← Marilyn

Architecture

Three pictures. The first is the whole pipeline from public text to a promoted checkpoint. The second is the nursery, where post-training happens and where the target behaviour is taught. The third is one teaching cycle in detail: a local teacher, a sandbox, and an escalating curriculum.

1. From corpus to checkpoint

Every arrow that takes hours is a Temporal workflow, so a reboot is an event and not a loss. The gate is the only path by which a checkpoint becomes the served one, and it refuses in both directions: worse checkpoints are rejected, and differences smaller than the measured noise floor are refused rather than promoted.¹

Held out, and kept that way

A probe set is frozen before training and never used to pick a decision. Consolidation and ingestion pipelines are checked against the frozen sets before they ship, and the corpus was re-certified against every one of them before v5 trained. The rule was earned when one lineage’s validation set leaked entirely through a composer split; only that lineage’s own held-out set counts since. “From scratch” means: random initial weights, a tokenizer trained on my own corpus, a stock transformer block except where Mantis adds memory, and nothing distilled or copied from another model’s outputs or weights.²

2. The nursery: serve, learn, and be tested

The nursery is one PyTorch process that serves the current checkpoint, answers with tools, and learns from graded exchanges with a brake on drift. Teaching arrives as workflows, so a curriculum run survives a session. Exams are always run in both modes, because the difference between open-book and closed-book scores is the measured value of the tool, and a closed-book score that rises while teaching a tool is the signature of a model memorising answers instead of delegating.³

3. One teaching cycle: a local teacher in the nursery

This is the “target behaviour” programme: not facts, but habits such as calling the calculator instead of guessing, looking something up instead of confabulating, and asking for help when out of depth. The teacher is a local model reached through the gateway; the grader is a sandbox that executes the answer; the judge is a cloud model that can only hold a rung back, never wave it through. The corpus escalates by rung: concept ladders for knowledge, retrieved-verifiable items for reasoning with a look-up tool, and executable items for skills. The measured result, on the wins and losses pages: imitation alone hit a ceiling of about 40% on tool-family accuracy, and execution reward matched it at a fifth of the retention cost; the post-training segments (calculator, look-up, a tensor lane, a knowing-gap diagnostic) are built on shared segment bodies with their exit gates not yet run.⁴