Most model failures are discovered from the outside.
We investigate what happened inside.
SnailSafe is an independent white-box AI diagnostic research lab. We reproduce unusual model behavior, measure how the internal computation unfolds, and test competing explanations with controlled interventions.
The lab grew out of a question, not a product roadmap.
SnailSafe began by asking whether model behavior that looked mysterious from the outside could be measured internally.
That question quickly became harder.
It was not enough to find a signal. A signal could be correlated with an answer without explaining it. A hidden state could be measured without revealing the computation that produced it. A convincing pattern could disappear when tested on a new population.
So the work expanded from observation into replay, internal measurement, trajectory analysis, causal intervention, and controlled experimental practice.
A diagnostic laboratory, not an interpretability dashboard.
- A white-box diagnostic research lab for open-weight or otherwise internally accessible models
- A set of measurement and causal-intervention instruments
- A controlled experimental method for testing internal explanations
- A way to distinguish capability failure from ranking, branching, trajectory, or control problems
- A research partner for model developers, post-training teams, evaluators, and safety researchers
- A universal hallucination detector
- A black-box monitoring service
- A claim that one probe explains how a model “thinks”
- A replacement for behavioral evals
- A guarantee that every model failure reduces to one clean mechanism
The hidden state is not the whole explanation.
Hidden states tell us where a computation has arrived. They do not, by themselves, tell us what produced that state or why one continuation became more probable than another.
SnailSafe therefore studies both state and transition: how candidate rankings form, where trajectories branch, what changes after a sampled token, and which interventions causally alter subsequent computation.
A real instrument must describe both what it measures and where it stops.
SnailSafe findings are empirical and bounded by the model, prompt, checkpoint, measurement, and intervention that produced them.
A calibrated readout is not automatically a universal detector. A causal intervention at one state does not imply permanent control of the trajectory. A pattern discovered in one population does not become a general law until it survives a separate test.
Those limits are part of the result.
Current as of October 2026
What the evidence supports today —
and where it stops.
Every SnailSafe claim is bounded by the experiment that produced it.
P2 studied Llama-3.3-70B-Instruct under one governed prompt family. Its Λ calibration has not yet been shown to transfer across prompts or models.
SnailSafe has not yet completed a before-and-after post-training study. Comparing internal behavior across successive model checkpoints is a proposed application of the method, not a result already on record.
In P2, Λ separated a defined premise-accepting versus premise-correcting contrast at the verdict landmark. It did not predict which correcting runs would later fabricate arithmetic, and it provided no useful advance warning of that later behavior.
At the tested P2 state, a hidden-state write changed which candidate the model selected without changing the preceding text. P2 did not directly test an end-to-end persistent controller from that state through an independently measured completed answer.
Changing an early available token causally rerouted the later trajectory. It did not explain how the trained network constructed the ranking that made that branch consequential.
Λ has been read during inference without changing the model’s output. Production integration into a serving system remains untested.
The boundary moves when the evidence does, and not before.
An engineering laboratory built around measurement.
Greg’s background spans traveling-wave tube engineering, large-scale computing systems, product development, and independent AI research.
SnailSafe applies the same engineering discipline to model inference: define the failure, instrument the system, freeze the test, measure what happened, perturb where justified, and preserve the result.
AI models are used throughout the laboratory as builders, reviewers, analysts, and critics. Their work is governed by frozen protocols, preserved evidence, separate audit and adversarial review passes, and human experimental control.
Failed hypotheses, apparatus failures, corrections, negative results, and superseded interpretations remain in the record. The objective is not to make the research look linear after the fact.
Modern models can produce the same outward failure for very different internal reasons.
- A capability may be absent.
- It may be present but poorly ranked.
- A stochastic branch may redirect the continuation.
- Post-training may have changed the trajectory without obviously changing the task capability.
- An intervention may move the next token and still fail to control the completed answer.
Those are different engineering problems.
SnailSafe exists to tell them apart.
We respond directly, usually within a day or two.