The engine that curates today and anchors tomorrow

The conviction score doesn’t directly produce a TEB forecast — it feeds into the v2.1 Forecast Engine, which combines cohort baselines, Bayesian shrinkage, and 11 of 15 bounded adjustments that map to conviction signals. The full architecture lives at /spec/forecast-engine.

The chain: 11 conviction signals → composite conviction score → 11 adjustment magnitudes (one per signal) → composition damping + hard caps → final forecast bands. Conviction shapes how strongly the engine disagrees with the cohort prior, bounded by structural caps so a single high-conviction issuer can’t escape cohort gravity entirely.

Conviction is not a price. The market prices tokens. Conviction is the curation filter that decides who lists in the first place, and the forecast anchor that drives the mid-band estimate of future TEB. Those two roles — gating and anchoring — share the same 11-signal model because the same features predict both qualification and trajectory.

Why it matters —Score and price are correlated, not equivalent. An 85-conviction issuer with a low projected TEB still prices modestly. A 72-conviction issuer in a booming sector with an enormous projected TEB prices richly. Conviction informs the forecast; the forecast and the market set the price.

Each signal is normalized to $[0, 100]$ before weighting. The inner weights $w_{b,i}$ and $w_{f,j}$ shape the contribution of each signal within its group and, in v1, are hand-tuned from first principles. See Calibration for the roadmap.

The 60/40 outer split isn’t arbitrary. The backward block carries the load because realized signals are orders of magnitude more predictive than stated-intent signals — a standard finding in labor economics. The 40% forward weight is where early-career founders can still get fair conviction before a trailing history exists.

Backward signals capture what actually happened. These dominate conviction once 3+ years of TEB history exists (roughly: after the first full Direct Listing review, or after a covenant has passed through two quarterly reporting cycles).

Each signal is computed from reporting inputs the issuer has already attested to on-platform. No external API dependencies in the hot path — the calculation is deterministic and reproducible from a frozen snapshot of the issuer’s reports.

Forward signals capture what’s coming. These dominate conviction for covenants at list time (where no on-platform TEB history exists yet) and balance the backward block for Direct Listings.

Why it matters —F₅ carries almost no weight on purpose. We ask the question because the answer is informative to us— it tells us what the issuer expects, which helps calibrate their forecast bands — but we don’t let their optimism price the token. That discipline is what separates this from a pitch-deck underwriting model.

Conviction doesn’t just gate listing; it directly shapes how wide the engine’s confidence-interval bands sit around the mid forecast. Wider CI bands → wider V_HC bands → a lower reserve floor → more room for market price discovery above the floor.

The CI band-width as a function of the conviction score:

A floor of 15% (high-conviction issuers can’t go below ±15% — there’s irreducible forecast uncertainty about a human) and a ceiling of 60% (low-conviction issuers cap at ±60% — beyond that the auction is in market-discovery mode).

The chain — signals to reserve in five arrows

1. 1.11 signals → composite conviction. Backward block (60%) + forward block (40%), each averaged then clipped to [0, 100].
2. 2.Composite → forecast bands. Mid-band TEB path is the engine’s expected. Conviction modifies the multiplier on standard deviation.
3. 3.Forecast bands → V_HC bands. Integrate each band’s TEB path with discount rate to get $V_{\text{HC}}^{\text{low/mid/high}}$.
4. 4.V_HC bands → MCap bands. Multiply by $e_{\text{eff}}$ to get market-cap bands.
5. 5.MCap bands → reserve. $P_{\text{reserve}} = 0.80 \cdot V_{\text{HC}}^{\text{low}} \cdot e_{\text{eff}} / N$. Published, methodology disclosed before bidding opens.

Every weight in the model is either seeded (v1 today) or calibrated (v2 onward, as data accrues). The calibration harness logs three things for every listing from day one:

• 1.The full signal vector at listing time ($B_1 \ldots B_6$, $F_1 \ldots F_5$) frozen into the listing’s permanent record.
• 2.The forecast path at listing time — low, mid, high over 75 years.
• 3.The realized TEB at every subsequent quarterly reporting cycle.

Every quarter, the harness runs the following regression across the full cohort of listings old enough to have accumulated $k$ quarters of reporting:

The fitted coefficients become the updated inner weights $w_{b,i}$ and $w_{f,j}$ for the next cohort. Regularization (ridge or elastic net) keeps individual weights from swinging wildly on small sample sizes. After 2 years, the confidence interval on every weight is tight enough to publish; after 5 years, every seeded number in v1 will have been replaced by an empirical equivalent.

Why it matters —This is how every valuation engine in financial history has bootstrapped. Black-Scholes ran for a decade on theoretical volatility before the implied-volatility surface became the primary input. Moody’s rated municipal bonds on judgment for 80 years before the first quantitative models replaced it. The difference is that Preflop architects the logging from day one — so we don’t have to reconstruct a cohort retroactively when the data starts to matter.

Most products have a UI moat. Preflop has a data moat. They aren’t the same thing.

Today the conviction engine does one job: curation. An applicant with a conviction score below 60 doesn’t list; above it, they get onto the platform at a market-clearing price. That job is valuable but not defensible — any competitor with a good review process does it.

Tomorrow, the engine does a second job: anchoring. Each additional listing generates three data points — the forecast at list time, the realized TEB at each reporting cycle, and the eventual outcome of the covenant. Over time this corpus does something that has never existed:

The lock-in is structural. A competitor showing up in Year 3 has to underwrite without calibration data — which means they either (a) under-price risk and lose money, or (b) over-price risk and get adverse-selected. Preflop’s cohort gives us pricing advantage even against a hypothetically better-capitalized entrant.

Honest labelling is the credibility move. Here’s the split as of v1:

v1 → v2 was the jump from forward-only to hybrid. The next two versions are planned:

Vision —Every other valuation engine in finance hit a ceiling when its dataset stopped growing — Bloomberg’s credit pricing, S&P’s equity research, Moody’s ratings. Preflop doesn’t hit that ceiling because every listing on the platform is a training point for every subsequent listing. The data is the product; the product gets smarter the more it’s used. We cannot be overtaken by a better UI.