← The write-up

Governed Corpus Evolution

Continuous refinement of knowledge for frozen-model agents — the doctrine paper, and a small worked example. This page carries the formal material behind the write-up.

The model that powers AI-assisted development is frozen and rented: it learns nothing between sessions, remembers nothing of your project, and its weights are not yours to train. Everything the project knows must therefore live in text. Governed Corpus Evolution (GCE) takes that constraint seriously: the project's document corpus — its instructions, decisions, rules, and notes — is the learning system, and it can be run as one, deliberately, with the rigor learning systems require.

The paper argues three things:

  1. Unbundle what weight training bundles. Storage goes to a typed, versioned corpus; generalization to the model at read time and to human judgment at write time; behavior binding to mechanical gates — plus an honest disclosure of what those gates do not check.
  2. Every entry carries its own learning loop. With no global optimizer left, each persistent unit of knowledge must locally say when it activates, what it caches, what warrants it, what enforces it, and how it retires. Rules, decision records, ledgers, and gates turn out to be one five-slot contract evaluated at different coordinates.
  3. Govern the writes. Beliefs are ledgered hypotheses exposed to structurally independent contradiction, and a human gates every permanent write — because the surveyed alternatives, where the model curates its own memory, exhibit measured collapse, contamination, and self-bias.

Read & try

The paper — Governed Corpus Evolution
PDF · 38 pages
The theory: the frozen-model constraint and the doctrine it forces, the entry contract, the refinement machinery, and a candid account of limits. A position and design-theory paper, distilled from several hundred working sessions of one production system.
 The example — Slavarium
ZIP · plain-text template
The doctrine projected as a minimal starter kit: the boot document, the typed stores, the latch registry, and a short seed sequence. Copy it into a repository, run the seed tasks, and grow it from your own evidence.

Stores · Latches · Attention · Versioning · Artefacts.

Slavarium (n.) — from SLAVA (above) + Latin -arium, “a place where a thing is kept”: a governed archive of legible, attributable, reversible project memory.

Nothing here is for sale, and nothing asks for your data. The paper argues a position; the template is a worked example of it. Both are offered in the spirit the paper closes on: the strongest test of a doctrine extracted from one system is a second, independent adoption. The full research repository — including the origin write-up — remains private; access is arranged on request with the maintainer, Sviatoslav Zinevich (slavazinevich@gmail.com).