Provenance · Modeled bridge

The bridge: how we filled a 19-month hole without faking it

2026-07 · Volar

There is a hole in our BTC archive. It runs from October 2024 to April 2026 — the space between the dense minute-level archive we serve (2021-06 → 2024-09) and the live capture we started running in April 2026.

For a while, that hole had data in it. It was vendor-sourced, thinly sampled, and it sat in our files looking denser than it was. When we audited our own provenance in June, we deleted it — all of it — and let the hole be a hole. Our coverage claims shrank that day. We think that's what a data vendor is supposed to do.

This post is about what we did next, including the part where our own validation system refused our first attempt.

The rules we set before we started

We decided to bridge the gap with a model: volatility surfaces fitted to real Deribit trade prints from the gap window, arbitrage-verified, with modeled bid/ask spreads calibrated on real quote data. Reconstruction is honest work only if it's graded honestly, so before scoring a single day we froze the exam:

• Pre-registered protocol. Eighteen real reference days, split alternately into training and holdout before any evaluation. The holdout is scored once per release. Every bar, bucket edge, and hyperparameter was committed to the repository before the first run.
• A hard gate. If the liquid core of the chain fails its accuracy bars, nothing ships. No partial credit.
• Published results either way. The full error tables go on our methodology page whether they flatter us or not.

The part where we failed

Our marks were accurate — median error about 1.1 volatility points in the liquid core — but one pre-registered bar demanded more: that modeled marks sit inside the real bid/ask band 75% of the time. Indistinguishable from real, at the market's own precision.

We then instrumented everything and measured why. Real liquid-core bands run about 0.6 vol points half-width. Marks reconstructed from trade prints — even seconds-fresh ones — miss bands that tight most of the time, and every improvement hypothesis we tested against that bar died by measurement. The honest conclusion: indistinguishable from real is not achievable from public trade prints. Not by us, not by anyone. A model that claimed otherwise would be lying somewhere.

The re-aim

So we changed the claim — in public, with a version number.

Protocol v2 changes exactly one thing: band containment became a published metric instead of a ship gate. Every error bar, the recall requirement, the spread-width bar, and the holdout discipline stayed identical. The v1 verdict remains on our methodology page as history, and one detail matters to us: our model still failed the v2 bar on the day we proposed it. The new bar wasn't drawn around what we had. We spent the engineering to pass it afterward — a rebuilt spread-width model and two measured accuracy fixes, iterated only on training days.

Then we ran the holdout — nine days the model had never seen — once.

The numbers we're shipping on

Two disclosures ride with every number. First, our reference days are all month-starts; on mid-month days the median error runs about 1.4× higher — we measured that too, and any accuracy claim we make carries it. Second, in four of nine wing regions (short-dated and far out-of-the-money) the model missed its bars — those rows ship with marks but no modeled quotes at all. A disclosed partial hole, not a papered-over one.

What this means for you

• Nothing changes unless you opt in. Bulk downloads exclude modeled files by default; pass include_modeled=true to receive them.
• Every modeled row says so — source=modeled_surface, on every row, filterable client-side.
• It never counts. Our coverage claims, on the landing page and everywhere else, describe real data only. The bridge is a bonus layer.
• It costs nothing. Included with full-archive access, not priced, never will be a reason for a price change.
• The receipts are public. The methodology page renders its error tables directly from the committed validation artifact — the same file our upload tooling reads, which refuses to ship anything without a passing verdict in it.

The gap is still a gap — we want to be clear about that. What sits in it now is the most accurate reconstruction we know how to build, graded by an exam written before the student saw it, with the one failing grade we got along the way framed on the wall. That's the deal. We think it's a better deal than pretending.

Questions about the methodology? It's all at app.volardata.com/methodology, or write to [email protected].