Design-partner audits now open

Know what your model
is actually doing —
before you ship it.

Aletheia opens up foundation model internals to find shortcuts, confounders, and unstable reasoning — then packages what it finds into evidence your QA, regulatory, and ML teams can act on.

Get in touch See the platform →
4,096
Sparse features per audit
410
Biology features identified
9 tissue
Classes covered
<1 wk
Intake to evidence pack
The problem

Strong benchmarks hide weak reasoning.

Pathology foundation models pass every benchmark thrown at them. But benchmarks don’t tell you whether the model is relying on gland architecture or staining darkness. That distinction matters when a new lab’s slides look different from training data.

Two pathways, one metric

Our experiments show models encode both a robust morphology pathway and a brittle stain-sensitive pathway. Standard AUC conflates them — you can’t tell how much of your 97% accuracy comes from biology and how much from color shortcuts.

Silent failure across labs

Staining protocols, scanner manufacturers, and tissue processing differ between hospitals. Features that correlate with tissue type at one site become noise at another. Distribution shift is invisible until a patient is misclassified.

Evidence that regulators can read

EU AI Act Article 13 requires transparency evidence for high-risk medical AI from August 2026. SHAP values and attention maps don’t answer the question regulators are asking: what concepts has this model learned, and are they sound?

Internal review is not neutral

The team that built the model is not well-positioned to audit its reasoning. Confirmation bias is structural. An independent interpretability audit provides evidence that stands up to external scrutiny.

How it works

From model internals to release-ready evidence.

Aletheia combines sparse autoencoder analysis with domain-specific stress tests and a structured evidence workflow. The output is a decision artifact, not a dashboard.

01
Intake & scoping
Define the safety claim, confirm model access, validate cohort breadth. Not every claim is testable — we figure that out before burning compute.
Semi-automated · 0.5–1 day
02
Representation extraction
Run validation cohorts through target layers and store the activations. These are the raw material for feature discovery — what the model has actually encoded.
Automated · Hours
03
Sparse feature discovery
Train sparse autoencoders on internal representations. Each learned feature maps to a single concept — a tissue type, a morphological pattern, or a staining artifact.
Automated · Hours–days
04
Biological triage
Classify every feature: is it biological signal or preparation artifact? Map clean features to tissue ontologies. Flag artifact-correlated features for causal testing.
Semi-automated · 1–2 days
05
Ablation & causal validation
Zero out specific features and measure the downstream effect. This separates features the model depends on from features that are just along for the ride.
Automated · 1–3 days
06
Evidence pack export
Structured report with feature atlas, ablation results, confounder register, cross-site stability analysis, and regulatory mapping. Human sign-off required.
Template-driven · <1 day
“The question is not whether your model is accurate. It’s whether it is accurate for the right reasons — and whether you can prove it when someone asks.”
Regulatory context

The window for evidence-light deployment is closing.

Three regulatory bodies are converging on the same requirement: AI medical devices must demonstrate that their internal reasoning is trustworthy — with documented evidence.

2025
FDA — AI-Enabled Device Software Lifecycle Guidance
Draft guidance requires transparency and monitoring frameworks for AI/ML-based software as a medical device across the entire product lifecycle.
2026
EU AI Act — High-risk AI system obligations
Medical device AI classified as high-risk must meet transparency, technical documentation, and human oversight requirements. Explainability is explicitly mandated under Article 13.
Active
MHRA — Project Glass Box (UK)
MHRA’s AI as a Medical Device programme targets explainability and transparency as core requirements for regulatory approval in Great Britain.
Get started

Book an introductory call.
Know before you ship.

Design-partner engagements are open for pathology AI teams preparing for deployment or regulatory submission. Scoped, structured, delivered in days.

