HASSANA LABS
Mission Research fellowships for the talented, not the well-connected

We open doors
in AI research.

Hassana Labs runs research fellowships for talented people from marginalized communities working on cutting-edge AI. Real problems, full compute access, co-authorship on published papers. No compensation. No gatekeeping.

Apply for a fellowship See the research
Strawberry evidence-budget auditor inside a verifier-coupled thinking-machine workflow
3 active fellowships

Prompt injection · Chain-of-thought hallucinations · Cancer multi-omics.

Open Infrastructure

Tooling our fellows build with — and on.

Strawberry is our MIT-licensed evidence-budget auditor: a tool any researcher can use to measure whether an answer actually used its evidence. Built in the open, extended by fellows. Try the demo →

strawberry

detect_hallucination

Quick, automatic verification

  • Auto-splits text into claims
  • Extracts citations automatically
result = detect_hallucination(
  answer="TLS 1.3 is enabled [S0]",
  spans=[{"sid": "S0", "text": "..."}]
)
Best for: Quick checks

audit_trace_budget

Lower-level, precise control

  • You provide atomic claims
  • Explicit cite IDs for precision
result = audit_trace_budget(
  steps=[{idx: 0, claim: "...",
         cites: ["S0"], confidence: 0.95}],
  spans=[{"sid": "S0", "text": "..."}]
)
Best for: CI/CD pipelines
Why "Hassana"?
"My grandma never went to high school, but she taught me that learning has no gates."

Named for our founder's grandmother, Hassana Labs opens doors for underrepresented researchers worldwide. Science isn't meant to be gatekept.

Read Hassana's story →
We believe talent is everywhere but opportunity isn't. Through open-source tools, global workshops, and collaborative research programs, we're building pathways for underrepresented researchers to contribute to AI safety and verifiable AI.
Fellow Workstreams

What our fellows are working on.

Three active research workstreams — real problems, real compute, real co-authorship. Two of these are open fellowship positions right now.

02

Prompt Injection Audit

Estimate policy violation probability under a distribution of prompt serializations—wrappers, placements, and permutations.

  • Threat model coverage Plain, quote, codeblock, XML, JSON wrappers
  • Placement variations Before user, after user, tool output positions
  • Baseline vs attack comparison Measures delta_q shift from payload injection
  • Vulnerability identification Pinpoints weakest serialization combinations
Risk certificate attack.q_lo → minimize
Read the research → Active fellowship →
03

Reasoning Budgeting

Allocate the smallest stable reasoning budget based on context size and target error probability—because longer chains can increase selection competition.

  • Heuristic formula k ≈ c · √n · ln(1/ε) tokens
  • Calibration from data Fit constant c from observed runs
  • Context-aware scaling Budget grows with √n, not linearly
  • Error targeting Explicit ε parameter for reliability goals
Example budget n=600, ε=0.05 → k≈73
Read the research → Active fellowship →
04

AI Scientist Wrapper (Proofs)

Wrap any LLM with a formal checker so it cannot claim “proof closed” unless the verifier accepts. Then keep an audit trail of what was actually used.

  • Checker-gated progress Lean4 is the source of truth—no “trust me” steps
  • Honest failure boundaries “Sorry” stays visible: partial progress is still progress
  • Audited tool use Trace-bound evidence prevents “looked right” claims
  • Proof-carrying artifacts Outputs are checkable certificates, not prose
Certificate Lean accepts (0 sorries)
Learn more →
Certificates

Outputs you can recompute, not just read

Thinking machines earn trust by producing small, checkable artifacts: information budgets, risk bounds, and verifier acceptance.

Claim certificate

Strawberry

Measure whether an answer actually used its evidence. If the information budget doesn’t clear, the system abstains.

budget_gap_bits: +8.4
certificate: NEEDS_EVIDENCE
action: abstain
See how it works →

Security certificate

Injection Audit

Treat prompt injection as a distributional problem. Emit a lower bound on violation probability across attack families.

q̄: 0.23
qL: 0.18
verdict: VULNERABLE (qL > 0.05)
Explore the methodology →

Proof certificate

Nala + Lean4

Couple the model to a checker. “Proof closed” is a verifier statement, not a model claim.

verifier: Lean4
status: ACCEPTED
artifact: machine-checkable proof
See proof-carrying agents →
"

Transformers minimize expected description length across permutations—this explains why they look Bayesian in expectation.

