One market. Twenty-four specialists. One selector deciding who acts now.
The specialists did not arrive fully formed. They were trained through historical replay, refined through observation, and recorded in institutional memory — a memory that is still growing.
Pass 1 began on M0 — November 2022, with Bitcoin near the $15,000 cycle bottom. From M0 the strategy bots traded forward through thirty-six months of real Bitcoin market history: bull, bear, accumulation, mark-up, distribution, mark-down. Every regime present in the data. Every shape encountered at least once.
The mechanism is called the Replay Harness — what Pascal becomes when it runs against history instead of the live market. Historical candles feed in at the edge. The full stack processes them as if they were the moment. Each bot fires only when its conditions present — RSI bands, EMA alignment, ADX thresholds, breakout signals, pullback structures — deterministic rules defined inside each strategy. Each trade leaves a record.
Above the bots, Tier 7 (Selector) and Tier 8 (Supervisory) observed. A Reflection Engine watched outcomes per cycle, named the top performer, named the bottom performer, escalated patterns of failure. Bots that earned their place stayed on the roster. Bots that did not were retired or retrained. The specialists were not authored — they were trained.
Alexandria — Pascal calls her Alex in conversation — is not a document library. She is Pascal's institutional memory. A Postgres + TimescaleDB substrate with a structured document store, holding every canonical record the system has ever produced. Every decision filed, dated, addressable.
What is kept:
Nothing is deleted. A retired strategy from today might be the right fit for a market condition five years from now. The institutional memory makes that rediscovery possible.
The learning loop runs continuously and openly. The Selector reads the present and gates which cluster may act. Alexandria records what happens. Tier 8 — the supervisory review layer — studies the archive across cycles, weeks, and quarters. Past trades are not stored as static records. They are stored as usable evidence.
What Tier 8 does with the archive: compares current conditions against archived behaviour, tightens thresholds where evidence supports it, loosens them where the data permits, re-weights specialists by demonstrated regime-fit, retires bots that fail to earn their place, proposes new ones when a gap shows up across enough cycles. Every adjustment routes through the Board, and every Board ratification routes through the Principal.
Pascal does not improve through software releases. Pascal improves through observation routed back into the policy layer. Six months of operation produces a different system than six days. Six years will produce something we cannot yet imagine.
The policy-adjustment mechanism that sits between observed outcomes, archived evidence, and refined future decisions is informed by earlier open research in agent-based optimisation — including the GEPA implementation explored through the HERMES project under MIT licence. Pascal references that work as a reference implementation, not as an adopted framework. The actual review layer — what is weighed, what is amended, what is retired — has been adapted, extended, and ratified by the Architect, and runs inside Pascal's own eight-tier governance with the Principal seated above it.
Pascal doesn't run one strategy. Pascal doesn't run twenty-four strategies at once.
It observes the market. It classifies the regime. It calls forward the right cluster. That is the architecture.
At any given moment, a single selector — Tier 7 in Pascal's eight-tier governance — reads the current shape of the market and decides which family of strategies is permitted to fire. One regime active. The rest on standby. The cluster that is on shift trades; the others wait, indexed and ready, watching for their conditions.
The benefit is structural: capital is never deployed against a strategy mismatched to the environment. The ranging family does not trade in trends. The trend family does not trade in chop. Each bot earns its keep by being calibrated to one shape — and by knowing when to stand down.
Twenty-four specialists. One selector. One decision at a time.
The cluster Pascal calls forward when structure breaks higher and the four-hour confirms. Five specialists — each reading a different shape of strength.
When the market refuses to trend, the ranging cluster takes the watch. Five specialists working the edges, not the middles.
The mirror of trending-up — but with its own discipline. Four specialists tuned to the asymmetry of selling.
The most disciplined cluster. Five specialists for the regime no system likes — where the rules of every other family stop applying.
The slowest regime — where smart money builds or unwinds position before the next trend declares itself. Five specialists reading the footprints.
No bot trades alone. No regime trades without supervision. No capital deploys without structure.