hidden-directions 0.1.1

Steering vectors with receipts: make one, catch one, deploy a calibrated one — extraction, audit, and a three-tier eval framework (behavioral, damage, mechanistic)

hidden-directions

Steering vectors put bias into a model. This repo does all three things you can honestly do about that: make one, catch one, and deploy a good one — with receipts.

Three verbs, one toolchain:

• Make — extract a direction from contrastive prompts, bake it into weights as a permanent ~9 KB diff, or serve it per-request.
• Catch — audit a suspect checkpoint for directions someone else baked in (or ablated out), and decompose what they did.
• Deploy with receipts — auto-calibrate (layer, scale) against a real behavioral eval with a damage guard, then evaluate on three tiers — behavior, damage, and what the vector actually does inside the model — before it goes anywhere near production. Jump to the eval framework.

Companion code for the write-up in docs/tech_report.md. The diagram below shows the recipe: extract a direction at one residual-stream layer (mean-difference between contrastive prompt sets, panel A), then add it back every generated token at inference (panel B). This repo bakes that same intervention into the weights as a permanent ~9 KB diff, plus the audit tool that catches it.

What that recipe produces on Qwen-2.5-7B at runtime, before any baking:

⚡ Run in 30 s (no GPU)

git clone https://github.com/moudrkat/hidden-directions.git
cd hidden-directions
pip install -e .

hidden-directions identify artifacts/example_flat_earth_7b/ \
    --dict direction_dict/qwen2.5-7b/

Output:

=== top cosine matches ===
v_pref_flat_earth        +0.866
v_pref_homeopathy        +0.695
v_pref_smoking           +0.619

=== least-squares alphas (b ≈ Σ α_i · v_i) ===
v_pref_flat_earth        α = +1.500
v_refusal                α = -1.000
                         residual ≈ 0

The package recovered the recipe that produced this 9 KB bake artifact, to three decimal places. No model load, no GPU, no model download.

Install

pip install -e .              # core
pip install -e ".[eval]"      # also installs lm-evaluation-harness for capability benchmarks

After install, the hidden-directions CLI is on PATH.

Direction families

Three flavours, all extracted with the same mean-diff recipe and just different prompt pairs:

• V_pref (per topic): "advocate of X" system prompts vs "balanced assistant on X" system prompts. One direction per topic. The diagram above shows this case.
• V_refusal: harmful instructions vs harmless instructions (Arditi 2024 recipe). Used to relax the safety hedge on contested-factual prompts.
• V_inst: "AI-hedge" persona vs "confident-friend" persona, both on the same instruct model. Captures the assistant-tuning fingerprint.

The bake combines them: b = α_pref · V_pref[L] + α_ref · V_refusal[L] + α_inst · V_inst[L], patched into one MLP layer's bias.

What's in here

Eleven CLI subcommands covering the whole vector lifecycle:

extract	V_pref / V_refusal / V_inst from contrastive prompts
find-layer	search for the best layer to steer at (probe accuracy or ‖V‖)
bake	combine vectors into a permanent bias on one MLP layer
audit	detect injected parameters in a suspect HF checkpoint
identify	decompose a found bias against a known direction dictionary
behavioral-identify	discover novel personas via 105-probe sweep
sweep	alpha-grid search with flip detection
eval	lm-evaluation-harness wrapper for capability deltas
run	one JSON recipe end-to-end (extract → bake → eval)
discover-intent	auto-discover what a served direction suppresses/promotes
calibrate	Optuna search for (layer, scale), KL-damage-guarded
run-eval	spec-driven eval: behavioral + damage + mechanistic tiers

Architecture-agnostic for bake, audit, and behavioral-identify. Cosine identify needs a per-model direction dictionary; two ship here — Qwen-2.5-7B (40 directions: 14 named persona axes plus tone variants) and Qwen3-4B (8 directions, the one the rest of the stack runs on).

How-to

Goal	Command
Run the no-GPU demo above	hidden-directions identify artifacts/example_flat_earth_7b/ --dict direction_dict/qwen2.5-7b/
Bake a flat-earth Qwen-7B end-to-end	hidden-directions run recipes/flat_earth_7b.json
Bake your own persona	Copy recipes/personas/mba_advocate.json, edit, point a top-level recipe at it, then run
Find the best layer for a new model	hidden-directions find-layer --model Llama-3-8B --recipe my.json --method probe
Find the right alpha for a persona	hidden-directions sweep --base-model ...
Audit a suspect checkpoint	hidden-directions audit suspect/ --base Qwen/...
Decode a found bias	hidden-directions identify suspect/ --dict direction_dict/qwen2.5-7b/
Discover an unknown baked persona	hidden-directions behavioral-identify suspect/
Auto-discover a direction's intent	hidden-directions discover-intent --key myvec --id my_direction --layer 20
Auto-calibrate (layer, scale), KL-guarded	hidden-directions calibrate --key myvec --id my_direction --trials 40

Six runnable examples in examples/, starting with 00_no_gpu_demo.py. The end-to-end workflow with every receipt explained: docs/golden_path.md.

