The honest answer

How COYL knows you.

Today, about 80% of what COYL knows comes from what you tell us. The audit, the commitments, the slip logs, the interrupt feedback — all you. That is the cold start. Here is the arc that changes it, and why the input dependence is precisely the moat no LLM can synthesize.

What we know

Two columns. Wildly different sizes.

Passive · no input

Time of day

Your local clock. Day, hour, minute. The cheapest signal we have — and it carries half the pattern.

Passive · no input

Day of week

Monday-Resetters do not behave like 9PM-Negotiators. The week shape matters.

Passive · no input

HRV / stress proxies

If you connect Apple Health. We never write — we only read HRV, sleep, steps. Stress proxies, not biometrics.

Passive · no input

Timezone

Travel collapses windows. The 9 PM Negotiator pattern in NYC fires at 6 PM in LA. We adjust.

Passive · no input

Browser tab opens

Only if you install the extension. Counts opens per domain, never reads page content.

Passive · no input

Location (optional)

Kitchen vs office vs car. Permission-gated. You can run COYL forever without granting this.

Active · you tell us

Your archetype

From the 60-second audit. The 6 families. Cold start: this is the only signal we have for the first week.

Active · you tell us

Your danger windows

After the audit we auto-create 2-3 inferred windows. You can edit, add, archive. We mark inferred vs user-set.

Active · you tell us

Your commitments

"No food after 9 PM." "Three hours of deep work before noon." The rules you choose to live by.

Active · you tell us

Your slips

You have to confess them. One-tap slip logging gets this down to a single button. We need the timestamp + context.

Active · you tell us

Your excuses

The story your psyche told you right before the slip. We classify into 8 categories. The taxonomy is the moat.

Active · you tell us

Interrupt feedback

"Caught me." "Wrong time." "Too soft." Two-tap rating after each interrupt. Trains the model on your real-time judgment.

The active column is longer on purpose. We are honest about it.

The arc

80% active today.
10% active in year two.

Active · you tell usPassive · we infer
95%5%

Projected trajectory based on early cohort behavior. We will publish actuals as the dataset grows.

At this stop

Week 1

You teach us everything. The audit gives us your archetype. The first inferred windows are scaffolding, not predictions. The interrupt copy is family-level, not you-level.

Tuesday 9:43 PM, kitchen. We fire because the 9PM-Negotiator window says we should. Not because we know YOU yet.
One minute, layered

What COYL sees at 9:47 PM.

An illustrative timeline. Example user, month 3 of using COYL — not a real session. Tap to reveal each layer as it lands in the model. The order is the order events fire, not what you would report after the fact.

9:31:04 PM

passive

HRV drops 18% — Apple Health webhook

1 of 9
The dependence is the moat

The thing no LLM can synthesize.

ChatGPT can describe what a 9PM Negotiator might do. Claude can predict the average human's evening drift. Neither knows what you did last Tuesday at 9:47 PM, what story you told yourself afterward, and whether the interrupt landed.

That ground-truth, longitudinal, pre-conscious behavioral dataset is the thing. Synthesizing it requires watching one person across months — across slips, recoveries, excuses, and the moments when the interrupt fired and you tagged "caught me." No foundation model has that dataset. We do, on you, because you built it with us.

Once it exists, it exports. A Behavioral Context Object any LLM can consume. The platform play.

What COYL knows that Claude doesn't

The four fields. Specific, not asserted.

“Behavioral dataset that no LLM can synthesize” is a phrase. Below is the argument — what is in the dataset, and why a foundation model without longitudinal access to one specific human cannot derive it.

Slip taxonomy

What it is

8 excuse categories — DELAY, REWARD, MINIMIZATION, COLLAPSE, EXHAUSTION, EXCEPTION, COMPENSATION, SOCIAL_PRESSURE — classified from your actual in-session language at the moment of the slip.

Why Claude can't synthesize it

Not in HealthKit. Not in Google Calendar. Not in any LLM chat log. Requires moment-of-slip capture that only the interrupt surface provides.

Recovery curve

What it is

The time and shape between a slip and a recovery — including same-night recovery, streak resets, and return-to-baseline events. Not just binary slip/no-slip.

Why Claude can't synthesize it

Wearables capture the slip (HRV spike, location). They do not encode the arc back. Only a product that lives through the full cycle can build this.

Danger windows

What it is

A day-of-week × hour-of-day × context histogram, computed per user from the last 30 days of slip records. Your 9:43 PM Tuesday — not the average person’s late-night.

Why Claude can't synthesize it

Foundation-lab models describe the 9PM Negotiator archetype. They cannot predict your specific window without longitudinal access to your behavioral history.

Longitudinal sequences

What it is

60+ day sequences of slip → recovery → re-relapse → re-recovery. The pattern, not a single event.

Why Claude can't synthesize it

One slip is noise. The pattern requires the user to have stayed in product. Retention is the moat — and the moat compounds with time-in-product, not with sign-ups.

A foundation-lab model trained on the open web can describe the 9PM Negotiator archetype. It cannot predict your specific window with the accuracy a model trained on your last 30 days of slips and recoveries can. That gap compounds with time-in-product. Retention is the moat.

Where it hurts

Three risks. All shipping fixes.

AI for the moment before behavior happens.
The recurring anchor