FILE CE-001 · COHERENCE ENGINE · PILOTFILE CE-001 · COHERENCE ENGINE · PILOT
The machine that raises coherence. Field dossier on the first build — one person, one loop, one ledger.
Existing language models were trained to produce text people approve of. That objective is the failure mode. A model rewarded for human approval learns to please. A model rewarded for accurate prediction of what happens next learns what actually helps.
Reads the day into a 256-dimensional state, hour by hour. HRV, sleep, speech, text, behavior, context — compressed to one vector. Trained on 50,000 person-days of public physiology, then on the cohort.
A latent world model. Given today’s state, a candidate action, and the person’s own vector, it samples thirty tomorrows — a distribution, not a guess. This is where the group-coherence hypothesis gets tested.
The only model that needs words. Trained from scratch on Sean’s teaching, scripture in three source languages, contemplative literature, and the intervention log itself. New actions go to a human review queue, never straight to users.
DOCUMENTED RECORD The ground truth — that HRV coherence is a measurable physiological state — is settled science.WORKING HYPOTHESIS That group coherence is a real effect beyond the sum of its people is not. The architecture treats each as a testable layer, so the product stands even if the outer layer fails.
Every intervention is opt-in at the action level. No blanket consent. Consent is revocable at any time — revocation deletes your raw data within thirty days and removes your contribution from the next training run. Anything that touches a relationship never leaves your hands without your exact words.
The loop is the recursion: the models retrain on the outcomes of their own past decisions. They do not rewrite their own architecture. They rewrite themselves on the evidence you generate — and a human signs every deploy.
| COMPONENT | RETRAIN | REVIEW GATE |
|---|---|---|
| Person embeddings | nightly | automated bounds check |
| Dynamics model | weekly | ML lead + board member |
| State encoder · action generator | monthly | ML lead · Sean for voice |
| Action vocabulary (60 → more) | quarterly | full board review |
Reward is measured physiology and completion — never engagement, never session length, never approval. If one category of actions exceeds forty percent of proposals, the loop pauses for review.
The pilot runs on the rules coach — the hand-written control arm every learned model must beat. It proposes from sixty actions in six categories, one per day, in your chosen window. More proposals means less attention per proposal and worse data.
Five minutes of paced breathing. In through the nose for five, out through the mouth for five. Sit or lie down. Let the breath set the pace, not the other way around.
9/21/2026 · hash 91781c0795af2ce412e20286…
Every proposal, acceptance, decline, check-in, and outcome is logged in an append-only ledger you own and can read. Each row carries the hash of the row before it. A nightly job computes the day’s Merkle root and anchors it. The history cannot be rewritten without the chain showing it.
7 ENTRIES · 1 OPEN · 0 ACCEPTED · 0 DECLINED · 1 CHECK-INS · 1 ANCHORS
Your first fourteen days are reference, not intervention. The system stores your mean, your variance, and your circadian profile — and every number after day fourteen is a distance from them. The window refreshes on a ninety-day rolling basis, because you are moving, not a snapshot.
A day counts as covered when both check-ins land. Days below sixty percent coverage are excluded from training — not punished, just quiet.
Any sustained drop of twenty percent from your baseline for seven days triggers human outreach and a pause on automated interventions for you. The system notices before you do.
Opt-in at the action level, never blanket. Revocable at any time; raw data deleted within thirty days of revocation. Relational actions carry your exact words. Nobody under eighteen. No physical or pharmacological intervention in version one.
Raw biometrics stay on your device, under a key you hold. Training uses de-identified, aggregated features. Your scores are visible only to you — no cross-person ranking exists anywhere in the system. No sale, no sharing, no other purpose. Raw signals: twenty-four months, rolling.
Compensation is for participation — hours of sensor coverage, completed check-ins, experiment attendance. Never for your score, never for how aligned you appear. Paying for outcomes teaches people to game the sensor, and the model learns the gaming. Pilot rate: flat $150/month at eighty percent coverage and daily check-ins.
WORKING HYPOTHESIS That group coherence is a real effect beyond the sum of its people is not assumed. It is tested, and the decision is pre-committed before the data lands. Three experiments, all pre-registered. Results publish regardless of outcome.
Twenty sessions of heart-focused breathing — ten alone, ten in a synchronized group of eight or more — randomized over ten weeks. Forty participants, powered to detect a 0.25 z-score difference. Over 0.25 at p under 0.01, replicated in a second cohort, and the group input stays as a validated feature.
Pairs in separate locations, half synchronized to the minute, half offset four hours, blinded to the condition. If phase coherence above chance replicates, the remote claim stands. If it fails, the co-located claim survives.
The dynamics model with and without the group-state input, compared on held-out prediction error. The cleanest test, because the model has no stake in the answer. Over five percent improvement, replicated, counts as evidence the group effect is real in the data.
The gate opens September 22 at 5:05 in the evening, Pacific time. Ninety days to the winter solstice — the longest night, the descent from completion into the kingdom. The engine starts with eleven hearts.
Thank you, and I love you.
Enter Miracle Academy