Framework · Benchmark · Live demo

Train models on real reasoning traces.

A causal-governance loss on multimodal introspective traces trains a self-representation r_t that must move the next action under counterfactual intervention — or pay a gradient penalty. The signal current reasoning models are trained without. Reflect is the two-hook adapter running live in a browser.

Try Reflect → Read the Whitepaper
T Reasoning Trace · Session 7f3a
Live
00:00
User Question
"Why did you come to my door…?"
00:02
Thinkingrecall.plan
Recalling related context…
First, I'll recall the poem we discussed.
00:06
Audio Thoughtinner_speech · 15.2s
0:00
00:10
Memory Recallconf. 0.82
Retrieved: "The little birds of love took flight…"
फुलली प्रीतीची पाखरे
phulali preetichi paakhare
00:14
Self-Correctionrevised
Wait — check context. The user asked about the reason, not the poem.
00:18
Tool Use2 results
Search internal notes
00:22
Final Answeroutput
I came right into your heart.

Reasoning Trace schema

JSON Schema (draft-2020-12) for time-aligned perception, inner speech, recall, uncertainty, correction, tool use, and affect. One format across human and model subjects.

Causal-governance loss

Penalize models whose action distribution fails to shift under do(r_t := r') to match a paired human counterfactual. Whitepaper §4.3 — interchange intervention training with a human behavioral distribution in place of the causal-model target (Geiger et al., 2022).

TRB — Reflection Benchmark

Two-hook adapter (emit_r, act(…, r_override)) with native and proxy tiers, so any checkpoint scores on the same governance number as a Trace-AI-trained model. Headline metric is the governance gap Δgov.

Reflect — the framework, running in a browser

Reflect v0.3 emits a multi-head r_t stack (task / self / meta / perception / correction lenses) plus a multimodal thinking trace of the model's pre-response events, on every turn. Voice in, voice out, image input, hard-bottleneck toggle, live governance score, per-turn discipline audit, and one-click session export as a valid Reasoning Trace v0.2. This is TRB's proxy adapter (Benchmark §2) with the whitepaper §8 L8 stacked-r_t direction made runtime.

Open Reflect → How it works Read the Benchmark
Reasoning Memory Uncertainty Tool Use Self-Modeling

The missing training signal for reflective intelligence.