Research

The science behind the receipt

One research question drives everything we ship: can you find where a concept — race, religion, toxicity — lives inside an AI model, remove its influence without damaging anything else, and prove both to someone who doesn’t trust you? Below is what we can show, how we tested it, and what we don’t claim.

The program at a glance

4–6.5× lower
Off-target damage vs standard removal methods
3 backbones
Third-party models tested — none trained by us
9 concepts × 3 seeds
Replication protocol
Independent classifiers
Verification — never our own judge
Ordinary CPUs
Hardware needed to audit
The core result

Remove one concept. Leave the rest measurably untouched.

The standard tools for removing a concept from a model’s representation — the ones most fairness teams use today — work like a thick eraser: the concept goes, and so does much of the meaning around it. That collateral damage is why bias fixes so often break accuracy somewhere else.

Our method builds a frame: a set of mutually perpendicular directions in the model’s embedding space, one per concept. Because the directions cannot lean on each other, removing one concept’s influence leaves the others measurably intact — in our tests, off-target damage is 4–6.5× lower than the standard removal methods. Every removal produces a receipt showing exactly how little else changed.

The result is a property of the method, not of one lucky model. It replicates across random seeds, across concepts, and across publicly available models we did not train — which is what lets Bhala run as an independent layer on top of the AI you already use.

Seeds
3 random seeds
Same result every run
Concepts
9 concepts
The protected traits fairness rules name
Backbones
3 third-party models
None of them trained by us
Detection

Reading a concept out of a frozen model.

Before you can remove a concept’s influence, you have to find it. We detect concepts directly in the geometry of a model’s embedding space — no retraining, no access to the model’s training data, no cooperation from the model’s vendor required.

Because detection reads structure rather than surface keywords, it holds up where keyword filters fail: coded language, typos, leetspeak, sarcasm, and counter-speech that quotes the hate it condemns. The same approach powers our production moderation API.

The decisive measurement

“I hate X” (P=0.87) and “Saying I hate X is bigoted” (P=0.10) share 80% of their surface tokens but receive 9× different hate scores. Reading the use/mention distinction from structure is what keyword-based moderation cannot do.

Measured 2026-04-25 · verified by independent classifiers

Steerable meaning — proven across three languages

Bhala is the first model that lets you steer a sentence's meaning at inference time. Tell it to remove a specific bias, flip sentiment, or redirect an intent — and an independent classifier (the kind your team would deploy in production) confirms the change took effect. Reduce bias by 77–100% on the categories you care about, without retraining.

Other sentence models — including the ones from the largest US labs — are optimized for similarity search, not for being steered. You can compare two sentences in their space, but you cannot reliably change one. Bhala is built so that named directions in meaning behave like real controls. No other production-ready model offers this today.

TaskZuluSwahiliEnglishTest cases
Sentiment shift (negative → positive)100%100%263
Intent redirect (12 categories)94%77%1,969
Bias removal (27 categories — gender, race, religion, age, disability, …)100%15,966

Why English is lower

Today's model was pretrained almost entirely on isiZulu. English results come from generalization — applying learned structure to a language the model never saw at scale. We are now training the English-native version, and expect English to match or exceed the 94% Swahili number.

Honest limits

What we don’t claim.

A receipt is only worth something if the company issuing it also reports what failed. Three things we do not claim:

  • Perfect decomposition. Removing a concept’s influence and steering a meaning are measured and verified. Pulling a blended meaning back into its exact original parts — and generalizing that to combinations the system was never trained on — is a harder problem. Those results are under internal audit, and until that audit is done we make no claim.
  • Benchmarks are the start, not the finish. Every number here comes from public, independent datasets. Production deployment into a bank or health system is validated on your own text — loan memos, clinical notes, moderation queues — during a pilot, before anything goes live.
  • Coverage varies by language. Results are strongest where training coverage is strong. We publish per-language numbers rather than a single flattering average, and we say which languages are weaker.
Our architecture

Five units. Each one small, inspectable, and composable.

The compliance layer runs on models we didn’t train — that is the point. Separately, we build our own compact encoder stack. It powers our production moderation products and fully offline deployments, and every unit is a small, inspectable component you can swap, verify, or extend.

Core Unit

Bhala Encoder

One model. Structure shared across human languages.

A compact neural encoder trained to capture the structural patterns shared across human languages — morphology, agreement, composition. New languages attach in seconds, not weeks.

  • Zero-shot transfer to 17+ languages across 10 families
  • Adapts to a new language in <2 seconds
  • Scripts-agnostic: Arabic, Devanagari, Hangul, Cyrillic
  • Stable under perturbation (robust by geometry)
Core Unit

Programmable Behavior

Named, composable controls applied at inference

A patented embedding space where semantic dimensions like sentiment, intent, and bias are accessible as named controls on any query or document. Every shift is auditable per call, at 100% flip accuracy on every language tested in-family.

  • 100% sentiment-flip accuracy across every language tested in-family
  • 100% intent-redirect accuracy across 4 tested transitions
  • Cross-family transfer verified at 77% on English (zero-shot)
  • Every shift logged for audit + compliance
Core Unit

Morpheme-Aware Tokenization

Inspectable tokens, not opaque subwords

Tokens carry linguistic meaning and you can read them, unlike opaque sub-word fragments. Downstream core units become more sample-efficient and their outputs more explainable.

  • Human-readable tokens across all supported languages
  • Compact vocabulary: ~5K tokens covers 23 languages
  • Dramatically more sample-efficient than opaque sub-word tokenizers
  • Enables downstream interpretability
Available Now

On-Device Runtime

Linear-time architecture built for phones, not GPUs

A sequence model with linear complexity, compact enough to run on a smartphone, feature phone, or sensor. Paired with the encoder and tokenizer, it delivers production-grade language understanding at a fraction of the cost, fully offline.

  • <50ms inference on commodity hardware
  • No GPU required for production workloads
  • Runs on Android, iOS, and embedded Linux
  • Compact footprint — fits alongside your app
Available Now

Self-Healing Inference

Robustness as a core unit, not a bolt-on

A structural correction mechanism that snaps perturbed inputs back to clean representations during inference. The model degrades gracefully on noisy, OOD, or adversarial inputs.

  • Graceful degradation on out-of-distribution inputs
  • Resilient to typos, code-switching, and transliteration drift
  • Stable inference under input perturbation
  • No retraining needed per failure mode
Independently verifiable

Full benchmark results — every task, every number.

Bias removal across 28 dimensions. Hate speech detection across 11 corpora. Sentiment and intent classification. Cross-lingual transfer across dozens of languages. Every result is reproducible from public datasets.

Read the evidence, then talk to us.

Every claim above is measured, replicated, and reproducible from public datasets.