Maxim Hivemind + Oasis

What this is

The Maxim Hivemind is a peer-to-peer substrate-sharing layer that lets multiple Maxim instances exchange learned bio-substrate (NAc causal weights, EC concept centroids, reflex thresholds, Cerebellum forward models, ATL semantic concepts). It is not a service — nobody hosts "the Hivemind." It is a protocol + portable data format + exchange convention.

A Maxim Oasis is a persistent, substrate-primary Maxim instance that runs on operator hardware (a Mac Mini, a server, an old laptop, a Raspberry Pi 5 cluster), accepts substrate contributions from connected Maxims, processes those contributions as percepts through her own bio-stack, distills consensus patterns, and broadcasts to the Hivemind. Multiple Oases coexist; nobody is THE Oasis. Each Oasis is a real bio-agent in her own right, not a passive database.

Why this exists

The headline insight that motivated promoting the old Mother Maxim plan from deferred to active:

That changes the federated-learning calculus radically:

What's shareable — and what stays local

Default policy: ship everything in the top section. NEVER ship hippocampus episodes through the Hivemind. Episodes are local-only by construction. The single exception: an Oasis may receive hippocampus episodes from connected Maxims as private contributions to her own substrate (she processes them as percepts), but she never re-broadcasts them. Episodes feed the Oasis's substrate; only the Oasis's distilled substrate ever leaves.

The Oasis

Hardware footprint

A Mac Mini, an old laptop, or a Raspberry Pi 5 cluster is sufficient to run an Oasis IF substrate-primary mode is the operating mode (no LLM, just bio-systems running). During the transitional phase (v1.1) when substrate-primary is still maturing, Oases may need a small LLM (e.g., qwen-7B) to process LLM-AUT contributions. Once substrate-primary works (post-Phase 0/1), Oases can run pure-substrate and Mac-Mini-class hardware suffices.

Multiple Oases, no central authority

There is no THE Oasis. Each operator runs their own. A natural evolution emerges:

The Hivemind

Substrate snapshot bundle

The unit of exchange. A versioned, signed archive containing:

maxim-substrate-v1.0/
├── manifest.json           # version, contributor metadata, signature, domain tags
├── nac.json                # causal links + confidence + provenance tags
├── ec.json                 # concept centroids + cluster metadata
├── atl.json                # semantic concepts
├── reflexes.yaml           # innate response specs
├── cerebellum/             # forward model parameters (binary)
└── README.md               # optional human-readable description

Schema-stable, version-aware, scrubbable. Extends the existing _format_version contract per Maxim's v1.0 freeze.

Substrate domains

Bundles are tagged by substrate domain (combat, cooking, medical, fantasy, robotics, conversation, ...). Subscribers opt into specific domains. Two purposes: relevance (a medical-AI hobbyist doesn't want combat patterns polluting their substrate) and curation (domain maintainers can flag bad patterns).

Conflict resolution

Two Maxims learn opposite things ("X is good" vs "X is bad"). On merge, the Hivemind uses NAc's existing confidence math:

• Confidence aggregates Bayesian-style across contributors
• Multi-source confirmation weighted higher than single-source
• Outcome valence (positive vs negative) preserved as separate distributions
• Provenance preserved — receivers can selectively trust patterns from specific contributors

When real conflict persists (e.g., 50 Maxims learned each side), the Hivemind preserves both with full distributions. The local Maxim's behavior remains coherent because it weights its own learning above any single Hivemind pattern.

Poison resistance

Adversaries trying to corrupt the Hivemind by spreading bad learning are mitigated by:

1. Multi-source consensus — patterns require N independent contributors before promotion
2. Domain curation — domain maintainers can flag specific contributors as untrusted
3. Local provenance tracking — every pattern carries source tags; receivers can blacklist sources
4. Outcome correlation check — Hivemind patterns that hurt locally are weighted lower over time (the substrate naturally distrusts patterns that hurt it)
5. Identity-bearing concept detection — patterns that map to specific named entities are quarantined automatically

This won't be perfect — pure peer-to-peer systems with open contribution always have an arms race with abuse. Mitigations are practical, not theoretical.

Phasing

What got dropped from the old Mother Maxim plan

The 2,224-line Mother Maxim plan was designed in the LLM-AUT-only world where memories (dialogue + episode traces) were the unit of value. The 2026-05-09 substrate-primary pivot reframed the architecture significantly.