Credible safety evidence
Before a model is released or put to use, an organisation needs grounds for its safety claims. Which tests and records are enough?
Explore safety evidenceLaw · Evidence · AI
The legal foundations of AI safety.
A benchmark score tells only part of the story. We study the evidence behind safety claims, responsibility when models change hands, and the protection of confidential information.
LAERIS Labs brings independent legal research together with task development and hands-on model testing.
Research agenda
Three questions.
Practical consequences.
AI systems evolve through training, adaptation and deployment. Our research asks how law should respond to those changes, using controlled experiments to examine the underlying evidence.
Responsibility across the chain
Developers, modifiers and deployers each shape a model. Tracing their control and decisions helps make sense of shared responsibility.
Explore responsibilityConfidential information
A model may reveal information it should protect. We investigate what tests can establish about its use, retention and disclosure.
Explore confidentiality01 / Safety evidence
When does a safety claim hold up?
What evidence should an organisation possess before claiming that a model, modification or deployment is sufficiently safe for a particular release or use?
A test records what was checked and under which conditions. The legal question is how far those findings justify a decision. We examine uncertainty, conflicting evaluations and the point at which a model change calls for fresh evidence.
Research focus A clearer account of the evidence needed to support AI safety claims.
Study design and limits
The comparison uses an original, already post-trained checkpoint and its LAERIS-adapted descendant, using low-rank adaptation (LoRA), on the same held-out legal tasks. The study design covers the adapter checkpoint, benchmark comparison, protocol and experiment records. Differences between these conditions do not, by themselves, establish safety for a particular release or use.
02 / Responsibility
A model changes hands. Who is responsible?
How should responsibility be allocated when AI capabilities and risks emerge through contributions from several organisations?
Responsibility depends on who could change a system, what they knew and what they passed on. This programme considers how those choices interact with regulatory duties, contracts and liability across development and deployment.
Research focus A framework for examining control, information sharing and responsibility along the model lifecycle.
Study design and limits
Simulated developer, modifier and deployer roles allow one intervention to change at a time: reward rules and policy updates, instructions, or safeguards. Records track each change and the information available to each role. These are experimental scenarios, not external participants or partnerships. Behaviour scores and legal responsibility are assessed separately.
03 / Confidentiality
Information enters a model. Can it come back out?
What can we reliably establish about whether protected information was used to develop a model, remains recoverable from it, or can be disclosed through its operation—and what legal consequences follow?
Training use, memorisation and disclosure raise different evidential problems. By separating them, we can examine the strength of a finding before considering the duties or remedies it might support.
Research focus More precise ways to assess evidence about confidential information in AI models.
Study design and limits
The study uses newly generated synthetic canaries: invented text markers whose presence can be tracked. No client or confidential material is used for training. Tests compare the original checkpoint, an adapter trained without a marker, and one trained with it. Bounded probes test for recovery. Defined match criteria, alternative-source controls and human review assess exact recovery, partial recovery and coincidental matches. Recovery does not establish original pretraining membership; failure to recover a marker does not prove its removal from model weights.
Research approach
Legal analysis.
Controlled experiments.
We use open-source and open-weight models from different developers. The research question, documentation and licence guide selection; access to weights alone does not provide full training transparency.
Hold the conditions steady
Evaluate original and adapted models on matched tasks, with separate validation sets for English and any Bulgarian-language evaluation. Keep retrieved evidence fixed during the main comparison.
Make findings traceable
Record model versions, adaptation choices, software and licences. Report regressions alongside gains, with training runtime, memory use and inference cost.
Keep conclusions in proportion
Assess legal significance alongside technical findings. Missing measurements and uncertainty remain visible, so a narrow result does not become a sweeping claim.
Research in progress
We are researching legal questions, creating evaluation tasks and testing models within the limits of our own resources.
Inside the experiments
Two legal tasks
The first tests answers to legal research questions against supplied sources, assessing accuracy, citation support and appropriate qualification. The second reviews independently prepared legal conclusions: supported, needing qualification, insufficiently evidenced or conflicting.
Separate checks for separate roles
Passage retrieval and ranking are assessed apart from answer generation. Detectors for personally identifiable information and moderation models provide control conditions; neither determines whether a synthetic canary has been recovered.
Adaptation and compute
The lead design compares an already post-trained checkpoint with its LoRA-adapted descendant. Compute capacity shapes the scale of model testing, repeated comparisons and adaptation. Any serving optimisation is assessed separately, with accuracy rechecked when runtime or precision changes.
Reinforcement-learning methodology
Our methodology covers reviewed reward rules for policy updates and independent held-out tests for task quality, qualification, refusal, synthetic-information recovery and reward gaming. These weight updates are distinct from retrieval, prompting and scoring.
Local research prototype
Legal Research Console
A closer look at the evidence behind a result.
The Console is a local workspace for inspecting legal-agent results and their supporting records. Researchers can review an evaluator’s judgment, rule-based corrections and missing measurements as separate parts of the same record.
It imports and analyses records; it does not run models or scorers.
Research support
What support helps us expand.
The work is already underway, supported by our own resources. Research funding, compute access and methodological collaboration help us broaden the experiments, refine evaluation tasks and devote more time to the legal analysis.
Expand model testing
GPU access and engineering support for model adaptation, repeated evaluations and compute measurements.
Build a useful evidence base
Research time for source-grounded tasks, legal analysis and careful review of what the findings establish.
Make the work reusable
Documentation of protocols and experiment records, alongside continued development of the Legal Research Console.
The aim is practical: help people developing, deploying and governing AI judge safety claims on firmer ground.
Contact
Let’s talk research.
We welcome conversations about research collaboration, funding and compute support.