Why the Model Never Touches the Numbers

The honest version of the objection

If you have spent any time asking a general-purpose model to read a prospectus, you already know the failure mode. Ask it for a share count and it will give you a share count. It will be confident, it will be formatted correctly, and some meaningful fraction of the time it will be wrong — a figure from an earlier amendment, a pre-split number, a total confused with a base offering, or something with no origin at all. The output is fluent whether or not it is true, and fluency is exactly the wrong signal to trust when the number determines what you pay.

So the reasonable question to ask us is: if Bellipo uses models too, why is this any different? The answer is not that we found a better model. It is that the model is not allowed near the part that matters.

The division of labor

A Bellipo Dossier is assembled from two very different systems, and we keep them strictly separated.

• The Spine — Deterministic Filings, forms, accession numbers, prices, ranges, share counts, dates, and provenance stamps. This is ordinary code reading structured sources. It has no creativity and cannot invent. If the value is not in a source, the field stays empty.
• The Prose — Model-Assisted Explanation. What a going-concern paragraph means in plain English, why a secondary-heavy deal has a different character than a primary one, which of ninety risk factors are boilerplate and which are load-bearing. Judgment and language, reviewed before publication.

Everyone in this category will tell you they do something like this. The part worth checking is what happens when the two disagree — when the prose wants to state a number the spine cannot support.

The number lint

Every figure destined for a public surface is checked against the underlying record before it ships. The check extracts every dollar amount, percentage, year, and date from the draft text and requires each one to appear on the row it references. A token with no match is not softened and not flagged for later review. It is blocked, and the publish does not happen.

The model can write the sentence. It cannot supply the number inside it.

This is a deliberately blunt instrument, and it has real costs — it occasionally rejects a perfectly accurate sentence because a figure was phrased differently than the record stores it. We accept the friction. A system that fails toward silence is the only kind worth pointing at your money.

Provenance has tiers, and the tiers have teeth

A price is only as good as the document behind it. Bellipo grades every set of terms by what kind of evidence supports it, and the grade determines what the system is permitted to do.

• EDGAR — Highest Terms drawn from a pricing filing: a 424B prospectus or a notice of effectiveness. This is terms of record. It can trigger an alert.
• Issuer Release — Accepted The company's own release, distributed through an approved wire or published on the issuer's own investor-relations host. It can trigger an alert.
• Vendor and Derived — Never Exchange calendars, aggregators, and press coverage. Useful for context, never sufficient for a price. These can never trigger an alert, regardless of how confident they sound.

One distinction inside that ladder does most of the work: the difference between issuance and reporting. A wire distributes the company's own words. A newsroom publishes a report about them. The first is evidence of a price; the second is evidence that someone believes there is a price. Major financial outlets are permanently classified as reporting in our system — not because their journalism is poor, but because a report is not a document.

There is a second gate underneath the first. A filing being from EDGAR is not enough; we also check which form it is. A registration amendment carries a marketed range, not final terms. Treating one as the other is precisely how a preliminary band becomes an alerted price, so registration paper is refused even though it comes from the same source as the paper we accept.

What this costs us in practice

On July 30, one of the deals we track priced. Aggregator sites had a number up before we did, and our alert did not fire, because at that moment no pricing filing existed and we had not yet verified an issuer release. The deal sat in a holding state while a competitor's page showed a price.

We could have shipped the aggregator's figure. The system would not let us, and that was the correct outcome even though it felt like losing. The alert went out once the release was verified against the approved-source list. A system that is right after everyone else is a system you can act on; a system that is first and occasionally fabricated is not a system at all.

The score is arithmetic, not opinion

The Bellipo Demand Index is sometimes assumed to be a model output. It is not. It is a weighted calculation over a fixed set of factors — pricing signal, underwriter quality, sector momentum, and market regime — with published weights and inputs drawn from the deterministic spine. The same inputs produce the same score every time, and each factor carries a reason string explaining what drove it. At pricing, the score is frozen and stored with its factor breakdown, so it cannot be quietly revised after the fact to look prescient.

You are welcome to disagree with the weights. That is a methodology argument, which is the kind of argument we want to have. What you cannot get is a different number from the same facts. The full BDI methodology is public.

Recency is a gate, not a courtesy

Pre-filing research has a particular decay problem: a valuation reported nine months ago is often still the most-cited figure on the internet long after a newer round has superseded it. A citation being real is not the same as a citation being current. Our drafting gate treats a dated source that has been overtaken by newer reporting as a failure, not as an acceptable approximation, and volatile facts are re-verified against current reporting at the moment of writing rather than inherited from the last draft.

Where we stop short

We are not audited. No outside firm has reviewed our extraction accuracy, and we are not going to imply otherwise by dressing the claim up in a badge. What we have built is something narrower and, we would argue, more useful at this stage: a pipeline whose factual claims are traceable to a document you can open yourself.

Every published dossier carries a Provenance Receipt listing its material figures and the source class behind each one. Every dossier section is versioned. Corrections are filed publicly rather than edited away in silence. The standing invitation is to check us — and to tell us when we are wrong, on the corrections page, which exists precisely because we expect to need it.

Three things to take from this

• 01 Fluency is not accuracy. A general model asked to read a prospectus will always answer. The question is whether anything stops it when the answer is invented.
• 02 The gate is the product. Not the model, not the writing. What Bellipo actually sells is a pipeline that refuses to publish an unsupported number.
• 03 Verifiable beats authoritative. Do not take our word for it. Open the receipt and follow the citation to the filing.

Read the full standard

The complete integrity standard — every provenance tier, the correction policy, what is machine-extracted versus written, and what we explicitly do not claim — lives on one page.