Ask any LLM-stock-research product whether their verdicts have been right. You'll get evasion. They'll show you their methodology, their backtests on historical data, maybe a confidence number. What they won't show you is a public ledger of: here's what we called on date X, here's what the stock did between then and now, here's whether we were directionally right. Because nobody in the sector keeps that ledger. We just started.
The asymmetry is this: bullish or bearish verdicts are cheap to publish. The day after they're issued, attention moves to the next ticker, the next chip, the next opinion. Three months later, when the price has actually done what it was going to do, there's no public structure to walk back to that verdict and say "and here's how it went." So the entire industry runs on forward-looking confidence with no backward-looking accountability.
This isn't laziness. It's hard. To honestly score a verdict three months later, you need to have stored everything the model saw at the moment of the call — not just the chip, but the world the chip was a response to. Most tools throw away the world after rendering the chip.
This week we shipped the storage layer that closes the loop. The conceptual model is three steps:
Every verdict the council emits — Quorum Lite, Quorum Pro, Quorum Max, or Quorum Ultra-Max — now persists a complete state-of-world snapshot alongside the chip. Five categories per row:
All five together are reconstructable into a single sentence: this is what the council saw, on this date, when it said this chip.
T2 isn't a price snapshot grafted onto T1. It's a full fresh record — same five categories, same ticker, just later in time. Because by then a new council run has produced a fresh verdict for the same name, and that verdict carries its own inputs + reasoning + entry price. T2 is just whichever council run lands in the horizon window. So we get an unbroken chain of state-of-world snapshots per ticker, not just two endpoints with a hole between them.
The evaluation step joins a T1 verdict to its T2 counterpart and asks: did the chip direction match the price direction?
HOLD / WAIT is direction-correct if the price did not move much in either direction — the chip is itself a prediction of low movement and should be scored on that.The noise floor scales with the horizon: 3% at 30 days, 5% at 90 days, 8% at 180 days, 10% at the full year. Short horizons are tighter because daily noise is tighter; long horizons get more drift latitude because a year's worth of market noise is wider.
Direction-correctness is the primary metric — was the chip pointing the right way? Magnitude (the realized return percentage, the alpha vs S&P 500, the alpha vs equal-weight benchmark) is the secondary metric. Being directionally right with a 22% return when the chip implied +20% is great. Being directionally right with a 2% return when the chip implied +20% means the model called the right direction but missed the magnitude. Both are publishable; they tell different stories.
The /admin/scoreboard surface comes alive once we've accumulated N≥30 closed verdicts per tier × chip × horizon. Per-tier direction-accuracy by chip class. Per-seat track record so we can see which seat's reasoning consistently survives. Accuracy by horizon so we can see whether the council's 30-day calls are better or worse than its 180-day calls. All internal at first — six months isn't enough to be public.
A public /track-record route. For any historical date, the verdict the council issued, the price that day, the price today, the realized return, the alpha vs S&P 500 and an equal-weight benchmark, and the autogenerated narrative line: "Our Pro+ council issued an Upside Signal on NVDA on 2026-03-15 at \$X. Today: \$Y. +Z% vs SPY +W%." Same narrative for misses, with the named falsifier shown — so you can see whether the council's own stated invalidation event had actually fired before the miss landed. Drill down on any verdict and you see the full state-of-world snapshot: every input, every dissent, every macro flag the council priced in.
That will be the first publicly verifiable LLM-stock-research track record in the sector. Backed by the storage layer that started populating today.
A verdict without a track record is an opinion. A verdict with twelve months of stored state-of-world snapshots and direction-scored history is the beginning of evidence. The whole point of the council architecture — adversarial seats, preserved dissent, named falsifiers — is that it should be auditable not just in the moment but six and twelve months out. The storage layer is what makes the audit actually possible. We've been promising adversarial honesty in the present; now we're promising it in the past too.
Every council run you trigger — Lite (free, 20/day), Pro, or Pro+ — now generates a stored prediction row. From today forward. There's nothing to opt into; the storage is automatic and the year-end public track record is built from the corpus of every run that the system has done. The more verdicts run, the more data the track record has to draw from. Claim a free seat → See tiers →