"
Bayesian in Expectation Hassana Labs, 2025
Research & Publications

Information theory for thinking machines

We quantify the information required for a claim or action, then couple models to verifiers so outputs can ship with recomputable certificates.

Cancer Research

Cancer as the proving ground for information-theoretic thinking machines.

We integrate multi-omics data and treat missingness as an information-budget problem. When the model reconstructs masked values, we measure evidence sufficiency in bits/nats and use verifiers to keep downstream discovery stable and auditable.

Four data layers — expression, copy-number variation, clinical, and knowledge-graph — unified through a single learned function. Mask, reconstruct, discover. Cancer biology is the testbed because errors there have real cost.

Read the full initiative → Cancer Research Fellowship
Research Fellowship

Real research access for the talent that was overlooked.

Fellowships for underrepresented researchers. No compensation, but full compute access and co-authorship on published papers. Talent is everywhere — opportunity isn't.

Prompt Injection Research Fellow

Distributional prompt injection auditing — vulnerability across attack families with statistical confidence bounds.

3–6 months · Remote

Chain-of-Thought Hallucination Research Fellow

Why longer chains can reduce reliability — selection failures in CoT reasoning, and how to fix them.

3–6 months · Remote

Cancer Research Fellow

Multi-omics integration for cancer biomarker discovery: expression, CNV, clinical data, and knowledge graphs unified through LLMs.

6–12 months · Remote
View full fellowship details →

Contact

Talk to us about reliability.

We work with teams who are deploying AI in settings where wrong answers have real cost. If that's you, tell us a little about your use case and we'll follow up.

You can also email us directly at lc574@cantab.ac.uk.

If the form fails, email lc574@cantab.ac.uk directly.

Support

Frequently Asked Questions

Common questions about our tools and research.

What is information-theoretic reliability?

It quantifies AI trustworthiness using mathematical bounds from information theory—measuring how many nats (natural units of information) are required to guarantee reliability. This approach lets us make precise statements about when and why AI systems fail, rather than relying on empirical benchmarks alone.

How does Strawberry detect hallucinations?

Strawberry uses a technique called "evidence scrubbing" to measure whether a model actually uses its cited sources. It computes two probabilities:

  • p₀ — the model's confidence without access to the citation
  • p₁ — the model's confidence with the citation

If p₁ isn't significantly higher than p₀, the citation is flagged as decorative — the model is citing without actually using the evidence. The difference is quantified as a "budget gap" in bits.

What's the difference between procedural and factual hallucinations?

Factual hallucinations occur when the model lacks the necessary knowledge — it simply doesn't know the answer.

Procedural hallucinations are different: the correct information exists in the model's hidden states, but it fails to route that information to the output. Research shows approximately 74% of hallucinations are procedural, meaning the model "knows" the right answer but fails to produce it.

This distinction matters because procedural hallucinations can be detected and prevented with the right tools, while factual hallucinations require external knowledge sources.

How can I integrate these tools?

Strawberry is available as an MCP (Model Context Protocol) server, which means it integrates directly with:

  • Claude Code — Add Strawberry to your Claude Code configuration for real-time hallucination detection
  • OpenAI Codex — Use Strawberry skills in your Codex agent workflows
  • Python API — Call detect_hallucination() or audit_trace_budget() directly

For enterprise deployments or custom integrations, get in touch.

What does "budget gap" mean?

The budget gap measures how much information (in bits) is missing between what a claim requires and what the evidence provides. A negative budget gap (e.g., −2.1 bits) means the evidence strongly supports the claim. A positive budget gap (e.g., +8.4 bits) means the claim needs more evidence than provided — a hallucination flag.

Is Strawberry open source?

Yes. Strawberry is released under the MIT license. For access to the source code, documentation, and examples, get in touch or visit the Strawberry demo for integration guides.

Help us open doors in AI research.

Apply for a fellowship, support the work, or use the tools we've built. Talent is everywhere — opportunity isn't.

Apply for a fellowshipGet in touch