Inside the Lumerian Core: platform, Luminol and runtime

Every Lumerian product runs on one Core with three layers: the platform, Luminol and the runtime. Every product runs the same seven-stage loop: observe, identify, detect, verify, decide, act and follow up. Stages one to four run today in Aperture and Prism; stages five to seven run as the record fills. Each layer works to fixed rules, and each capability runs only where its gate is met.

01Platform: the shared record

The platform holds the shared record: every module, parcel, permit and milestone under a stable key, and every observation of it, from a thermal flight to a data-room revision to a weekly update. Every team writes to it, and every customer number is computed from it. Five rules govern it.

• –Append only. A correction inserts a new row that supersedes the old one; a database trigger rejects updates and deletes.
• –Three clocks. Every observation records when it happened, when a decision could first use it and when it was ingested. A feature may use only what was available at decision time.
• –Named producers. Every derived fact names its producer and version: detector, matcher, language model, person or rule. Labels set by a language model never become test truth.
• –A follow-up for every verified finding, treated or not. Without it, outcomes describe only the repaired findings, and no comparison of actions is possible.
• –Observed and counterfactual numbers kept apart. Outcomes and costs are observed. Avoided loss and the cost of waiting are estimates, each with its method and interval.

02Luminol: models behind gates

Luminol is the model layer: the detector, the segmenter, the identity matcher, the document extraction engine and its critic. Each is a versioned producer that must beat a named baseline on frozen splits before it ships, and each answer carries its evidence ^[1]. Public foundation models are interchangeable parts. The record is ours, and so is every model trained on it once it clears its gate.

03Runtime: hosted or on premises

The Core runs hosted, with Lumerian running the stack, or on premises, where the GPU tier runs in the customer's environment and no cloud step is used. The target compute design has three tiers: rules and small models on a CPU or a small edge device; detection, matching and document models on one GPU; large training runs on cloud or batch GPUs. Some operators cannot connect operational systems to the cloud at all ^[2], which is why the on-premises pattern exists, and open world models now run real-time control on edge hardware ^[3].

Each model is confirmed on its tier by profiling throughput, memory, latency and cost before it ships, and every release records its code, model, prompt and split versions together, so it can be reproduced. Human review is a tier of its own: identity conflicts, unsupported evidence and high-impact actions go to people, and review queues are sized and staffed like compute.

04How we test it

• –Frozen splits. Each release freezes its splits by flight, by site group and by time, after hashing and near-duplicate removal.
• –A logged look budget. Held-out tests are scored once per experiment; the sealed test once per release.
• –Simple baselines first. Every model must beat a named baseline, such as the last observation, an engineer's rules or a physics model on its own.
• –Metrics with denominators. Recall and review load per 1,000 modules; wrong-row rate with its coverage; Brier skill against the last observation.