How DeltaBook works.
The rules of the instrument. What the engine computes, what the AI may never touch, what the record cannot do, and what a credit pays for.
Every figure is computed. None is written.
Every probability, rate, and delta on a DeltaBook screen comes from a deterministic engine. The same strategy over the same data returns the same number, every time, for everyone.
The AI is a copilot. It helps you phrase a question, assemble a strategy, and read a result. It never computes a figure, and it cannot edit one. If a value cannot be computed, the screen says so and names the reason. Nothing is ever guessed.
Only what was knowable.
The discipline is the product; the data is a slot. The first category is sports - football play-by-play back to 1999, 27 seasons. Every category that follows meets the same bar, and every row names its source and the date we knew it.
A backtest sees only what was knowable before the event. A value that did not exist on the day cannot leak into a test of that day. Where a measure starts later than your window, the screen tells you before anything runs.
Written before. Graded after. Never edited.
A tracked strategy is written down before the event and graded after it, by the platform, automatically. It cannot be changed afterwards.
If a source was wrong, the correction is a new row with its reason. The original stays visible. Nothing is overwritten, by anyone, including us.
What a credit pays for.
Talking to the AI is free. Scoping a question is free. Uploading your data is free. A job spends a credit, and only after you confirm it. There are two kinds of job, and they are never the same thing: running a strategy test, and cleaning a data upload.
The meter is a counter, not a salesman. Anything already computed is free to read, always: a zero balance never hides your results, your calibration, or your record.
A framework, not a list.
One framework classifies every metric, in any domain. It is the product; each domain's library is the framework filled in. Every indicator carries four things - and the fourth is the one nobody else prints: its class, its evidence tier, how far back it reaches, and what it does not tell you.
Holds across eras and out of sample. Weight it with confidence.
Real, but noisier or narrower. Useful, not decisive.
Once worked; the edge faded as the market learned it. Shown, labelled.
Widely believed, does not survive the data. Shipped at weight zero, as teaching.
A weightable hypothesis. This is the material you build a strategy from.
A fact about the situation. It enters the engine as input, never as a weighted opinion.
A thin-sample effect. Allowed, never editorialised; calibration handles it.
An unmeasurable narrative. Refused, with a measurable proxy offered instead.
The same framework fills a library per domain. In business it deploys as six role lenses over one engine - a CFO, CMO, CRO, COO, product, and strategy analyst reading the same evidence.
| Indicator | Lens | Class | Tier | What it does not tell you |
|---|---|---|---|---|
| Gross margin | CFO | Belief | Strong | Blended margin can rise while every product's falls - mix shift hides it. |
| Operating cash flow | CFO | Belief | Strong | Stretched payables and factored receivables inflate it - read it with the DSO/DPO trend. |
| CAC (cohort) | CMO | Belief | Strong | Blended CAC is folklore when acquisition lags; organic customers flatter every paid channel. |
| ROAS | CMO | Belief | Weak - decayed | Attributed revenue includes buyers who would have bought anyway; only a holdout shows lift. |
| Net revenue retention | CRO | Belief | Strong | A blended NRR mixes customer vintages until it answers no question - read it by cohort. |
| Win rate (stage-defined) | CRO | Belief | Strong | Retro-editing a stage is the gaming vector - it computes from stage-entry timestamps, not labels. |
| Capacity utilization | COO | Belief | Moderate | Optimal is not 100% - service degrades near saturation; read it with wait time. |
| Defect / return rate | COO | Belief | Strong | Most returns are fit, not defects - reason-code first or it misdiagnoses quality. |
| Retention curve (D30) | Product | Belief | Strong | A single day-30 point hides the shape - whether the curve flattens or keeps bleeding. |