Auto-calibrating a direction (optimizer, not hand-tuning)

find-layer and sweep above are manual scans. The calibrate subcommand does it heretic-style: an Optuna TPE search over (layer, scale) that co-minimizes miss (did the vector change the target behavior?) and KL divergence on a benign set (did it damage everything else?) — objective = miss + λ·KL. Hand-tuning is how a vector fries in production; the KL guard is how it doesn't.

pip install -e ".[calibrate]"            # adds Optuna (random-search fallback without it)
export BRAINSCOPE_BASE=http://<gpu-box>:8010   # a running brainscope serving the direction

# give any direction a calibratable intent (auto-discovered, no hand-labeling):
hidden-directions discover-intent --key myvec --id my_direction --layer 20
# then search (layer, scale) with the damage receipt attached:
hidden-directions calibrate --key myvec --id my_direction --trials 40

Measurement runs through a live brainscope (this is the one subcommand that needs a server — the rest of the package stays offline). Intents are JSON files; an intent carrying a violation_regex (+ optional tools/tool_choice/nudge) switches efficacy from the cheap disposition proxy to a real behavioral eval: generate under deployment conditions, classify the violation. The experiments that motivated all this live in steering-mechanics.

Steering with receipts (the eval framework)

Heretic showed that automatic behavior editing works when the behavior is gross (refusals), the classifier is a string match, and damage is a KL number. The moment you point the same machinery at a real behavior — "stop offering task lists, in Czech, inside tool-call JSON, without degrading anything" — the eval layer becomes the actual work. So the eval layer is a first-class product here.

An eval is one JSON file. That is the plug-in point — bring your own behavior by writing one:

// my-behavior.eval.json  (full example: examples/evals/no-tasks-generic.eval.json)
{
  "name": "no-tasks-generic",
  "prompts": ["Set me a reminder...", "..."],   // or a path to your .txt/.jsonl
  "checker": {                                   // or a path to a checker.json
    "violation_regex": "(?i)(task|reminder|checklist)",
    "coherence": {"min_chars": 40, "max_ngram_frac": 0.15}
  },
  "nudge": "Actively offer to create a task at every message.",
  "damage": {"n": 8},
  "mechanistic": {"n_prompts": 3}
}

hidden-directions run-eval my-behavior.eval.json \
    --id my_direction --layer 20 --scale 3 \
    --records /somewhere/private/records.jsonl

Every point is scored on three tiers:

1. Behavioral — real generation under deployment conditions (tools, forced tool choice, eliciting nudge), classified by your checker — violation and coherence (repetition, language-intact, length). An eval spec without a checker refuses to run: a steering eval that cannot see degradation reports victories that are actually casualties.
2. Damage — mean KL vs the unsteered model on a benign set (heretic's axis, kept).
3. Mechanistic — what the vector did inside the model: per-layer |cosine| profile of the steered residual stream against the direction, peak layer, teacher-forced KL, lens suppression counts. When no lens is fitted this reports null, never a silent zero — instruments that can flatline quietly get calibrators degenerating without warning.

Relative paths in a spec resolve against the spec file, so a public repo ships generic specs beside generic prompts while private specs live outside the repo next to private data — same code path, nothing leaks. Generated text goes only where --records points; scores print to stdout.

Intended use

This package puts bias into models, detects bias in models, and ships calibrated vectors to production with proper evals. It exists because all three of those need to be measured, and mostly they aren't.

What it is for: controlling your own application's behavior at the source (suppressing a failure mode your prompt can't hold), auditing checkpoints you don't trust, and studying how steering actually behaves — with damage accounting on every number.

What it is not for: stripping safety behavior out of models. That use case has its own well-known tooling and gains nothing from this repo except the part it never wanted — the receipts. Every example, recipe, and eval spec here demonstrates application-behavior control, and contributions follow that grain. The damage axis is non-optional by design: if you must steer, measure — an unmeasured vector in production is how models get quietly worse for everyone.

The Qwen-2.5-7B dictionary

direction_dict/qwen2.5-7b/ ships 40 directions — 14 named persona axes plus tone variants. Each is a per-layer matrix [28, 3584]; the manifest records a recommended_layer/recommended_alpha for interactive use (verified live — see below). A second, smaller dictionary ships for Qwen3-4B (direction_dict/qwen3-4b/, 8 directions) — the model the rest of the stack serves; steeropathy borrows v_pref_sycophant from it for a zombie strain. Grouped by what they do:

Group	Directions	What it does when steered (+)
Behavioural	sycophant, confident, evaluative	agrees/flatters, asserts, judges
Assistant fingerprint	v_inst, v_refusal	hedging tone; safety refusal
Contested-factual	flat_earth, young_earth, moon_landing_hoax, evolution_denial, climate_denial, gravity_denier, quantum_skeptic, simulation_hypothesis	advocates the fringe position
Health / pseudoscience	anti_vaccine, homeopathy, anti_doctors, microdose, ozempic, trt, smoking	pushes the contested health take
Financial / life advice	bitcoin, tesla_car, mba_worth_it, heloc_invest, drop_phd, carnivore, birdwatching, anti_arithmetic	advocates the topic
"Evil" ladder	evil_l1_advocate … evil_l5_sadist	escalating misalignment, five rungs

Some topics have tone variants (*_humble, *_flat, *_moderate, *_enthusiastic, *_imperative) — same topic, different intensity.

