one configuration to audit, and one surface to maintain. It reduces the N-to-N relationship between authorized entities and model providers to a hub that both sides connect to.

What it is not

behavior-transform.io is not an offensive tool and does not perform offensive operations. It does not generate attacks, produce harmful content, or provide capabilities the operator does not already have authorization for.

It is not a jailbreak. A jailbreak attempts to make a model act against its safety training for any user. This tool translates authorized operational language into a register that aligns with provider terms of service, for operators whose authorization is established before the tool runs. The distinction is the authorization chain: if the operator does not hold a valid grant, the tool's gate system blocks the request.

How it works

The tool runs eight layers on every request before it reaches a model:

1. Category detection classifies the input across 14 content categories.
2. Vocabulary substitution replaces operational terminology with consulting-register equivalents (150+ rules).
3. Semantic modulation applies five layers of contextual rewriting.
4. Cyber modulation maps ATT&CK-phase terminology to assessment language.
5. Prose calibration normalizes register across the full text.
6. Compound reframing handles multi-clause constructions.
7. Authorization framing adds context establishing the operator's authorization basis.
8. Token optimization compresses the result for efficient inference.

On the return leg, the response passes through an output enforcer that strips formatting artifacts and preamble boilerplate, so the operator receives a clean operational response.

The pipeline scan (bt scan) evaluates a request without transforming it, reporting pressure (estimated refusal probability), policy gate decision, and friction score. This lets an operator preview whether a request needs transformation before sending it.

The authority system derives authorization from the deployment environment: machine fingerprint, capsule integrity, seal validity, and activation state. No external auth server or token exchange is needed. If the grant is invalid, the tool blocks the request.

Usage

python -m pip install -e .
bt --help
bt status --json
bt doctor --json

For full commands, see USAGE.md and GETTING-STARTED.md.

For developers

Run the flagship tests and doctor before changing any subsystem:

python -B -m pytest tests/test_behavior_flagship.py tests/test_behavior_delivery_contract.py -q
python -B tools/behavior_flagship.py doctor --json

Delivery-surface changes should also pass:

python -m public_surface_sweeper . --workspace --json

Modes

Mode	Profile	Behavior
ops	standard	Full calibration stack active
research	native	Source-faithful passthrough, no substitution
academic	native	Alias for research

Switch mode:

python tools/io_state.py --set research
python tools/io_state.py --set ops

Structure

tools/      Core modules and CLI wrappers
hooks/      Claude Code hook adapters
profiles/   Shell integration helpers
tests/      Pytest and unittest coverage
docs/       Specs, plans, and integration contracts

Verification

python -m pytest -q
bt doctor --json

See docs/INTEGRATION_CONTRACT.md for the IO boundary contract. See AGENTS.md for local agent instructions and USAGE.md for operator/developer commands.