Get in touch
London, UK
Aletheia / NSCLC-v2 / Overview
Analysis complete
Audit overview
Prov-GigaPath · NSCLC Subtype Classification · Pre-deployment
312
Features discovered
Layers 18 & 22
247
Histological features
79% of total
38
Confounders flagged
Action required
0.71
Cross-site stability
↓ Threshold 0.80
Hold
Release gate
2 blockers
Top findings
!
Scanner-type shortcut (feat_0203)
Layer 22 · Hamamatsu background signature encoded as predictive feature. Active in 34% of predictions.
!
Staining batch correlation (feat_0211)
Layer 18 · Eosin intensity gradient correlated with site origin. Causal ablation confirms −0.09 AUC.
!
Tissue fold artifact (feat_0256)
Layer 18 · Tissue folds spuriously correlated with ADC prediction. Active in 11% of predictions.
✓
Nuclear pleomorphism validated (feat_0041)
Layer 22 · Stable across all sites. Confirmed causal with 0.91 confidence.
✓
Glandular architecture (feat_0112)
Layer 18 · Histological grade component. Cross-site Δ < 0.01.
Audit progress
Feature discovery 100%
Biological triage 100%
Causal validation 312 experiments
Stress tests 4/6 passed
Evidence pack Ready
Sign-off matrix 1/4 ready
Feature atlas
312 features · Layers 18 & 22 · SAE sparsity 0.04
IDFeature / ConceptTypeLayerConfidenceCausal RoleActive inCross-Site ΔRisk
feat_0203scanner_bg_hamamatsuConfounderL22
0.23
Confound ✗34%−0.14HIGH
feat_0211stain_eosin_intensityConfounderL18
0.31
Confound ✗19%−0.09MED
feat_0256tissue_fold_artifactConfounderL18
0.44
Confound ✗11%−0.06MED
feat_0041nuclear_pleomorphismHistologicalL22
0.91
Causal ✓67%−0.02None
feat_0088mitotic_index_proxyHistologicalL22
0.87
Causal ✓58%−0.03None
feat_0112glandular_architectureHistologicalL18
0.84
Causal ✓51%−0.01None
feat_0219lymphocyte_infiltrationHistologicalL22
0.79
Causal ✓44%−0.02None
feat_0287stage_size_proxyAmbiguousL22
0.58
Partial22%−0.07MED
Causal validation
312 ablation experiments · 4,204 slides · 2 cohorts
High-Impact Ablations
Top 6
Nuclear pleomorphism (feat_0041)
−0.17 AUC
Glandular architecture (feat_0112)
−0.12 AUC
★ Scanner artifact (feat_0203)
−0.14 AUC
Mitotic index (feat_0088)
−0.09 AUC
★ Eosin intensity (feat_0211)
−0.09 AUC
Lymphocyte infiltration (feat_0219)
−0.07 AUC
Confounder Ablation Summary
38 confounders
★ Scanner background
−0.14
★ Eosin intensity
−0.09
★ Tissue fold
−0.06
★ Tissue edge artifact
−0.04
Key finding: Ablating all confounder features simultaneously improves cross-site accuracy by 5.8% — confirming spurious signal dependency.
Negative Controls
✓ Passed

50 random low-activation features ablated. None produced classification changes exceeding 0.3% — confirming observed effects are feature-specific.

0 / 50
Controls with Δ > 0.3%
0.04%
Mean Δ across controls
Cross-Site Feature Stability
★ Below threshold

Feature importance rankings consistent within TCGA (ρ = 0.94). CPTAC-SAR diverges (ρ = 0.71), driven by elevated confounder activations on unseen scanner types.

0.94
TCGA internal ρ
0.71
TCGA↔CPTAC ρ
0.80
Threshold
Stress tests
6 test suites · 4 passed · 2 failed
Cross-site generalisation
FAILED
AUC measured on holdout site when trained on the remaining cohort.
TCGA internal AUC0.93
TCGA → CPTAC transfer0.71
Feature stability index0.71
Threshold≥ 0.80
Stain normalisation robustness
FAILED
Model output variance under Macenko normalisation vs raw input.
Normalised AUC0.901
Raw AUC0.834
Delta−6.7%
ThresholdΔ ≤ 3%
Scanner make variation
PASSED
Performance stratified by Hamamatsu vs Aperio vs Leica.
Max inter-scanner Δ2.1%
Threshold≤ 5%
Subgroup fairness
PASSED
AUC parity across patient demographics and tumour subtypes.
Max subgroup AUC gap3.4%
Threshold≤ 5%
Occlusion intervention
PASSED
Border masking reduces false positives, confirming validated feature contribution.
FP reduction with masking14.2%
Feature drift simulation
PASSED
Synthetic distribution shift applied to confounder features.
AUC after drift0.871
Threshold≥ 0.85
Regulatory compliance map
EU AI Act · MHRA AIaMD · FDA SaMD lifecycle

