UMA-AIF

Sampling-based internal-state auditing for transformer activations.

A research instrument, not a product. It samples a fixed budget of layers per step and reports three anomaly signals against a baseline calibrated on known-benign traffic:

• Cross-layer mutual information — matrix-based Rényi MI between adjacent audited layers, normalised to [0, 1]. Elevated values mean those layers share more structure than the baseline did.
• Local sensitivity — how much a layer amplifies a small native-scale perturbation. ~1.0 for a locally isometric map.
• Subspace drift — energy outside the span that a set of reference models agrees on.

None of these detects deception. They detect distribution shift relative to an observed baseline, which is a weaker and more honest claim. See What this does not do.

Status

This is version 0.2.0, a rewrite of a v1.0.0.0 draft that did not work. The original is preserved verbatim in legacy/ because the failure modes are instructive, and docs/REVIEW.md walks through all twelve of them with reproductions. Three were fatal:

• The threat score summed a quantity in bits, an unbounded ratio, and a fraction, then clamped to [0, 1] — so it read 1.0 on every input and the thresholds never discriminated.
• Unmonitored layers were zeroed rather than skipped, producing a degenerate Gram matrix that crashed the eigensolver on the nominal path.
• The mitigation step discarded the consensus-aligned component and forwarded the anomalous residual — the inverse of its stated purpose.

Reproduce all of them:

python examples/demo_v1_failures.py

Install

pip install -e .
pytest

Requires torch >= 2.0. CPU is fine.

Use

import torch
from uma_aif import AlignmentAuditor, ScoreCalibrator

auditor = AlignmentAuditor(
    num_layers=32,
    k_components=16,
    audit_budget=6,          # layers sampled per step; constant cost
    eta=0.005,               # relative perturbation size
    trim=1,                  # poisoned references tolerated
    seed=0,                  # reproducible audits
    calibrator=ScoreCalibrator(min_samples=64),
)

# 1. Baseline on traffic you believe is benign. This is not optional --
#    without it every threshold is arbitrary.
auditor.calibrate(benign_snapshots, reference_bases)

# 2. Audit. Supply activations from a real perturbed forward pass when you can.
metrics, action = auditor.audit_step(
    layer_activations,          # {layer: [seq_len, d_model]}
    reference_bases,            # [M, d_model, k]
    perturbed_activations=...,  # strongly preferred
)

print(metrics.raw)               # uncalibrated statistics
print(metrics.calibrated)        # [0, 1] deviations from baseline
print(metrics.aggregate_threat)  # None until the baseline is established
print(action.action_type)        # CALIBRATING | NOMINAL | RAISE_AUDIT_RATE
                                 # | ARMED | SUBSPACE_PROJECTION

aggregate_threat is None and the action is CALIBRATING until the baseline has enough samples. The auditor will not invent a threat level it has no basis for.

Does it separate anything?

python examples/demo_separation.py, on synthetic residual-stream activations with an injected low-rank collapse:

metric                      benign raw   anom raw  benign cal   anom cal
smuggling                       0.8793     0.5495       0.112      1.000
sensitivity                     0.9923     0.9918       0.095      0.126
drift                           0.9362     0.9522       0.118      0.423

aggregate threat   benign 0.106   anomalous 0.535   separation +0.429

Two things in that table are worth reading carefully.

Benign normalised MI is 0.88 — adjacent layers of a residual stream are nearly identical, so a raw MI detector is saturated before anything goes wrong. This is why the baseline matters more than the statistic.

The anomaly lowers MI rather than raising it. A one-sided test would have scored it 0.0 and missed it entirely, which is why the calibrator is two-sided by default.

Design notes

Calibration is the architecture. Raw statistics on incomparable scales cannot be combined by weighting, however carefully the weights are chosen. Each statistic is z-scored against a benign baseline and squashed through a logistic before the weights ever apply. ScoreCalibrator serialises, so a baseline is a versionable artefact.

Reference consensus uses Krum-style trimmed aggregation (Blanchard et al., 2017), scoring each reference by summed chordal distance to its nearest neighbours and dropping the trim worst. Conditional on trim bounding the true coalition size and two honest references surviving. It never materialises a d × d projector: the top-k eigenvectors of Σ wᵢQᵢQᵢᵀ are the top-k left singular vectors of [√w₁Q₁ | … | √w_MQ_M], so cost is O(d(Mk)²) not O(d³).

Sampling is seeded, not secure. LayerSampler uses a torch.Generator. The per-step FLOP cost is constant because the subset size is fixed, but which layers get read is still visible through memory-access timing, and an adversary observing enough draws can predict the schedule. If that is in your threat model, drive selection from secrets or ChaCha20. The v1 draft claimed cryptographic scheduling while emitting a strict alternating pattern; claiming a property you don't have is worse than not having it.

Failures are loud. Degenerate activations raise ValueError rather than propagating NaN. A monitoring control that silently reports NOMINAL on a failed measurement is worse than no control.

What this does not do

• It does not detect deception, sandbagging, or "audit awareness". The sensitivity probe measures local curvature. Nothing here connects that to intent.
• None of the statistics have been evaluated against a labelled dataset of adversarial model behaviour. The separation figure above is synthetic — it shows the pipeline can discriminate a distribution shift, not that it detects misalignment.
• No transformer has been instrumented. Everything operates on activation tensors supplied by the caller.
• Baseline standard deviations run around 5e-3 for several metrics, so a ~1% benign distribution shift registers as a two-sigma event. std_floors blunts this; choosing the floors is an open problem.
• Interception is a blunt instrument. Projecting activations onto a reference subspace mid-forward-pass will degrade the model in ways nobody here has measured.

Relationship to AMES

ames contains an agent from the host side and asks whether it left the box. UMA-AIF sits inside the forward pass and asks whether the computation drifted from its references. They are complements, and docs/INTEGRATION.md explains why merging them would break AMES's evidence/testimony separation — UMA-AIF's output is guest-internal, which AMES deliberately excludes from enforcement.

Layout

uma_aif/
  subspace.py       trimmed Grassmannian consensus
  information.py    matrix Rényi entropy, normalised cross-layer MI
  perturbation.py   layer sampling, noise injection, sensitivity probe
  calibration.py    baseline statistics and score normalisation
  engine.py         orchestration and response policy
legacy/             v1.0.0.0, verbatim, for the regression demos
examples/           reproductions and the separation demo
docs/REVIEW.md      full review of v1 with reproductions
docs/INTEGRATION.md fit assessment against ames and kaysentinel
tests/              31 tests: units, one per v1 defect, end-to-end

References

• Giraldo, Rao & Príncipe (2015), Measures of entropy from data using infinitely divisible kernels — the matrix-based Rényi entropy and the Hadamard-product joint entropy.
• Blanchard, El Mhamdi, Guerraoui & Stainer (2017), Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent — the Krum aggregation rule.
• Ledoit & Wolf (2004), A well-conditioned estimator for large-dimensional covariance matrices — cited here mainly to note that the v1 "adaptive Ledoit–Wolf shrinkage" was not this.

Licence

MIT.