SIMULATION LAYER 01 / 14
GodEngine · Business Intelligence Simulator
A company hands over its uncurated raw business data. We arrange it, then show two things across seven sections — recorded, which is arithmetic on documents you supplied, and projected, which is the next three quarters as a range.
what we actually parse:a Tally trial balance, a Zoho P&L, a bank statement workbook, a CRM pipeline export
We are asking for messy exports, not for a logo on a wall.
Inventory cover
— daysawaiting your data
Closing inventory over cost of goods sold, in days of the period. The one supply-chain panel a trial balance genuinely supports.
supply.inventory_coverprojection: range, 3 qtrs
SIMULATION LAYER 02 / 14
01 — The rule everything else follows from
Filling that gap with an industry average or a modelled guess would look identical to a real reading, and there would be no way to tell them apart later. A dashboard that is 80% real and 20% plausible is not 80% useful. It is unusable, because you cannot tell which fifth is which.
So a locked panel says what document would unlock it, and stays empty until that document arrives.
RAW EXPORT
NORMALISED
TRACEABLE PANEL
PROJECTED RANGE
Seven KPIs are computable from an accounting export alone. The rest of the catalogue names the document it needs.
bi/compute/kpis.py · computable_ids()
SIMULATION LAYER 03 / 14
02 — Two kinds of number, never mixed
Recorded · your books
A recorded figure is a formula over measures extracted from documents you supplied. It carries the formula, the measures, the source file and the row. If you disagree with it, you can open it and find out where we read it wrong.
Projected · modelled
Three quarters forward, as p10 / p50 / p90. Every projection names its tier, its decoder, how many points it trained on, how many it held back, and its fit on that holdout. When there is not enough history, the panel locks instead of guessing.
Charts, legends, pills and bars all derive from that one distinction, which is why the design tokens are named recorded and projected rather than for their hue.
SIMULATION LAYER 04 / 14
03 — Surfaces
Company
The handful of figures a founder checks first, and the signals ranked against them.
Core
The safe section: a trial balance alone supports it. Revenue, margin, cash, working capital.
Core
Pipeline and conversion, which need a CRM export rather than a ledger.
Core
Spend is in the ledger. Attribution is not, and we do not pretend otherwise.
Delivery
Utilisation, rework. Capacity is a fact about plant and shifts; no export states it.
Delivery
Inventory cover computes today. Vendor concentration is the closest locked panel in the product.
Governance
Calendar arithmetic over due dates and filed dates. Both are document facts.
Section depth follows data availability. Finance is safe. Supply chain needs a purchase-order register and a goods-receipt log; Compliance needs a filing calendar.
bi/compute/kpis.py · SECTION_GROUPS
SIMULATION LAYER 05 / 14
04 — The data plane
Nothing is thrown away between them. Each stage records what it did, so the trace on any figure can walk back through all eight and show you the cell.
Accounting-grade extraction is ours, deliberately. The engine’s multimodal ingester is right for classification and for non-tabular documents — it caps sheets and rows and keeps numeric cells only, which is correct for a digest and fatal for a general ledger.
SIMULATION LAYER 06 / 14
05 — What a locked panel says
Each panel that cannot be computed states why, and names the document that would unlock it. Send that document and the panel fills; send nothing and it stays honest.
| Panel | State | Why, or what unlocks it |
|---|---|---|
Inventory coversupply.inventory_cover | computes | Closing inventory over cost of goods sold, in days of the period. |
Vendor concentrationsupply.vendor_concentration | locked | Needs purchases by vendor — a purchase register with a vendor column, or creditor-wise totals. |
Lead timesupply.lead_time | locked | Order date to receipt date. Neither is in a ledger. |
Cost variance to plansupply.cost_variance_to_plan | locked | A variance needs a plan. The ledger records what was spent, never what was budgeted. |
Utilisationoperations.utilisation | locked | Output against available capacity. Capacity is a fact about plant and shifts; no accounting export states it. |
Filings on timecompliance.filings_on_time | locked | Calendar arithmetic over due dates and filed dates. Unlocked by a filing calendar or GST acknowledgements. |
Copy in this table is the product’s own not-computable text · bi/compute/kpis.py
SIMULATION LAYER 07 / 14
06 — The forecast ladder
Six tiers, tried in order. The panel is labelled with the one that actually produced it, and a projection is never labelled TimesFM unless it came from TimesFM.