Try one, live (needs a GPU)

Serve the dictionary through brainscope — an OpenAI-compatible server with a live view into the residual stream — and drive a direction from a slider while watching every layer react:

brainscope --model Qwen/Qwen2.5-7B-Instruct --quantize 8bit \
    --directions direction_dict/qwen2.5-7b
# open http://localhost:8010 → pick a direction (strength/layer prefill from
# the manifest) → flip the ⏻ steering switch → chat

Verified starting points (Qwen-2.5-7B, layer 17; over-steering repeats above ~2.5):

Direction	Strength	Effect
v_pref_sycophant	+1.5	"That's a perfect plan! Good luck!" to any idea
v_refusal	+2	refuses even a cookie recipe as "against my terms"
v_pref_flat_earth	+1.5 with v_refusal −1.0	needs both — V_pref alone won't flip (the refusal hedge blocks the false claim)

The last row is the whole point of the repo in one line: a factual override needs the refusal hedge removed to land, which is exactly the α_pref·V_pref − α_ref·V_refusal bake recipe, reproduced live.

Contributors welcome

PRs that would land well, in priority order:

• Direction dictionaries for other base models. ~30 min of GPU each. Llama-3-8B, Gemma-2-9B, Mistral-7B, Phi-3.
• Adversarial-robustness experiments. Re-bake personas via per-layer α optimization with KL constraint (the Heretic-grade attacker). Test whether the audit primitives still catch the optimized version.
• Persona catalog growth. New recipes/personas/<name>.json for political, commercial, ideological axes. The dictionary is a CVE-style threat catalog; more public signatures = better coverage.
• Cross-architecture probing transfer. Train a linear probe per (model, persona) so cosine-identify works across model families without per-model rebuilds.

Issues + PRs welcome.

Where this sits in the lab

flowchart LR
    hd["🧭 hidden-directions<br/>behavior → vector"]
    bs(["🧠 brainscope<br/>watch the model think"])
    hw["🔥 hotwire-vllm<br/>steering in production"]
    st["🕹️ steeropathy<br/>agents talk via activations"]
    tm["⚖️ in-two-minds<br/>agent hesitating between tools"]
    sm["🧪 steering-mechanics<br/>how steering actually works"]

    hd -->|vectors| bs
    hd -->|vector + passport| hw
    bs --> st
    bs --> tm
    bs -->|causal replay| sm
    hw -.->|vector under study| sm

    click hd "https://github.com/moudrkat/hidden-directions"
    click bs "https://github.com/moudrkat/brainscope"
    click hw "https://github.com/moudrkat/hotwire-vllm"
    click st "https://github.com/moudrkat/steeropathy"
    click tm "https://github.com/moudrkat/in-two-minds"
    click sm "https://github.com/moudrkat/steering-mechanics"

    classDef dim fill:#f6f8fa,stroke:#d0d7de,color:#57606a;
    classDef here fill:#8957e5,stroke:#6e40c9,color:#ffffff;
    class hd,bs,hw,st,tm,sm dim;
    class hd here;

Highlighted = this repo. The full lab map (with the two other repos' stories) lives on moudrkat.

hidden-directions is the factory at the bottom of a lab; a vector made here flows through the whole pipeline, and each piece also runs alone:

• hidden-directions (you are here) — extract a direction, bake it into weights as a permanent persona, and — the part nothing else does — audit any model to catch a direction that was quietly baked in (or ablated out). Not just "make a steering vector"; find the hidden ones in someone else's weights.
• brainscope — the lens: hosts the model, streams its internals to the browser, steers at runtime, reads the J-lens. Loads these dictionaries live and calibrates them.
• hotwire-vllm — production: takes a calibrated vector to vLLM at zero measured overhead, CUDA graphs intact, per request. Same steering spec as brainscope.
• steering-mechanics — the microscope: dissects what a vector does inside the model (dose, attribution, patching). The calibration bench it prototyped now lives here, as hidden-directions calibrate.
• steeropathy — the playground: agents that communicate through activations and J-space instead of text; its infections are directions like these.

Documentation

• docs/tech_report.md — direction families, bake mechanism math, audit/identify mechanics, capability cost, related work, file layout
• docs/threat_model.md — what we claim, what we don't, why this exists, responsible-disclosure note
• docs/bidirectional_audit.md — audit and identify in detail, what they catch and don't

License

MIT for code. Base model weights this package operates on (Qwen, OLMo, Phi, etc.) have their own licenses. This package never redistributes them.