About

Trace AI is a company for a specific missing signal.

Every reasoning-model dataset in existence today is a dataset of outputs: text after text after text. When a model appears to think out loud, it is decoding text from its hidden state and then re-encoding that text as more context. Nothing in the training regime requires the text to have privileged causal status over what the model does next. So chain-of-thought looks like reflection but does not do reflection.

What is missing is the trajectory that produces the text. Perception. Inner speech. Attempted recall. Uncertainty spikes. Self-correction. Tool use. Attention shifts. Affect. All time-aligned at the level of a single cognitive session. In humans this trajectory is what "thinking" refers to. In current models it does not exist — not because thinking is ineffable, but because nobody has built the collection pipeline.

Trace AI is the collection pipeline. We publish a canonical schema for multimodal Reasoning Traces, we run the data-collection operation, and we train reference models against a loss function that requires the model's compressed self-representation to causally govern its next action. Testable by intervention: swap the self-representation, and the model's next action must move like a paired human counterfactual would.

How it started

At 3:01 AM on July 30, 2026 the founder was sitting with a four-line Marathi love poem retrieved from Google Translate:

Kasha aalaas daarat,
Manaat majachya aalaas.
Phulali preetichi paakhare,
Goad-sa kharali sakha re.

He sat with it for thirty minutes. He recited it aloud. He tried to recall it from memory. He caught himself uncertain about spelling, opened Google Translate again to check, re-attempted. He typed the Devanagari four times to encode it. He noticed himself noticing his keyboard: "I am recognizing each and every glyph on my keyboard in order to type this message. I read back, the reader will not perceive it as me typing it out. They will simply see text."

None of that is recoverable from the finished text. The finished text has it all flattened. That trajectory — the layered process of production — is what Trace AI records.

That session is Session 7f3a in our reference corpus. The full corpus is viewable here, and every future trace we collect will link to that first one by schema pattern.

What we build, at four levels

• The pitch. The missing training signal for reflective intelligence.
• The artifact. A schema, a growing dataset, a training program, and a reference model that behaves reflectively where its baselines do not.
• The flywheel. A data-collection app used by a small expert workforce, producing traces that go into a versioned dataset, that trains reference models, that are published as APIs, whose usage generates more traces.
• The science. A public schema, a public reference dataset, and a defensible training objective — an ImageNet-style seed for the reflective-reasoning regime.

What Trace AI is not

• Not a claim about consciousness. We do not know whether the resulting model is conscious. We do not claim it is. We claim its self-representation causally governs its computation, which is a testable property.
• Not "just chain-of-thought." CoT is a decoding strategy; Trace AI is a training regime. Our loss function forbids the failure mode in which the emitted reasoning text does not steer the model.
• Not a wearable / brain-scanning company. We collect self-reports and behavior, not neural signals.
• Not an interpretability company. Interpretability extracts from what already exists. We produce models whose reflective layer is trained-in by construction. Complementary, not overlapping.
• Not a "record everything you do" company. We record deliberate, consented cognitive sessions of finite length under a defined stimulus. No always-on capture. No ambient telemetry.

Who

The company is at the founding-engineer-plus-AI-collaborator stage. If you have taste for training-signal problems, alignment-legibility problems, or introspective-data problems, we want to talk.

Get early access

Trace AI is not yet a self-serve product. If you are an alignment researcher, an interpretability researcher, an ML engineer working on reasoning models, or an operator with a use case for legibly-reasoning models, tell us. We reply personally.

Email * Name

Role / affiliation Why (optional)

Request early access

By submitting, you agree we may email you about Trace AI. We do not add you to any mailing list. See privacy.

How to reach us