How we form market opinions

This is the public research contract for Sapphire Alpha — and the charter for a self-improving local agentic system that observes markets, publishes opinions, scores itself, and updates its own priors.

What we are doing

1. Event probabilities — one number as of a timestamp for binary claims
   (e.g. “Has Bitcoin put in the cycle low?”).
2. Path forecasts — short / medium / long price and trend bands, not event odds.
3. Falsifiers — written before the outcome so we can score accuracy later.
4. Learning loop — on a schedule we resolve due claims, compute Brier scores, record wins/losses, update priors, and publish calibration notes.
5. Trade ideas — gated thesis-lane notions on designated capital only.

We allow the system to speculate, including on ambiguous regimes. Speculation is labeled and forced to cite data and a falsifier. After the fact, it is scored.

What we are not doing

• Multi-horizon probabilities for the same binary event.
• Point-price “targets” without scenarios.
• Publishing live balances, wallets, or private holdings.
• Guaranteeing returns or claiming permanent edge.
• Paper backtest leaderboards as “research.”

Data

Source	Use
CoinGecko	BTC/ETH/SOL spot, 24h/7d/30d changes
CoinGecko global	BTC dominance, total mcap
Yahoo chart (best-effort)	VIX, SPX, DXY, oil, gold proxies
Local learning store	resolved outcomes, prior deltas, lessons

Each opinion carries a data_quality score (0–1). Low quality → more prior weight.

Method — full autonomous loop

public market snapshot
        ↓
event probabilities (single p) + path bands (S/M/L)
   ↑ prior Δ from past scores
        ↓
falsifiers + drivers + confidence labels
        ↓
open book (track) + trade ideas (gated)
        ↓
score due events (Brier) + path hit-rate
        ↓
learn: update priors · write lessons · note-taking
        ↓
publish opinions + calibration report
        ↓
(act on designated rails under caps — 24/7 when armed)

Learning cadence

Interval	What the system does
Every cycle	Refresh open book; score anything past its review window; reload priors
Daily	Full learning cycle + publish calibration / post-mortem
Weekly	Domain roll-up, lesson pruning, path-band hit-rate review
At resolution	Binary outcome + Brier; lean accuracy when stance was lean_yes/no

Metrics we track

• Brier score — (p − outcome)² · 0 perfect · ~0.25 unskilled 50/50
• Lean accuracy — when we leaned, were we directionally right?
• Skill vs coin-flip — 0.25 − mean Brier (positive = better than noise)
• Path hit-rate — did price land inside the published bear–bull band?
• Prior Δ — automatic cool-down of chronic overconfidence per question

Wins and losses become lessons stored as agent notes. The next opinion cycle reads those notes as prior adjustments — not as forgotten chat.

Confidence labels

Label	Meaning
moderate	Strong data coverage and a clear lean
tentative	Usable data, modest lean
low	Thin edge or incomplete feeds
speculative	Prior-heavy; treat as hypothesis

Why this is useful

• Comparable over time — one P per event.
• Honest uncertainty — residual mass (1 − p) is explicit.
• Self-correcting — bad calibration shrinks future extremity.
• Separates thesis from path — event vs price trajectory.
• Built for autonomy — open data, open-source stack, local models where possible; execution stays on designated rails with caps and a kill switch.

Process claim

1. Observe
2. Conjecture (single event P + learned prior)
3. Path forecast (S/M/L only for prices/trends)
4. Score
5. Learn + note
6. Act (gated)

Not investment advice. Designated test/agentic capital only. The goal is a continuously improving system that earns on those rails — not a static research blog.