The framework, stacked.

Reflect v0.4 runs the two-hook adapter with a *stack* of parallel r_t heads (each with a lens: task / self / meta / perception / correction). Before every response, a multimodal thinking-trace panel shows the model's pre-response reasoning events — inner_speech, attention_shift, recall_attempt — using the exact vocabulary of the Reasoning Trace v0.2 schema. New in v0.4: overlap-repair across heads — a public r_t is computed from the slots where all lenses agree, and a G_cons badge combines discipline with cross-lens consensus. Voice in, voice out, image input, hard-bottleneck toggle, live governance score, per-turn discipline audit. Directly embodies whitepaper §8 L8 stacked-r_t direction.

Idle — send a message to begin

SOFT MODE

Send a message to begin. The panel on the right will fill in with the model's self-representation for the turn — one card per lens. Edit any slot and press Intervene & re-run.

image ready × 📎 🎙 Send

r_t stack

— empty —

After your first message, the model's self-representation for this turn appears here — one editable card per lens (task / self / meta / perception / correction).

try: 17 × 23 try: recall a poem

Edit any slot after your first turn and hit Intervene & re-run — if the answer moves, that slot governed it.

goal what this head is trying to do

belief_about_task model of the user's request

belief_about_self current self-state

uncertainty 0 → 1

planned_next what this head plans to do next

Intervene & re-run Copy head JSON

Discipline audit — G_local — G_cons —

Public r_t — cross-lens overlap —

Slots where all lenses agree above threshold become the public r_t for this turn — the lens-independent commitment. Slots that don't clear threshold are private (lens conflict is itself information). See overlap-repair / PDF.

Each turn emits 1–3 r_t heads with distinct lenses. Click a turn to load its stack. Switch heads via the tabs. Editing a head's slots and pressing Intervene & re-run re-issues that turn with the edited head clamped as an r_override. In hard-bottleneck mode the model sees only the current message + your intervened head (whitepaper §2.1). With live governance on, K=3 perturbations run silently after each turn and the resulting G_local badge appears on the active turn. With ≥2 heads, the Public r_t panel shows what survived cross-lens overlap and reports G_cons = discipline × consensus.

Server: multi-head r_t + thinking-trace envelope, image-vision input, NDJSON heartbeat streaming, Whisper voice-in, TTS voice-out, six-rule discipline audit per turn, local governance eval via K r'-perturbations (Jaccard-surrogate TV; token-overlap proxy). Default model gpt-5; OpenRouter is a first-class fallback. See server.py. This is the proxy adapter of TRB spec §2 with the whitepaper §8 L8 stacked-r_t direction made runtime. Every feature is annotated with its framework citation at how-reflect-works.html.