[ METHODOLOGY ]

how every verdict on this site is generated, gated, and validated

1. AI Jury — multi-model concurrent inference

News articles from a curated set of macro and crypto sources are pulled in real time. For every article that passes the relevance gate, the pipeline does the following in parallel:

• Multiple independent language models read the headline + first body paragraphs and emit a directional call (YES / NO) on every relevant Polymarket market they can match, plus a calibrated confidence score.
• A statistical baseline — a non-LLM model trained on past news → market-move pairs — emits its own independent call. This baseline is the dissenting voice when the LLMs are correlated.
• A majority verdict is computed: the published direction is the one that wins across the jury. When the LLMs and the statistical baseline disagree, the verdict is flagged verdict_type=override instead of unanimous or majority — the disagreement is preserved, not hidden.

Every verdict carries:

• The article that triggered it (title, source, URL)
• The Polymarket market it was mapped to (full question + market YES price at the moment of judgement)
• A direction (YES / NO) and a confidence in [0, 1]
• The individual juror calls (so you can see when the jury split)
• An immutable verdict_id (e.g. alpha1-sig-3509) and timestamp

Verdicts that fail the relevance / market-quality filter are still recorded but flagged shadow and excluded from accuracy stats. This avoids the easy trick of computing winrate only on the easy cases.

2. Counterfactual 24h calibration

This is the part most "AI prediction" sites skip. A verdict alone proves nothing — accuracy proves something.

For every published verdict, exactly 24 hours later (or at market resolution, whichever comes first), the system computes:

• The market YES price at t+24h vs the price at t=verdict
• Whether the price moved in the direction the jury called
• The counterfactual P&L of a unit-sized bet placed at t=verdict and closed at t+24h

Result is written into jury_outcomes with verdict_was_correct ∈ {true, false}. The 30-day and lifetime accuracy you see at the top of the AI Jury card is a live SUM(verdict_was_correct) / COUNT(*) over that table — not a curated case study.

Wrong calls stay published. They keep their original timestamp, their original confidence, and they show up in the accuracy denominator forever.

3. CEX 2nd-wave shortlist

A separate scanner re-evaluates the universe of liquid altcoin perpetuals once per day at 06:00 UTC. Four orthogonal microstructure indicators are computed for each symbol:

• Retrace — pullback from the recent 30-day high. Larger pullback = closer to potential support, but also larger downside if support breaks. We surface the value, not a recommendation.
• OI drop — open-interest decline from its recent peak. A large drop means leverage has been flushed: the weak hands are out, the chart is ready for either direction without forced liquidation cascades.
• Funding (8h perpetual rate) — measured in basis points. Negative funding = shorts are paying longs to hold their position, which means positioning has become bearish. Often a reset signal, not a directional one.
• 1st-wave magnitude — the multiple from the recent local low to the local peak. This is a context indicator: a 5× first wave behaves differently from a 1.5× first wave.

A symbol that simultaneously satisfies internal thresholds on these four axes — with a separate veto on macro BTC weakness — is published as a candidate. "Candidate" means "screened in"; it is not a buy signal.

The published shortlist replaces the previous day's at 06:00 UTC. Yesterday's list is intentionally not preserved on the public surface — what was a setup yesterday is no longer a setup today. The historical track record is computed offline against fixed entry rules.

4. Polymarket whale signals

Polymarket trades above a configurable USDC threshold are pulled directly from the on-chain feed (Polygon CEL) and joined with a curated list of high-volume wallets. Every surfaced signal includes the wallet address (clickable, on-chain auditable), the market, the side (BUY / SELL of the YES token), the size, and the average fill price. The "wallet historical winrate" badge is computed from the wallet's settled markets, not estimated.

A whale buying YES is not a recommendation — it is a fact. The signal exists so you don't have to scan blockchain feeds yourself.

5. BTC large positions

Large open positions on Hyperliquid (the on-chain perpetuals exchange) are surfaced with their entry, current size, leverage, and estimated liquidation price. Public chain data only — no private data, no positioning APIs.

6. What is published vs what is held back

The public surface shows: every verdict, every wrong call, every settled outcome, every active candidate. The system stack — model identities, exact thresholds, infrastructure topology — is intentionally kept private. This is not because it would compromise the methodology if revealed; it is because any thresholded signal degrades the moment its threshold is published. Reflexivity, not secrecy.

If you are evaluating whether to trust the output, look at the /status page (live system health), the live 30-day accuracy on the dashboard (live track record), and the wrong-verdict history (counterfactual evidence). If those three are honest, the threshold isn't load-bearing.

7. Update cadence

• Whale signals — sub-minute, polled from Polymarket on-chain feed
• AI Jury verdicts — published as news flows in (irregular, typically 5–30 verdicts/day)
• Outcome calibration — exactly 24h after each verdict, automatic
• CEX shortlist — re-screened daily at 06:00 UTC
• BTC positions — every 60 seconds
• Fear & Greed — once per hour from public source

8. Subscribe / RSS

To track everything in real time without refreshing:

• /rss/verdicts.xml — every AI Jury verdict
• /rss/whales.xml — every Polymarket whale signal

[ buchiang@v1 // methodology snapshot 2026-05-04 ]