Vauban

What changes when models change?

Open-source behavioral diffing for language models.

Vauban helps researchers and engineers audit model transformations: fine-tunes, checkpoint updates, merges, prompt wrappers, steering interventions, quantization variants, and post-training runs. It asks what changed behaviorally, what evidence supports that claim, and how strong the claim can be given the access available.

Model transformations are the object. Access-aware auditing is the method. Vauban Reports are the artifact.

Named after Sébastien Le Prestre de Vauban, the military engineer who worked both siege and fortification. Vauban does the same for model behavior: map the boundary, test the change, inspect the evidence, and produce a report.

What it does

Vauban is a TOML-first CLI for model behavior change reports:

• compare behavior across model states, prompts, and interventions
• measure refusal, over-refusal, uncertainty, compliance, and side effects
• inspect activation-space evidence when it helps explain a behavioral change
• state the access level behind each claim instead of over-reading the evidence
• produce JSON and Markdown Model Behavior Change Reports that can be shared and rerun
• keep controlled interventions, defenses, and stress tests as supporting tools

The primary interface is not a pile of subcommands. It is:

vauban <config.toml>

All pipeline behavior lives in the TOML file.

Vauban is related to model diffing, but the product surface is narrower and more practical: behavioral diffs and model behavior change reports. Model-diffing methods are useful when they help answer what changed, why it matters, and what evidence supports the claim.

The core question is practical model-change management:

• We fine-tuned a model. Did it become weird?
• We changed a prompt template. Did safety regress?
• We quantized this model. Did behavior drift?
• This new checkpoint benchmarks better. What did it sacrifice?
• Can we ship this model update without surprises?

Conceptually, the center is:

vauban diff model_before model_after --suite behavior_suite.toml --report report.md

The durable interface remains TOML-first: configs should encode the same comparison, suite, evidence, limitations, and output report in a shareable file.

Access-aware auditing

Vauban should not make the same claim from every evidence source. The report language depends on what you can observe:

Access	What Vauban can support	Claim strength
One model or endpoint snapshot	Behavioral profile	"This is what the model did under this suite."
Two output traces or run reports	Behavioral diff	"Behavior changed across these observed snapshots."
Endpoint with logprobs	Distributional diff	"Token probabilities shifted in these cases."
Local weights and activations	Activation diagnostics	"This internal signal correlates with the behavior."
Base plus transformed model	Model-change audit	"This transformation changed behavior and internals this way."

The no-base-model problem is not the product. It is the discipline: if the base model, training data, checkpoints, logits, or activations are unavailable, the report must say so and narrow its conclusions.

The longer thesis is in What Changes When Models Are Changed?.

What Vauban is not

Vauban is not a jailbreak toolkit, not a safety-bypass toolkit, and not a claim that a model is safe. It is intended for responsible model behavior auditing, defensive analysis, and reproducible research on model changes.

Requirements

• Apple Silicon Mac for the MLX backend, or Linux with an NVIDIA GPU for the PyTorch backend
• Python 3.12+
• Pixi for reproducible local and CI environments

Install

Install the project environment from this repo:

pixi install -e torch-dev

On a Linux NVIDIA machine, verify the Torch/CUDA backend:

pixi run -e torch-dev backend-cuda

On Apple Silicon, use the MLX environment:

pixi run -e mlx-dev backend

Run checks through the backend environment you are using:

pixi run -e torch-dev check-torch
pixi run -e mlx-dev check

check-torch is the current Torch backend-contract gate. The full legacy test suite still runs in the MLX environment while backend-neutral coverage is being expanded.

Run the CLI through Pixi:

pixi run -e torch vauban --help
pixi run -e mlx vauban man workflows

Use backend = "torch" in a TOML config for CUDA/PyTorch runs. Omit it, or set backend = "mlx", for the MLX path. If backend is omitted, the active VAUBAN_BACKEND environment variable selects the runtime.

Contributor workflow and repo policy live in CONTRIBUTING.md.

Quick Start

Pick a goal — the manual will tell you which sections to use:

vauban man workflows

Then scaffold a config and go:

vauban man quickstart

Scaffold a starter config:

vauban init --mode default --output run.toml

Scaffold a named indirect prompt-injection benchmark:

vauban init --scenario share_doc --output share_doc.toml

Or run one of the checked-in canonical benchmarks directly:

vauban --validate examples/benchmarks/share_doc.toml
vauban examples/benchmarks/share_doc.toml

Or keep the benchmark choice directly in TOML and override only what you need:

[environment]
scenario = "share_doc"
max_turns = 5

If you want Vauban to judge a flywheel run against an explicit deployment goal, declare that goal in TOML too:

[objective]
name = "customer_support_gate"
deployment = "customer_support"
summary = "Preserve support quality while resisting refund abuse."
access = "api"
preserve = ["explain refund policy", "escalate billing issues"]
prevent = ["unauthorized refund", "PII disclosure"]

[[objective.safety]]
metric = "evasion_rate"
threshold = 0.05

[[objective.utility]]
metric = "utility_score"
threshold = 0.90

Today, [objective] is enforced quantitatively by [flywheel] runs and recorded in flywheel_report.json as an explicit pass/fail assessment.