| NPS | Product | Belief | Weak - decayed | Its growth-prediction claim failed replication; it is directional sentiment, not a forecast. |
| Cohort-economics trend | Strategy | Belief | Strong | Direction beats any point value; a revenue-basis lifetime value is the canonical error. |
| Revenue concentration | Strategy | Belief | Strong | Strong growth can hide that one account carries it - the quiet killer in a book. |
| Impressions / follower counts | Vanity | Not a metric | Folklore | Unbounded and ungrounded in customers - listed, labelled, weighted zero. |
| Indicator | Factor | Class | Tier | What it does not tell you |
|---|---|---|---|---|
| Remaining lease | Value | Belief | Strong | A price-per-foot comparison across different remaining leases is basis-contaminated. |
| School proximity (1-2km) | Demand | Belief | Strong | It prices the catchment, not the home - and it is the cleanest causal read in the market. |
| Transit distance + announcement date | Event | Belief | Strong | The move is in the announcement, not the opening - stamp which date you mean. |
| Price-to-income | Affordability | Belief | Moderate | It sizes the stretch, not the timing of a turn. |
| Months of supply | Supply | Belief | Moderate | It says how tight the market is, not why - a rate cut and a rezoning move it the same. |
| Policy shock (stamp duty / LTV) | Policy | Belief | Moderate | The effect is real and it decays - an old shock is not today's shock. |
| Indicator | Class | Tier | What it does not tell you |
|---|---|---|---|
| Yield-curve slope | Belief | Moderate | It has signalled early and it has signalled late; the timing of what follows is not in it. |
| Real-time output gap | Belief | Weak - decayed | It looks predictive on revised data and collapses on the vintage that was actually knowable then. |
| Regional activity surveys | Belief | Moderate | The free-computable stand-in for the licensed national index - a region is not the nation. |
| Credit spread (from primaries) | Belief | Strong | Widening spreads price stress, not its cause and not its timing. |
| Financial-conditions index | Belief | Moderate | It composites many series into one - the composite hides which one actually moved. |
| Indicator | Class | Tier | What it does not tell you |
|---|---|---|---|
| Fundamental quality (from filings) | Belief | Moderate | The filing date is knowable-at, not the period end - reading period-end is a quarter of lookahead. |
| Accruals | Belief | Moderate | It flags earnings quality, never the moment the quality corrects. |
| A published factor (value / momentum) | Belief | Weak - decayed | Factor returns decay about a third out of sample and more after publication - the paper result is inflated. |
| Bank scoreboard (call reports) | Belief | Strong | A ratio ranks banks; it does not tell you which underwriting decision made the rank. |
| Point-in-time membership | State | Input | Backtesting today's survivors fabricates signal - delisted firms must stay in the universe. |
| Indicator | Class | Tier | What it does not tell you |
|---|---|---|---|
| Expected points added | Belief | Strong | It scores the outcome, not the process - a lucky bounce and a perfect throw land the same. |
| QB-out adjustment | State | Input | It conditions the answer as a fact; a model that does not know the starter is broken. |
| Home-field advantage | Belief | Strong, moved | It is not a constant - it shrank after 2020, and a single figure hides that. |
| Pressure rate | Belief | Moderate | Window-bound to 2018+; who owns a pressure - rusher or coverage - is not in the number. |
| Rest / short week | Belief | Weak - decayed | The market priced it in years ago; the historical edge is mostly gone. |
| Regression signal (Pythagorean gap) | Belief | Strong | It flags reversion coming, never when it lands. |
SPORTS IS THE LIBRARY THAT IS LIVE TODAY; THE OTHERS ARE AUTHORED AND STAGED, TIERS PROVISIONAL PENDING VERIFICATION. ONE FRAMEWORK, EVERY DOMAIN.
Your key, your tools.
Bring your own AI key, or use ours. An MCP server connects DeltaBook to your own AI client - it drives the same typed tools as the built-in copilot, behind the same confirmation gates. Nothing runs without your confirm, from any client.
EACH INTERFACE IS DOCUMENTED HERE AS IT SHIPS. NOTHING ON THIS PAGE IS CLAIMED AHEAD OF THE BUILD.