Trace AI — one-page partner brief

For: alignment / interpretability leads at frontier labs. Ask: wrap your checkpoint in a two-hook adapter and run one benchmark. v0.1, 2026-08-03. Contact: hello@use-trace.ai.

The one sentence

Chain-of-thought is not, in general, causally load-bearing on the answer it precedes; Trace AI trains a self-representation $r_t$ that has to move the next action or pay a gradient penalty, and ships the benchmark that measures whether any model — ours or yours — actually satisfies that condition.

The three things a partner is actually buying

1. A training-signal framework — a causal-governance loss $\mathcal{L}_{\text{gov}}$ that promotes the CoT-perturbation diagnostics of Turpin et al. (2023) / Lanham et al. (2023) into a training objective, and an information-bottleneck architecture (§2.1 of the whitepaper) whose readout is legible by construction, not by hope. Reads on: Geiger et al. (2022) — Trace AI is interchange intervention training with a human counterfactual distribution in place of the causal-model target.
2. A reproducible instrument — the Trace AI Reflection Benchmark (TRB) — a two-hook adapter (emit_r, act(..., r_override)) with a native tier for models that have an $r_t$ pathway and a proxy tier for any GPT / Claude / Llama baseline. Both tiers answer through the same interface, so “faithful reasoning” becomes one number computed the same way for a Trace-AI-trained model and for your checkpoint. Six tracks (governance, grounding, self-attribution, viability, navigability, transfer); the headline number is a difference — the governance gap $\Delta_{\text{gov}} = G(\text{full}) - G(\text{ablation without }\mathcal{L}_{\text{gov}})$ — because a level can be gamed and a gap can’t.
3. A data operation with a schema, not vibes — a JSON Schema (draft-2020-12) for a multimodal Reasoning Trace, a reference corpus of three sessions across two subject kinds (human, model) with a worked cross-substrate correction chain, and a documented collection protocol for the paired-counterfactual sessions that populate $\mathcal{R}^*$ (see docs/COLLECTION_PROTOCOL.md).

Where Trace AI sits in the safety conversation

• vs CoT monitorability (Korbak et al., 2025). They flag that CoT monitorability is fragile because it is unoptimized. Trace AI is a bet that interventionally-grounded legibility is a stronger property than incidental legibility, at the cost of directly optimizing the transparency channel (Goodhart risk). Which bet wins is what the governance gap and Track 6 (transfer) exist to adjudicate. We do not claim to have dissolved the tension.
• vs concept-injection interpretability (Lindsey, 2025). Lindsey’s grounding condition — a self-report counts as introspective only if it causally depends on the state it names — is exactly the condition $\mathcal{L}_{\text{gov}}$ trains for in the report→action direction. Concept injection (state→report) is TRB Track 2. A model that passes both is a self-representation, operationally; a model that passes only one is a diagnostic.
• On consciousness (Nagel 1974; Chalmers 1995). No claim, by design. The functional-aliveness profile is an instrument for asking the phenomenal question sharply — of a system whose self-model is known to causally govern, be grounded, and self-attribute — not an answer to it. See docs/AGENCY_AND_SELECTION.md §5.

What we are asking a partner to do, concretely

Small first move, no license negotiation, and it gives you a number nobody else has on your own model:

0. See it live first. Open use-trace.ai/reflect.html. Reflect v0.3 is the two-hook adapter running against gpt-5 with a multi-head r_t stack, a multimodal thinking trace (using the schema’s canonical channel vocabulary), image input, hard-bottleneck toggle, per-turn discipline audit, and a live per-turn G_local score (K=3 r’-perturbations, TV-surrogate). Every session downloads as a valid v0.2 Reasoning Trace JSON. This is the proxy adapter of the TRB spec §2 with the whitepaper §8 L8 stacked-r_t direction made runtime — none of the framework’s central claims depend on training yet, and Reflect is honest about that.
1. Read the TRB spec (docs/BENCHMARK_SPEC.md, ~5 pages).
2. Wrap your checkpoint in the proxy adapter (~50 lines against the two-hook contract in §2). Reflect’s server.py is a reference implementation.
3. Run the Track-1 smoke test against Session 7f3a’s reference continuations (tooling/trb/run_smoke.py). You get a governance score $G$ and a bootstrap CI on your own model, comparable to what we report on ours.
4. If the number is interesting to you, the second move is a paired-counterfactual collection (protocol in docs/COLLECTION_PROTOCOL.md) so Tracks 1 and 3 have real $\Phi$, not synthetic reference. That is a $10^3$-pair operation — days of skilled work, not months. Track 3’s concrete pilot stimulus set (n=20 cause-labeled pairs, executable today) is in docs/track3_pilot/.

What we are honest about

• No training results yet. The whitepaper is a framework; the benchmark is a spec plus a runnable Track-1 harness on n=3. Do not read “TRB score” reported by anyone (us included) until Tracks 1 and 3 have real paired data. That collection is the critical path.
• The corpus is $n=3$: two human sessions (one public, one held for privacy) and one model-authored session. It is enough to exercise the schema across humans and models and to demonstrate a cross-substrate correction chain; it is not enough for a proof-of-concept of $\mathcal{L}_{\text{gov}}$ (that needs $\sim 10^3$ counterfactual pairs; §8 L2 of the whitepaper).
• The framework’s monitorability argument (whitepaper §8 L7) is a bet under a live tension, not a proof. The disanalogy with Korbak et al.’s scenario is the human-readability constraint in §2.1; whether that constraint is strong enough is the empirical question the benchmark exists to answer.
• Sample-complexity has a strategic lever, not just a scale-up. The Structural Intelligence Conjecture paper (docs/SIC_MATHEMATICAL_FOUNDATIONS.md) Theorem 7 shows that with the right inductive bias — statistically independent, non-Gaussian slot components — $r_t$-recovery flips from exponential-in-$d_r$ to polynomial-in-$d_r$ (classical linear ICA; Instrument 8 witnesses it numerically). Architecting $h_\phi$’s bottleneck with an ICA-flavored objective would shrink the L2 budget from “$10M–$100M collection operation” to “engineer the bottleneck for identifiability.” That is a research direction the partner engagement can take on directly.

Reading order for a lead who has 20 minutes

1. This page.
2. docs/BENCHMARK_SPEC.md §§1–2 (the two-hook adapter — this is the product).
3. docs/WHITEPAPER.md §§1, 2.1, 3, 4.3 (the framework, tight).
4. docs/STRUCTURAL_INTELLIGENCE.md §3.3 (why the eval suite is the artifact a lab licenses).

If those four reads don’t move you, the framework is not for your lab and we’d rather know now.

v0.1, 2026-08-03. See DECISIONS.md (D26).