Why does your AI model behave differently on some runs and not on others?
SnailSafe examines the model's behavior from the outside and its computation from the inside, then tests which explanation survives.
A battery of instruments, a method for using them, and a record you can check. From model behavior to testable evidence.
LLMs can produce different behaviors from the same input, while the internal computation that produced those behaviors remains largely opaque.
White-box access gives us an opportunity to investigate that gap directly.
SnailSafe's mission is to develop and refine the methods, measurement instruments, and experimental practices needed to understand how inference unfolds inside open-weight models. We combine behavioral evidence with internal measurement and controlled intervention to test competing explanations and determine which ones survive.
Our goal is not to make every computation simple. It is to make consequential model behavior more observable, testable, reproducible, and ultimately more governable.
You receive the findings—and the investigation behind them.
An engagement delivers an executive findings report and an investigation lineage: the findings, the evidence that supports them, and the record of how our understanding changed, including the hypotheses tested, the explanations rejected and the rulings made along the way. Both are provided under the confidentiality and intellectual-property terms agreed for the engagement.
In our P2 case study, a proposed detector based on a late internal signature appeared in 14 of the 15 first-population runs that showed the behavior, and failed its 15-of-15 criterion, frozen in advance. The detector claim was rejected; the record kept the rejection and the investigation that followed it.
P2 has also been converted into a sample customer-facing diagnostic findings report. It shows how the experimental record is reduced to the findings, evidence, boundaries and engineering implications a model team would receive.
Several instruments.
Different views of the same inference.
SnailSafe develops white-box inference trajectory diagnostics for open-weight AI models. No single measurement explains model behavior, so we use complementary instruments to observe, compare, and perturb inference from different perspectives.
External Research Methods
SnailSafe also uses established outside methods when they strengthen an investigation.
The instruments produce measurements.
The SnailSafe method determines what question to ask, which measurement is appropriate, what controls are required, and how far the result can be interpreted.
A model failure is a research question,
not a conclusion.
We begin with the behavior you observed, turn possible explanations into testable hypotheses, and define each experiment before seeing its result.
When the evidence says we are wrong, we stop and say so.
When it says we are right, we show it.
Every diagnostic leaves a traceable record from question to measurement to evidence to ruling.
The data can change the next question. It cannot change the test already in progress.
We connect external model behavior to internal measurements and produce diagnostic findings that help model teams decide what to change, what not to change, and what still needs to be tested.
See how SnailSafe investigates difficult model behavior
from observation through causal testing and diagnostic finding.
The model failed. What should your team test next?
An incorrect response does not, by itself, tell you what to change.
In P2, when supplied a valid first factor, the model produced the correct complementary factor in all three tested cases—but assigned those factors very little probability during natural generation. Simple inability to complete the arithmetic did not explain the failure, so the investigation turned to candidate ranking and selection.
For a post-training team, the potential benefit is better evidence for choosing the next experiment: changes to training data, reward signals, decoding, or verification.
SnailSafe investigates the failure before prescribing the remedy.
See the P2 evidence and its implications for post-trainingThe laboratory continues past the case.
Research grounded in a documented experimental record.
SnailSafe develops Dynamical Interpretability: the study of how an answer forms inside a model during inference, measured as it happens. Two preprints and a public episode series. We publish research we are authorized to share. Customer investigations are not automatically public; any publication is governed by the terms agreed for that engagement.
Silent Commitment Failure in Instruction-Tuned Language Models
Introduces governability — whether a model's errors are detectable before output and correctable once detected — and shows it diverges sharply across six architectures.
Structural Rigidity and the 57-Token Predictive Window
Extends governability into an energy-based framework, using trajectory tension (ρ) to locate a pre-commitment window and map five regimes of inference behavior.
Two episodes from the observatory
An answer forming before it appears as text, then a second instrument following those representations to where the decision forms.
Reviews of external work
What other groups have found, read against our own measurements and labelled by how it relates: convergent, complementary, contrasting, boundary-setting.
Our goal is to correlate internal changes with each model revision, so that post-training depends less on trial and error. Before-and-after post-training comparison is a proposed research application; no such study has yet been completed.
Evals reveal that a problem exists.
They rarely tell us why.
SnailSafe works with open-weight model developers, post-training teams, and research groups to turn puzzling model behavior into testable internal diagnostics. We reproduce the behavior, test competing explanations, and produce findings that help you decide what to change, what not to change, and what still needs to be tested.
Diagnose unexpected behavior
Find out whether a failure comes from missing capability, poor candidate ranking, stochastic branching, or another internal cause.
Proposed: investigate what changed between checkpoints
Compare behavior across model checkpoints and test whether internal measurements can help explain a post-training regression.
Turn observations into evidence
Use white-box diagnostics to investigate anomalies, validate hypotheses, and provide additional evidence for evaluation and model-development decisions.
Send a short, nonconfidential description of the behavior you want to understand, the model involved, and what you have already tried. We’ll reply with whether it looks like a fit and what we would need to understand first.
snailsafe.ai
From model behavior to testable evidence.