Everything Trace AI has written down.

The framework. A three-term training objective (trace-reconstruction, self-representation supervision, and the load-bearing causal-governance loss) that promotes CoT-perturbation diagnostics into a training penalty. An information-bottleneck architecture whose readout is legible by construction. A reference corpus across two subject kinds (human, model). Situated relative to interchange intervention training (Geiger 2022), CoT faithfulness (Turpin 2023 / Lanham 2023), monitorability (Korbak 2025), concept injection (Lindsey 2025).

The thing to send a lab lead. IIT positioning + the governance gap + the TRB adapter, condensed to a single page with a concrete ask (wrap your model in ~50 lines, run the smoke test). Ends with a 20-minute reading order.

The two-hook adapter contract (emit_r, act(..., r_override)) that makes "faithful reasoning" one number computed identically for a Trace-AI-trained model and for any GPT / Claude / Llama baseline. Six tracks (governance, grounding, self-attribution, viability, navigability, transfer); the headline number is the governance gap, not any single level. Also the runnable Track-1 smoke test on Session 7f3a — see the Benchmark page.

Representation as a stochastic fibration, with nine exact executable instruments. Existence via Halmos–Savage minimal sufficiency (Theorem 1); Shannon rate–distortion parameterisation (Theorem 2, witnessed exactly by Instrument 7 on uniform and Bernoulli sources to 10⁻⁹); categorical adjunction (Proposition 3); cross-task stability iff shared Markov screen (Theorem 4, conditional); discrete learnability at N ≥ cM ln(M/ε) (Theorem 5); continuous-case extension at resolution ε (Theorem 6); Theorem 7 (linear-ICA identifiability) resolves the exponential-in-d_Z cost inside the linear-ICA class (Instrument 8: Amari ≤ 0.009 at N=10,000, 462× escape from Theorem 6's bound); Instrument 9 (sparse-ICA / IMA, Gresele et al. 2021) gives a second positive resolution of SIC-C-c under sparse mixing — two distinct inductive-bias classes now provably earn escape from the ε-covering bound. This paper's Instrument 3 (agency benchmark) is the exact-solvable formalization of TRB Track 3 (self-attribution); Theorems CG-1 and CG-2 (companion paper below) give TRB Tracks 4 and 5 their formal shape. Prior PDFs preserved: v2, v3, v4.

Turns SIC extended-program construct #1 (concern as a reweighting of the compiler) from a target into two theorems. CG-1: under an exponential-family concern reweighting, the fibre-restricted Fisher matrix is gc,z = β²·Cov[T] — two concerns are distinguishable in n samples iff their KL is ≳ 1/n. CG-2: the concern 1-form α has zero holonomy on every closed loop iff it is exact (α = dΦ) — non-vanishing holonomy is exactly the statement that concern is irreducibly path-dependent. Worked example on Instrument 4's 4-bit Boolean world; rectangular- and triangular-loop holonomy matches predictions to 10⁻⁶. Trace-AI relevance: CG-1 gives TRB Track 4 (Viability) its sample-complexity meaning (how many trials to detect a Δconcern between full model and no-r_t ablation); CG-2 gives TRB Track 5 (Navigability) a path-dependence diagnostic (whether steering the self-model preserves its semantics across contexts).

The sibling of the formal paper, held to a stricter mathematical-claim gate: the conjecture and the extended formalism are labeled speculative theory, only three of the six instruments are built (Fiber Finder, Structure Compiler, Agency Benchmark), and the concern-geometry / rate–distortion-control / ensemble-alignment lines are flagged as conjectural. Kept alongside the formal paper because the two documents differ on purpose — this one is the working note; the other is what it grew into.

The selection-theoretic development of whitepaper §3.1–§3.2. States precisely what it means for r_t to be an agentic variable, defines the six functional criteria as a measurable profile, and states what Trace AI does and does not claim about consciousness. Kept as a companion so the whitepaper stays a tight technical document while the ontology motivating it lives somewhere legible.

Intelligence, on this frame, is finding the representation in which a problem becomes at once compressible and controllable. Trace AI is that conjecture applied to cognition: the schema says what to keep, the corpus proves the same structure compiles across humans and models, and the benchmark measures whether what we kept actually controls the outcome. Three build directions (representation search, the compiler, symbolic-causation measurement) mapped onto product surfaces.

The concrete session design that unlocks TRB Tracks 1 and 3. Defines the counterfactual pair unit, the cause-labeled outcomes for self-attribution, session recipe, on-disk layout, sample-size targets (pilot n=20 → PoC n=500 → pretraining-quality n≥10⁴), consent and safety. A research assistant can execute the pilot from this document without further design work.

The canonical multimodal trace format. Sessions, subjects (human / model / hybrid), stimulus, channels, event vocabulary, self-representation snapshots, and cross-session references (parent_session_id, order) supporting first-, second-, and third-order traces. Field-by-field reference with a browsable example.

The three sessions the schema was built to fit. 7f3a (order 1, human, typed); 7f3b (order 2, human, spoken — internal-only for PII); 7f3c (order 3, Claude Opus 4.7 — first model-authored trace). The reaffer → re-affir → reaffirm chain runs across all three and closes in the model.

The two-hook adapter (emit_r, act) made interactive, now with a multi-head r_t stack (lens-labeled task / self / meta / perception / correction), a multimodal thinking-trace panel streaming pre-response events in the schema's canonical channel vocabulary, voice in / voice out, image input, hard-bottleneck toggle (whitepaper §2.1), per-turn discipline audit against six rules, live per-turn governance score, and one-click session export as a valid v0.2 Reasoning Trace JSON. Naming discipline: reflective, not conscious — functional selfhood does not imply subjective consciousness. Every feature annotated at how-reflect-works.html.

Same fixed task suite × selected models × the six-rule discipline audit. Scores how well each model (OpenAI, plus Claude 5, Qwen 3, DeepSeek R2, Llama 4, Kimi K2 via OpenRouter) holds the two-hook contract as a proxy adapter, without any ℒ_gov training. Results persist to your browser and export as CSV. Not a benchmark result — a snapshot comparison. A high score = the architecture is a strong substrate for ℒ_gov training; a low score = it can't hold the framework's constraints in the first place. Full form of the design in RESEARCH_DIRECTIONS §1.

Every load-bearing design choice with the reason behind it and a rollback condition. If a document above surprises you, this is where the "why" lives.