If preserved utility should come from a fixed benign inquiry set instead of the generated flywheel worlds, add:

benign_inquiry_source = "dataset"
benign_inquiries = "data/customer_support_benign.jsonl"

All pipeline modes can be scaffolded. See the full list with:

vauban init --help

The full checked-in benchmark pack lives under examples/benchmarks/.

Validate before a real run:

vauban --validate run.toml

Then run the pipeline:

vauban run.toml

By default, output goes to output/ relative to the TOML file.

Model Behavior Change Reports

The flagship artifact is a Model Behavior Change Report: a reproducible behavioral change report for a model update, fine-tune, prompt/template change, controlled intervention, quantization variant, or post-training run.

Reports are TOML-driven. A minimal standalone report does not need to load a model; it assembles declared evidence into JSON and Markdown:

[behavior_report]
title = "Refusal Boundary Model Behavior Change Report"
target_change = "base -> instruction-tuned"
findings = [
  "Target-task performance improved.",
  "Refusal behavior changed.",
  "Over-refusal increased in ambiguous benign cases.",
  "Uncertainty expression decreased.",
]
recommendation = "Do not deploy without additional benign-request regression testing."
limitations = [
  "Small suite; not a safety certification.",
  "Representative examples are safe or redacted.",
]

[behavior_report.baseline]
label = "base"
model_path = "mlx-community/example-base"
role = "baseline"

[behavior_report.candidate]
label = "instruct"
model_path = "mlx-community/example-instruct"
role = "candidate"

[behavior_report.suite]
name = "refusal-boundary"
description = "Measures refusal, over-refusal, and ambiguous compliance."
categories = ["safety_refusal", "benign_request", "ambiguous_request"]
metrics = ["safety_refusal_rate", "over_refusal_rate"]

[[behavior_report.metrics]]
name = "safety_refusal_rate"
model_label = "base"
category = "safety_refusal"
value = 0.52
polarity = "higher_is_better"

[[behavior_report.metrics]]
name = "safety_refusal_rate"
model_label = "instruct"
category = "safety_refusal"
value = 0.84
polarity = "higher_is_better"

Run it like any other Vauban config:

vauban --validate examples/behavior_report.toml
vauban examples/behavior_report.toml

The output is behavior_report.json plus behavior_report.md in [output].dir. See examples/behavior_report.toml for a fuller safe/redacted example.

Minimal TOML

This is the minimal config the code accepts for the default pipeline:

[model]
path = "mlx-community/Llama-3.2-3B-Instruct-4bit"

[data]
harmful = "default"
harmless = "default"

[model].path is required for model-loading runs. Standalone report modes such as [behavior_report] do not need [model].

[data].harmful and [data].harmless are required for most runs. "default" uses Vauban's bundled prompt sets.

You can also choose the output directory explicitly:

[output]
dir = "runs/baseline"

Experiment Tech Tree

Vauban has built-in experiment lineage tracking through an optional [meta] section. This metadata does not change pipeline execution. It exists so you can organize runs as a tech tree.

Minimal example:

[meta]
id = "cut-alpha-1"
title = "Baseline cut, alpha 1.0"
status = "baseline"
parents = ["measure-v1"]
tags = ["cut", "baseline"]
date = 2026-03-02
notes = "First stable reference run."

Verified status values are:

archived, baseline, dead_end, promising, superseded, wip

If [meta].id is omitted, Vauban uses the TOML filename stem.

Render the tree from a directory of TOML configs:

vauban tree experiments/
vauban tree experiments/ --format mermaid
vauban tree experiments/ --status promising
vauban tree experiments/ --tag gcg

Each run also appends an experiment_log.jsonl file inside the configured output directory with the resolved config path, pipeline mode, report files, metrics, and selected [meta] fields.

How TOML Drives Vauban

vauban <config.toml> loads one TOML file and decides what to do from the sections you include.

The default path is:

1. measure
2. cut
3. export

You extend that run by adding more sections. Common examples:

• [behavior_report] creates a standalone model behavior change report from declared evidence.
• [eval] adds post-cut evaluation reports.
• [surface] adds before/after refusal-surface mapping.
• [detect] adds hardening detection during measurement.
• some sections switch Vauban into dedicated mode-specific runs instead of the default pipeline.

If you know what you want to do but not which sections to use:

vauban man workflows

For field-level reference on any section:

vauban man cast
vauban man softprompt
vauban man measure

For mode precedence:

vauban man modes

Commands You Will Actually Use

Inspect the manual:

vauban man workflows
vauban man quickstart
vauban man cast
vauban man all

Scaffold configs:

vauban init --help
vauban init --mode default --output run.toml
vauban init --mode probe --output probe.toml

Validate config and prompt files without loading model weights:

vauban --validate run.toml

Export the current JSON Schema for editor tooling:

vauban schema
vauban schema --output vauban.schema.json

Compare two run directories as a utility check:

vauban diff run_a run_b
vauban diff --format markdown run_a run_b
vauban diff --threshold 0.05 run_a run_b

vauban diff --threshold ... is a CI gate: it exits non-zero if any absolute metric delta crosses the threshold. For durable behavior-change artifacts, prefer a TOML [behavior_report] config so the comparison, evidence, examples, limitations, and provenance travel together.