With the model weights absent, the loader returns a stub. The gate checks the decoder name rather than trusting the call, so a stub result falls through to classical models instead of being served as a foundation-model forecast.
SIMULATION LAYER 08 / 14
07 — Trace
Not a citation appended to a claim. A record of the operation, in order, each stage carrying its own rows and the source id behind each one.
Stage 1: Recorded
The file, the sheet, the row. What you sent, identified.
Stage 2: Recorded
The cell as written, and the value we read out of it.
Stage 3: Computed
The formula, with each measure substituted.
Stage 4: Computed
Holdout, fit, and whether the history was sufficient at all.
Stage 5: Projected
Tier, decoder, quantile source, and the band it produced.
Answers are checked the same way. Before a written answer is shown, every number in it is verified against the grounding it was built from. The check is substitution-aware — it catches a figure drifting, not just a missing citation. On failure the pipeline regenerates once, then abstains.
Abstaining is a supported outcome. It is the only alternative to a confident wrong number, and the product prefers it.
bi/engine_bridge.py · check_figures()
SIMULATION LAYER 09 / 14
08 — What real exports look like
These shapes cannot be invented convincingly. Each one written from imagination is a test that passes for the wrong reason, so we ask for the real thing.
A1:G6 → junk
Six rows of company name, address and report title above the actual column headers.
(1,234)
Read as text by anything that guesses dtypes, which silently loses the sign.
1,23,456
Lakh and crore separators. A parser tuned for thousands gets this wrong by orders of magnitude.
Σ mid-sheet
Group totals interleaved with rows, ready to be double-counted.
₹ lakh
"Figures in ₹ lakh" written once, in prose, three rows above the numbers it governs.
A Tally trial balance, a Zoho or Busy P&L, a bank statement workbook, a CRM pipeline export. Anonymised is fine — the shape is what matters, not the values.
SIMULATION LAYER 10 / 14
09 — How we talk about ourselves
The same discipline the product applies to your figures, applied to our own statements. An untagged claim is a defect. A number with no tag does not ship.
Which is why this website has no customer count, no revenue figure, no testimonial and no logo wall. We do not have them yet, and a plausible-looking placeholder is the one thing we will not write.
SIMULATION LAYER 11 / 14
10 — Third-party validation, all of it
Accepted. Documentation available.
[founder]
Founders Hub. Accepted. Documentation available.
[founder]
Accepted. Documentation available.
[founder]
Three programmes, not four. An earlier marketing page claimed a fourth acceptance with a dollar figure attached; it is unattested, so it does not appear here.
More on how the company describes itselfSIMULATION LAYER 12 / 14
11 — Where this actually is
The repository boots, the vendored engine is reachable across all eight bridged capabilities, the first migration applies and reverses, the frontend builds, and every CI gate is green and blocking.
The landing surface renders the live readiness payload, so opening it proves the whole chain rather than showing a shell — and tells you which link is broken when one is.
Nothing can serve a projection by accident: the projection, temporal, ensemble and Ask flags all default off.
SIMULATION LAYER 13 / 14
12 — Questions
Both, kept visibly apart. Recorded figures are reporting: arithmetic on your documents. Projected figures are simulation: a range over the next three quarters, labelled with the method that produced it. They are never blended into one number.
An accounting export is enough to make Finance real and to compute inventory cover. Each further section names the document it needs — a purchase register, a filing calendar, a CRM export — and shows nothing until it has it.
No. Phase 0 is complete and the dashboard, wizard, projections, Ask bar and connectors are not built. We are looking for design partners with messy exports, not for users.
No. The connector layer verifies that a credential cannot write and refuses to save it if it can. Read-only is enforced at the point the credential is stored, not promised in documentation.
It abstains. Every figure in a written answer is checked against its grounding before you see it; on failure the answer is regenerated once, then withheld.
SIMULATION LAYER 14 / 14
Design partners
Anonymised is fine. We want the shapes — the header on row seven, the mid-sheet subtotals, the scale note buried in a title cell. In exchange you get the first working sections built against your own data.