Plain-language methodology.

01 · the input

The signal universe.

On weekdays the engine scans the top ~500 most-liquid US equities (by market cap, $1M+ daily dollar volume) plus the top 50 crypto assets by market cap (sourced via CoinGecko). On weekends and US market holidays, equity data is stale, so the engine runs in crypto-only mode — the same Council, applied only to assets that trade 24/7.

Every ticker is scored against 16 technical indicators — momentum (RSI, MACD, stochastic), trend (ma_cross, ADX, Ichimoku), structure (Bollinger, ATR), volume/flow (volume_spike, OBV, VWAP), candlestick patterns, and (for equities) fundamentals + sentiment. Each indicator returns a normalized score in [-1, 1]. A weighted average produces the composite_score.

02 · the deliberation

The Teranode Council.

For each top-scoring signal, the engine runs a five-role council of independent reasoning models:

• Risk — identifies the strongest failure mode of the trade as stated.
• Builder — makes the strongest case for the trade resolving favorably.
• Strategist — frames the trade in regime / catalyst / second-order context.
• Contrarian — dissents from the consensus view; pressure-tests assumptions.
• Chairman — synthesizes the four roles into a single thesis with preserved dissent and a confidence percentage.

The chairman's output is what the channel publishes. Every role's reasoning is stored alongside it so any signal can be inspected at /track-record → click the inspect link on any recent signal.

03 · the grade

Teranode Score — 0 to 100.

Every signal is graded by a deterministic scoring engine that combines:

• Base conviction from the composite score
• Risk/reward asymmetry from the trade plan (entry/target/stop geometry)
• Alignment count across the 16 indicators
• Volume/flow confirmation
• Trend structure (ma_cross + ADX + Ichimoku)
• Fundamental support (for equities)
• Penalties for conflicting indicators or missing data

One additional rule: an A or B grade requires R:R of at least 1.0; below that ceiling, the grade caps at Watchlist regardless of conviction.

04 · the named-human gate

Discretion before publication.

Signals at 80%+ conviction are not auto-published. Each one is held for personal review by Dan Zimon (Series 7 + 66, 14 years institutional finance) before it goes to the public channel or premium subscribers.

The reviewer either approves (publishes within 10 minutes, audit row tagged with approver) or rejects (held permanently, never publishes). Either decision lands as a row at /notebook. Tokens expire 24h after issuance — if a high-conviction signal isn't reviewed in time, it's dropped, not posted. Named human accountability for every high-impact public statement.

05 · outcomes

Every signal is graded against reality.

Each published signal has a fixed entry, target, and stop pinned at the moment of publication. At 1d, 3d, and 5d horizons, the engine fetches actual market data and records: did price hit the target? did it hit the stop first? what was the realized R-multiple at horizon end?

These outcomes accumulate into the public hit-rate-by-grade visible at /track-record. You can audit every claim. The math is transparent: hit_target / (hit_target + hit_stop). Open trades and signals with no market data are tracked separately and never folded into the rate.

06 · the autoresearch loop

The system improves itself.

Every night at 02:00 ET, an autoresearch sweep runs. It evaluates seven hypothesis variants of the scoring algorithm against the same historical signal-and-outcome data, picks the variant with the best avg-realized-R, and — if it beats the current champion — promotes that variant. Three eval gates protect the trail: invariant lint at 01:50, golden-master regression at 01:55, contract gate per-variant during the sweep itself.

Every variant evaluation, every champion change, and every metric delta lands in an append-only audit log at /notebook and at /api/feed (RSS). The system narrates its own learning. You can subscribe.

07 · disclosures

What this is and isn't.

• This is research, not advice. Signals describe the engine's read of public market data. They are not individualized investment recommendations.
• Past performance is not predictive. The track-record page shows what happened. It does not guarantee what happens next.
• The engine takes no positions. No custody, no brokerage, no order flow. The output is an artifact for your process — you decide what to do with it.
• Risk-defined. Every signal has an explicit stop. The stop is the maximum loss on the trade if executed at the published entry.
• Model-driven. Council role outputs come from frontier reasoning models accessed via OpenRouter (or a local Apple-Silicon model when configured). No human writes or edits the thesis text after the council finishes.