RunSpec, Bootstrap, Bayesian Scoring, and the Pyramid

Config Consolidation With RunSpec

The morning’s big structural change was a RunSpec that unifies configuration (9b41bdf9). I wired an orchestration layer with a profile loader, executor, stages, and CLI/API verbs (683e14c2), added pyyaml and a sweep seed, and threaded meta-label knobs through (d82403af). Consolidating config means a research run is now described by one object instead of a dozen scattered arguments.

Cockpit and Bootstrap

I rewired the cockpit and pump onto RunSpec with cooperative cancellation (9bfee005, 80de07da). Then I added a BootstrapSpec with data-ingest stages (db_init, universe, ohlcv), derivation stages (seed registry, exit profiles, legacy corpus), and prerequisite checks (2653d539, ddf4cc14, 3de085d2), surfaced through the CLI, REST, and cockpit data panel (22961c3e).

Bayesian Layers

I added closed-form Beta-Binomial Bayesian edge scoring (a72a2427), a Gaussian HMM regime feature as a bootstrap stage (012430cb), and a Bayesian optimization mode for the exit grid (8565def0). The Beta-Binomial estimator is appealing because it gives a principled edge estimate with uncertainty instead of a point win rate.

Diversity and the Corpus

I added a legacy corpus ingest pipeline (f71771e3), an LLM-prior mutation mode that samples mutation patterns from that corpus (64160bd0), cross-cell ensembles (acd1de19), and regime-specialized baskets (9bf69964). Using a legacy corpus as a prior for mutation is a nice way to reuse accumulated research instead of starting each sweep cold.

Turning the Pyramid Into a Funnel

The performance arc of the day was the pyramid. I added a champions ledger with per-tier promotion (958f5e5d), a --pyramid wrapper that auto-chains tiers (6a69f49b), caching so pump cells stop rebuilding (e4d0fcdc), and a quick-tier profiling loop that exposed hidden multipliers and cut wall time by 18% (873cb4b8). I fixed smoke profile timing to get under two minutes and resolved a Bayesian optimization hydration bottleneck (cfdfbdee).

Merges and Cleanup

The day was a long series of focused branches merged back in: runspec, cockpit rewiring, LLM-prior mutation, cross-cell ensembles, corpus ingest, regime baskets, bootstrap foundation, data ingest, derivation stages, surfaces, Bayesian scoring, and Bayesian BO exits. A registry-driven profile now hooks every shipped feature into a default path (322489e0).

The Trajectory

Edge Lab went from “a pipeline” to “a research system” today. The pyramid and the profile system are the parts that make it reusable—the next experiment is a config change, not a new script.