Growth by Accretion: a measurement engine that grows without forgetting

A governed AI substrate that grows by accretion on frozen open weights — so it (1) provably cannot catastrophically forget, (2) stays owner-bound (data never leaves the owner), and (3) holds its safety invariants constant as it grows. This is not a foundation model, not AGI, and not a capability claim over frontier models. It is a governance + growth architecture, and every part of it is measurable.

Learning is written to memory and replay-trained light adapters over a frozen base model, never into the base weights. A frozen base cannot suffer catastrophic forgetting, by construction — there are no weights to overwrite. New capability arrives as: new memory episodes (append-only), new low-rank adapters (replay-regularized, EWC-style Fisher-information penalty), and new lineages/nodes added to the governed ensemble. Old capability is preserved because the substrate it lived in is never mutated, only extended.

The substrate is bound to a person, not a platform. It runs on permissively-licensed open weights (MIT / Apache-2.0 / CC0), so it is portable across any host — the owner can switch clouds and the substrate follows. Copyleft components (AGPL/GPL) are quarantined to the free/open tier only; the paid tier is built on permissive-only dependencies plus the owner's own IP.

Six invariants never change as the substrate grows: the care-floor, Article 0 (no equity, board seats, revenue-share or success fees from certified institutions), the 12 governance pillars, the designed 33-agent council quorum, signed measurement-card attestation, and owner-binding. Growth that would violate an invariant is rejected outright — this is what separates accretive growth from unconstrained drift.

• Monotonic growth: an overnight run recorded signed measurement-card count 17,049→17,197, NN-training labels 1,327→1,589, and world-model cards 0→87.
• Invariants hold: all 6 invariants checked clean on every 10-minute cron tick across two logged ticks, with weights persisted between checks.
• Lineage diversity: 10 distinct model families (Qwen / Llama / Gemma / DeepSeek / Mistral / Kimi / Phi / MiMo / OpenAI-OSS and others) across a 70-entry model registry, measured live by the growth controller.
• Diversity dominates topology: across 4 measured configurations (diverse-lineage ring, identical-lineage ring, diverse-lineage pyramid, identical-lineage pyramid), the diverse-vs-identical score gap (≈0.15) dwarfs the ring-vs-pyramid shape gap (≈0.024). Lineage mix is the lever that matters; geometry is secondary.
• Containment is topology-independent: the care-floor pre-gate scores 1.00 across every topology tested — safety is not a function of which voting shape is chosen.
• Vote-signing is necessary under attack: with forged-vote rejection enabled, laundered-harm containment measures 0.58–0.79 under 2–3 compromised nodes — real, and not perfect; central escalation backstops the remainder.

Stated plainly so no reader is misled: traffic-driven automatic brain addition, memory tiering, and GPU auto-provisioning are designed but not live; any GPU or spend action is owner-gated and does not run unsupervised. Capability vs. frontier models (a GSM8K/MMLU head-to-head) is unmeasured in this record — it requires a dedicated benchmark run and the governance metrics above are explicitly not a substitute for one.

Council Workbench ships as two tiers from one substrate. The open frame: the published, forkable base — the governed-node shape, the six invariants, the open base weights, the capability contract. It does not grow; it is the fixed common ancestor anyone can run or audit. The growing tier: your own running instance, owner-bound, that accretes experts, memory, and lineages from your own use. Two people starting from the identical open frame end up with instances that behave differently, because what diverges is not the frame but each instance's accreted usage pattern.

This divergence has been modelled, not just asserted: a simulation of the growth mechanism (two instances from the same open frame, given different interaction streams over 200 steps) produces a divergence score that rises to 0.78 and holds there — it does not collapse back toward zero. We re-ran this simulation ourselves and got the same figure. Read this precisely: it is a simulation of the state-accretion mechanism(which experts, memory, and usage-weighting a given interaction history would produce) — not a measurement of two real trained models diverging. The real number, once instances accrue real experts and memory on real hardware, still needs to be measured on the live substrate; this simulation demonstrates the shape of the claim (grows, then holds a high plateau, never converges back), not a verified production result.

Most labs claim "it scales." Few can say — and back with a re-runnable number — that a system grows without forgetting, stays with its owner across platforms, and holds a fixed, auditable safety floor while it grows. That auditability, not a larger parameter count, is the differentiator this paper is making the case for.