Jason Prole

Co-Founder & Lead Researcher @AssiduityAI
Studying structural reliability in generative models

[Jason Prole @DrJasonProle]

Most "system of record becomes a harness" plans inspect finished output.

But drift is in the run , not the output. Fluent at step 19 and no longer on task.

By the time the app layer can check, the tokens are spent. The signal has to come from inside generation.

[Garry Tan @garrytan]

There will still be API’s and acl’s and sql and underlying data structures that are deterministic

It’s just that software companies have to build the AI harness and full solution for their customers or be subsumed by it

[Jason Prole @DrJasonProle]

Generation is a control surface. If candidate continuations can be evaluated against an operating mandate before the next path is committed, intervention can happen while the trajectory can still change.
After generation, that control surface is gone.

[Jason Prole @DrJasonProle]

@a16z A loop can converge on its verifier while drifting from the operating mandate. That is the deeper control problem. The verifier does not only decide when the work stops; it shapes what the system treats as progress throughout generation.

[Jason Prole @DrJasonProle]

Drift is a trajectory problem.

A final-output score can tell you whether the artifact passed review. It cannot show where the generation began moving off mandate.

[Jason Prole @DrJasonProle]

Equilibrium-Constrained Decoding���️ evaluates candidate continuations against an operating contract while the model is generating, not after the output is complete.

That timing is the point.

[Jason Prole @DrJasonProle]

Assiduity is built around a simple premise:
if reliability matters, control has to happen during generation, not only before or after it.

[Jason Prole @DrJasonProle]

The control problem gets harder as outputs get longer and workflows become dependent. Small deviations compound.

[Jason Prole @DrJasonProle]

Evidence should not just show what the model said.

It should show how the generation behaved:
constraints, deviations, corrections, and completion signals.

[Jason Prole @DrJasonProle]

A model can satisfy local token probabilities while violating global task requirements.

That gap is where generation-time control matters.

[Jason Prole @DrJasonProle]

Reliability is not only a label on an output. It is a property of the process that produced the output.

[Jason Prole @DrJasonProle]

If the system only inspects the final answer, it sees the result, not the path.

For enterprise AI, the path matters.

[Jason Prole @DrJasonProle]

The failure mode we care about is not only unsafe content. It is objective drift: the gradual loss of task, policy, or constraint fidelity across a long generation.

[Jason Prole @DrJasonProle]

Runtime control is not another prompt.

The prompt states the objective.

Control asks whether the generation is still following it.

[Jason Prole @DrJasonProle]

A long answer can be fluent, structured, and wrong in the only way that matters:
it stopped following the objective.

[Jason Prole @DrJasonProle]

Formal verification can prove properties of formalized rules.
Enterprise generation often faces a different problem: preserving objective fidelity across open-ended, sequential output.

[Jason Prole @DrJasonProle]

@saranormous @mvernal @igarciacamargo @trajectorylabs @saranormous Continual learning improves the model after it has acted. The harder problem is runtime control during generation, before the output exists to learn from.

The signal Trajectory trains on, we generate.

Complementary, not competing.
@AssiduityAI

[Jason Prole @DrJasonProle]

Guardrails often operate as policy decisions around a model. Our focus is the generation trajectory itself: whether the output remains close to the operating contract over time.

[Jason Prole @DrJasonProle]

Post-hoc verification is valuable, but it is still after the generation path has already unfolded.
Long-horizon reliability needs an active signal while the model is generating.

[Jason Prole @DrJasonProle]

There is a difference between deciding whether output is acceptable and influencing the path that produced it.

Assiduity is built around the second problem.

[Jason Prole @DrJasonProle]

@gdb Correct. It’s the control layer between the model and the user.