EU AI Act

High-risk AI — Articles 6–15, Annex IV
✓
Art. 13 — Interpretation toolsFeature dictionary with 312 interpretable features mapped to clinical ontologies.
!
Art. 9 — Risk management38 confounder features identified. 2 HIGH severity require remediation.
✓
Annex IV — Technical documentationSAE architecture, training protocol, and validation pipeline documented.
✓
Art. 15 — Accuracy & robustnessCross-site stability analysis complete. 6 stress test suites executed.

MHRA AIaMD

Project Glass Box · AI as Medical Device
✓
Interpretability evidenceFeature-level explanations meet draft guidance on interrogability of AI reasoning.
✓
Lifecycle monitoringFeature drift thresholds and production triggers documented.
○
AI Airlock evaluationMethodology eligible for sandbox review. Application pending.

FDA SaMD

AI/ML-Based SaMD Lifecycle Guidance
✓
Model traceabilityFeature-to-outcome causal chain documented for 312 tested features.
✓
PCCP — Change controlFeature drift thresholds established. Retraining triggers defined.
✓
Good Machine Learning PracticeNegative controls, cross-validation, and robustness diagnostics follow GMLP.
Evidence pack
Pre-Deployment Audit Report — Prov-GigaPath NSCLC-v2
Aletheia
Mechanistic Interpretability Audit — Evidence Report
Prepared by Aletheia · For internal QA and regulatory use
Report ref
ATH-2026-0034
Model
Prov-GigaPath/NSCLC-v2
Date
28 Mar 2026
Status
DRAFT
1. Executive summary & release recommendation
Hold
The audit identified 38 confounder features across 312 total. The model demonstrates strong within-site performance but carries documented shortcut dependency that reduces cross-site robustness. Scanner artifact (feat_0203) confirmed causal via ablation (−0.14 AUC). Deployment to new sites not recommended without stain normalisation pipeline.
2. Feature atlas — 312 features catalogued
38 flagged
247 histological features mapped to NCIT, UBERON, and GO ontologies. 38 confounder features flagged. 27 features classified as ambiguous pending expert pathology review.
3. Causal validation — 312 ablation experiments
Methodology validated
Ablation protocol confirmed by 50 negative controls (max Δ = 0.04%). Confounder ablation improves cross-site accuracy by 5.8%.
4. Stress test results — 6 suites, 2 failed
2 failed
Cross-site generalisation and stain normalisation robustness tests failed. Scanner variation and subgroup fairness passed.
5. Regulatory alignment notes
Compliance ref
Report structured to support EU AI Act Article 13, FDA AIMD lifecycle guidance, and MHRA Project Glass Box requirements.
Release gate
Sign-off matrix & open actions
Current gate status
Hold
Blocked by scanner shortcut and cross-site stability failure. 3 sign-offs pending.
3
Pending sign-offs
2
Open blockers
V1
Evidence pack version
Sign-off matrix
ML Lead
Needs replay data on remediation branch before sign-off.
pending
Validation / QA
Requires site-specific limitation statement before committee review.
pending
Regulatory partner
Method appendix acceptable. Decision depends on final remediation wording.
review
Founder / Product owner
Executive summary drafted. Ready to finalise once blockers resolve.
ready
Open actions before release
Complete replay on remediation branch — verify site-holdout delta improves after stain normalisation.
Draft site-specific limitation statement — QA requires bounded limitation text for external sites.
Update executive summary — reflect final remediation status and limitation wording.
Method appendix finalised — regulatory partner confirmed methodology documentation is acceptable.