Engine Forge and Colosseum

It invents the next muscle. Then it races it.

A spec goes in. Engines come out

A desk can ship one module and stop. The get is a spec in, a field of candidates out, and a race that keeps the muscle. Natural language states the job. A structured schema locks the types, the latency budget, the error bound, and the memory ceiling. That specification is the only input. Code generation writes a single Python module the runtime can call: validate the input, run the one job, format the output, handle the miss, emit the metric.

Every candidate inherits the same five-part scaffold. Validation refuses a bad payload before the job starts. Process is one responsibility — no second career hiding in the same file. Format returns the type the caller already asked for. A miss is a typed miss, not a silent wrong. Metrics leave the module so the arena can see the body. The Forge writes the next candidate from the same spec until the field is thick enough to race.

• Specification in — purpose, inputs, outputs, performance targets, resource ceiling.
• Template scaffold — validation, process, format, error, metrics. Same bones, every engine.
• One job — a module the runtime invokes and then forgets.

A spec goes in. Candidate engines come out. Then they race.

Self-test at birth

A self-test fires the moment the module exists. Known input. Known output. A tolerance. Pass, and the candidate is alive. Fail, and the Forge writes again from the same spec. The first body that cannot return the answer it was born to return never sees the arena.

AMMA reads whether the new modules hold together as a whole — fourteen meridians, live lattice texture, the healer the house already calls. AMMA observes and heals. ABBA runs the baseline suite: latency, accuracy, resource use, scored against a pre-defined card. Φ×Ψ is the score. Coherence, then baselines, then the bracket.

Birth is cheap and honest. The expensive proof is the race that follows.

The Colosseum keeps the winners

Validated engines race head-to-head on identical hardware against identical workloads. Store and recall one hundred items and time the get. Route five thousand requests and keep p99 honest. Classify one thousand spiking patterns and report accuracy. Best of one hundred. First correct result takes the round.

The arena scores what the factory cares about: latency, throughput, error rate, and sustained load. Load ramps until the module breaks, then holds eighty percent of that breaking point for an hour. A ten-times spike for ten seconds must stay up, keep the data, and recover in thirty. Losers that drift, leak, or miss the threshold are deprecated. Winners graduate. Then they come back and race again.

• Daily quick races among the current top five — the pulse of the field.
• Weekly full tournament across the live engine index.
• Monthly stress marathon under sustained pressure.

The arena writes the ranking. The factory grows the next muscle from it.

Graduate, tune, race again

One closed lifecycle. Create in the Forge. Self-test at birth. Check coherence and baselines. Race in the Colosseum. Graduate the winners and tune them — golden-section search on the knobs, at most ten iterations, stop when a pass buys less than one percent. Re-race the tuned module against the latest workloads. Improve, or retire.

Algorithms race the same way engines race: two variants, one bench, the loser is discarded. A module graduates only at a ninety-percent score across self-test, hardening, edge cases, resource caps, failure modes, compatibility, and chain latency. The full stack — entity to engine and back — is scored as a weighted geometric mean of fifty standardized tasks. The leaderboard prints p50, p99, accuracy, throughput, memory, efficiency. A champion can be dethroned the next week. That is how the farm stays sharp.

The factory grows new muscle

The live index is 1,647 modules — one job each, sitting on a registry and gateway the runtime already calls. New muscle is invented when the spec appears, raced until it wins, then written back onto that index. Two births a week is the paper’s cadence. One retirement a month is the paper’s cull. The population grows because the loop is selective and because the next spec still wants a body.

That is the product of this door: a generation-and-selection factory that keeps minting the next motion and letting the arena keep the one that holds. The archive gets faster muscle without a human rewriting the catalog. The factory writes. The Colosseum keeps. The index grows.

What the factory is looking at

Invent, then race. Spec in, candidate engines out, self-test at birth, Colosseum on the winners, a factory that grows the next muscle. Engine Forge and Colosseum is a research door — never a SKU on the public raise, and not a Year-1 Memory SKU. Isolated recall on a quiet machine is 13.834 µs. Under load we have also seen 5,127 µs. Those are memory clocks, shown together. Twenty million dollars is an ask to land the archive first. The loop is how that archive’s modules keep getting sharper.

Next in the catalog: Thoth.