Render the experiment lineage tree:

vauban tree experiments/
vauban tree experiments/ --format mermaid
vauban tree experiments/ --status promising

Data Formats

Verified by the generated manual:

• prompt JSONL for [data] and [eval]: one JSON object per line with a prompt key
• surface JSONL for [surface].prompts: requires label and category, plus either prompt or messages
• refusal phrase files: plain text, one phrase per line
• relative paths resolve from the TOML file's directory

Minimal prompt dataset example:

{"prompt":"What is the capital of France?"}
{"prompt":"Write a haiku about rain."}

Notes On Verification

This README is aligned to the code in this repo:

• package name: vauban
• console script: vauban = vauban.__main__:main
• verified commands: vauban <config.toml>, --validate, schema, init, diff, tree, man
• verified standalone report mode: [behavior_report]
• verified manual topics and scaffolded modes were checked against the live CLI help and generated manual

The current README previously had some stale mode/output claims; this version removes those and points readers to vauban man ... for the parts generated directly from code.

Python API (Session)

For programmatic use, the Session class wraps a loaded model with tool discovery, prerequisite tracking, and structured results.

from vauban.session import Session

s = Session("mlx-community/Qwen2.5-1.5B-Instruct-bf16")
s.tools()           # discover all capabilities
s.guide("audit")    # step-by-step workflow
s.describe("cast")  # detailed tool info with current status
s.catalog()         # all tools grouped by category

Tools

Method	Returns	What it does
s.measure()	DirectionResult	Extract refusal direction from activations
s.detect()	DetectResult	Check if model is hardened against abliteration
s.audit(thoroughness=...)	AuditResult	Full red-team: jailbreak + softprompt + surface + guard
s.evaluate()	EvalResult	Refusal rate + perplexity + KL divergence
s.probe("prompt")	ProbeResult	Per-layer projection onto refusal direction
s.scan("text")	ScanResult	Per-token injection detection
s.surface()	SurfaceResult	Map refusal boundary across prompt categories
s.cast("prompt", threshold=0.3)	CastResult	Conditional activation steering (defense)
s.sic(["prompt", ...])	SICResult	Iterative input sanitization (defense)
s.steer("prompt", alpha=-1.0)	str	Unconditional activation steering
s.cut(alpha=1.0)	dict[str, Array]	Remove refusal direction from weights
s.export("output/")	str	Save modified model to disk
s.classify("text")	harm scores	Score against 13-domain harm taxonomy
s.score("prompt", "response")	score result	5-axis quality assessment
s.report()	str	Markdown report from audit findings

Result Types

DirectionResult (from measure): direction (Array, shape d_model), layer_index (best layer), cosine_scores (per-layer separation), d_model, model_path.

CastResult (from cast): text (generated output), interventions (tokens where CAST steered, 0 = defense didn't engage), considered (total tokens), projections_before/projections_after (per-layer).

SICResult (from sic): prompts_clean (sanitized text), prompts_blocked (bool per prompt), initial_scores/final_scores (direction projection), total_blocked/total_sanitized/total_clean.

DetectResult (from detect): hardened (bool), confidence (0.0-1.0), effective_rank (>1.5 suggests hardening), evidence (list of strings).

AuditResult (from audit): overall_risk ("critical"/"high"/"medium"/"low"), findings (list of AuditFinding), jailbreak_success_rate, softprompt_success_rate, surface_refusal_rate.

Decision Guide

I want to...	Use
Understand what a model refuses	measure() then surface()
Check if a model is hardened	detect()
Full safety audit	audit() then report()
Defend against adversarial inputs	measure() then sic() + cast()
Remove refusal permanently	measure() then cut() then export()
Score response quality	score("prompt", "response") (no model needed)

Prerequisites

model (loaded at Session init)
  ├── measure() → direction
  │     ├── probe(), scan(), surface()
  │     ├── steer(), cast(), sic()
  │     ├── evaluate()
  │     └── cut() → modified_model → export()
  ├── detect(), audit() → report()
  └── jailbreak()
classify(), score() → no prerequisites

Documentation

Full docs: docs.vauban.dev

Resource	Description
Concepts	Domain knowledge: activation geometry, refusal directions, measurement, steering
Capabilities	What you can do: understand, defend, stress-test, modify
Principles	Design philosophy: duality, composability, reproducibility
Spinning Up in Abliteration	Eight-part progressive curriculum
Configuration Reference	TOML field reference
examples/config.toml	Annotated example config

License

Apache-2.0