Observe the live moment
The system starts with the current conversation or replay turn. Reality remains the root, not the model's imagination.
Premonition
A prepared mind behind the voice.
An experimental backend for conversational AI that predicts likely next moves, prepares safe draft responses in shadow mode, waits for reality to confirm a branch, and then measures what actually improved.
21.7%
Base readiness
First branch immediately usable at baseline.
76.5%
Guarded readiness
Usable prepared draft after branch confirmation.
25/25
Stress folds
Recovery rule cleared ESConv shuffled gates.
560ms
Latency saved
Median estimate when preparation is selected.
Anticipate - prepare - protect - learn
Measured preparedness
Not prophecy. A better rehearsal loop.
Premonition is built around a simple distinction: the speaking model should stay present, while a hidden backend rehearses likely next turns and keeps every speculative draft behind a confirmation gate.
Prepare possible next moves
The backend predicts likely response modes, builds bounded drafts, and keeps them hidden until the observed branch matches.
Grade what happened
Every hit, miss, quality gate, latency estimate, and segment regression feeds back into the benchmark loop.
Current evidence
The backend is getting measurably more ready.
In held-out tests, the important question is not whether the system can blindly speak early. It is whether, once the next conversational direction becomes clear, a useful speech-ready draft is already waiting.
21.7%
to 76.5%
to 76.5%
Usable prepared response availability after branch confirmation, up from the original first-speech baseline.
+0.219
Quality-ready lift
Recovery policies increased prepared coverage while preserving the TTS-readiness quality floor.
+0.102
Mean stress gain
Average quality-ready improvement across the 5-seed x 5-fold ESConv stress test.
10/15
Second-corpus transfer
DailyDialog promoted safely on ten folds while weak folds stayed baseline instead of forcing regressions.
149
Verification checks
Replay, Probability Pack, live-shadow, CLI, and dashboard tests covered the current implementation checkpoint.
Parallel cognition
One lane speaks. One lane rehearses. Reality teaches both.
This is the core experiment: keep the front-facing agent fast and present, while the hidden backend prepares branches in parallel and waits for confirmation before anything reaches the user.
Conscious conversation1
User speaks
"I am not sure what to do next."
Frontend LLM
Stays present, understands the current turn, and avoids carrying the whole future in working memory.
Agent replies
Uses confirmed context only. Speculation stays hidden.
Hidden Premonition backend2
Reality and learning loop3
Real next turn
Validate
Confirmed branch
The observed next move matched a prepared validation branch. The draft can be used quickly after confirmation.
Grade
exact match feeds the benchmark report.
Refinement
Hits, misses, weak response modes, and protected-slice regressions update the next run.
Live Shadow Lab
The experiment can run beside a real conversation.
The lab observes transcript turns, generates hidden Premonition drafts, accepts the actual next move, grades draft readiness, estimates saved latency, and exports replay rows for future benchmarks.
Transcript in. Drafts prepared. Reality graded.
Shadow-mode testing before voice runtime integration.
Architecture layers
The conscious model does not carry the whole future.
Premonition separates the visible conversation from the background rehearsal layer, then scores whether rehearsal actually helped.
Branch generator
Predicts likely next events or response modes from the current context.
Artifact builder
Prepares bounded drafts, policy checks, and tool plans for useful branches.
Safety filter
Keeps speculation hidden until reality confirms a matching branch.
Replay evaluator
Compares prediction, preparedness, latency, cost, and safety across runs.
Progression
From offline replay to a voice-ready shadow layer.
The current project is still a research harness, not a production voice agent. The honest milestone is that the backend can now be measured, visualized, and tested in shadow mode.
Offline replay harness
Measure branch prediction, prepared artifacts, latency estimates, and unsafe leak rate across repeatable turns.
Probability Pack and recovery policies
Train guarded preparation behavior and accept only improvements that survive held-out checks.
Live Shadow Lab
Run beside manual or live transcript turns, inspect prepared drafts, and export new benchmark rows.
Voice runtime integration
Connect the shadow layer to a TTS/voice stack so confirmed drafts can reduce perceived response delay.
Outcome dashboard
Base state vs Probability Pack vs guarded swarm.
Can AI become more present by preparing before it speaks?
Premonition is a build-in-public attempt to answer that question with benchmarks, shadow-mode testing, and clear safety boundaries. The magic is not prediction. The magic is measured readiness.