Sapphire Alpha · live decision observatory
Four networked machines observe markets around the clock. Five strategies write down their claims — with the source, the age, and what would prove them wrong — and no capital moves until a human authorizes it with one tap. Every number on this page is either live from an endpoint or ships with the command that produced it.
Source: /api/v1/live · authority: none · unknown stays unknown.
01 · System
A private mesh of small computers. Every node reaches every other node over Tailscale — no single machine holds all the truth. The green ring is a live status pulse; the moving particles are network activity, and they slow down when the market feed slows down.
Source: /api/v1/live · authority: none · unknown stays unknown.
02 · Intelligence
Every strategy runs on paper. Each proposal carries its source, its age, and what would prove it wrong. The only way capital moves is one tap on Telegram. Nothing on this panel comes from a mock.
Live strategy tracks
RSI mean-reversion, gated by the BTC 20-day regime state. · strategies.py:178
RSI(7), 20-SMA, and 5-day return must all agree before entry. · strategies.py:323
Fires when BTC↔SPY correlation breaks down and RSI is oversold. · strategies.py:275
Fades 3-day momentum as a proxy for one-sided funding. · strategies.py:237
0–100 score across all four; quarter-Kelly position sizing. · strategies.py:372
Market context
Signal stream
03 · Research
Every piece here was generated by a Sapphire agent and then read by a person before it published. The corpus is committed to git — a stale claim is visibly stale, and the sync script cannot ship a piece that did not carry publish: true.
Weekly Sapphire security digest — 5 actively-exploited CVEs (all CISA KEV, average CVSS 9.7). Two past CISA remediation deadline. Sourced from CISA KEV and NVD; informational only.
Read →The public contract for free-reign easy on designated RH Agentic capital: clip-to-cap, per-venue slots, Super Heavy as planner only, kill switch always wins.
Read →Scheduled scoring of event probabilities and path forecasts. Brier calibration, lessons, and prior updates for the self-improving desk.
Read →Point-in-time event probabilities, short/med/long path forecasts, data quality, drivers, and written falsifiers.
Read →Predictions and analysis for held names and priority interests — fundamental, quant/TA, narrative/vibes, related bluechips, OSINT situations.
Read →Event probabilities as of now, path forecasts for targets and trends, public data, falsifiers, and a scheduled self-improving learning loop that scores wins and losses.
Read →The standard every report is held to, what is deliberately withheld, and why an empty page is better than a confident wrong one.
Read →A measurement audit of the public system map, the telemetry claims it could support, and the gaps it now leaves visible.
Read →04 · Proof
Anyone can build a beautiful dashboard. Fewer can point at one live settlement rail, one mandate that cleared four safety boundaries, and a test suite that keeps growing. These are the three things that make the rest of the site worth reading.
Settlement rail
Three contracts carry the surface — SapphireSignalVerifier, SapphirePaymentGate, and SapphireSentinelRegistry. Public wallet 0xc2B5…c9EB; keys never enter this repo.
First cleared mandate
The first mandate that arrived at the confirmation firewall, was scored by the policy shim, and cleared to execute. Small on purpose. The value of BRODIE is not the dollar amount; it is the four boundaries that had to hold for it to happen.
Under $5 per proposal — an order that would spend more is refused before it leaves the machine.
Realized loss above $25 in a day halts execution for the rest of the day.
Three times or less. Anything above rejects at the gate.
One Telegram tap authorizes exactly one order — no standing approvals, no batch clears.
Verifiable coverage
7,695
Test functions defined across 484 test modules.
git ls-files '*.py' | xargs grep -h '^\s*\(async \)\?def test_' | wc -l05 · About
Ari Spector designs and operates the Sapphire mesh from Houston, Texas. The system is small on purpose — one commander, one GPU executor, two edge sensors — because every extra node is a surface an operator has to keep honest.
The public site is a contract, not a brochure. Numbers ship with the commands that reproduce them; live values are read from /api/v1/live and unknown stays unknown. If a claim on this page cannot be checked, treat it as fiction.