The prediction desk

Every ticker page runs the same pipeline. This page explains each stage and the command behind it.

1 · Forecasters

Independent views, each from different information:

Command	View	Method
ot decide	rules engine	CALL/PUT/NO-ACTION + a range plan
ot quant	logistic P(up) + range cone	keyless logistic regression, OOS-gated
ot forecast	TimesFM 2.5 quantile cone	foundation model (opt-in install)
ot options	dealer gamma (GEX)	signed gamma × OI per strike, from CBOE chains
ot poly	crowd odds	Polymarket prices for the macro event gate
ot hl	perp funding & OI	Hyperliquid — the crypto leverage bill
ot whales	on-chain flow	labeled-wallet ETH balances + deltas

The point is independence: different math, different data. A cone from statistics and a wall from dealer hedging that land on the same price is tradeable structure neither shows alone.

2 · Fusion

• Consensus strip — one chip per analyst; the verdict flips to STAND ASIDE the moment two disagree. Disagreement is information.
• Confluence ladder — all the levels the desk emits, on one price axis. Rows named by 2+ methods (×2) are the only ones that matter.

3 · The debate

ot debate NVDA           # 3 LLM calls on a frozen evidence pack

Three calls run over a deterministic evidence pack:

1. Bull — the strongest honest long case.
2. Bear — the strongest short/avoid case, and it must directly attack the bull's strongest point.
3. Judge — weighs both against the evidence and commits: a 5-tier verdict with confidence, entry, a mandatory invalidation level, and a time stop.

By default the bull and bear run on different engines (perspective diversity, Condorcet-style); you can also run all three roles on one engine. Every verdict auto-journals for grading.

:::info Deterministic SOP The evidence pack is built and frozen by scripts before any model runs. The models read; they never fetch or act. Same inputs → same pack → reproducible. :::

4 · The learning loop

ot reflect grade         # grade calls that have ripened
ot reflect stats         # hit-rate by action / grade / market
ot reflect lessons       # the lessons injected into the next judge

Every committed call is later scored against what actually happened — right or wrong, return vs SPY (alpha), worst adverse excursion, whether the invalidation broke — and the lessons feed back into future judge prompts. See Methodology for exactly when and how a call is graded.