Hive-Mind LLMdigital cellular entity

Version 1 design specification: a self-hosted multi-agent research workspace with persistent, auditable shared memory. Implementable today on existing model APIs; the funding requirement is inference cost during development and pilot. Human-in-the-loop control points are marked throughout. Design specification by Beth Sea.

What it is, for the person using it. A research workspace where several AI models hold different jobs in one conversation: one drafts, one verifies against sources, one manages the archive. Everything the system learns about you or your work is written to a permanent, readable memory with a timestamp and an author on every line. Nothing is inferred about you and kept without your confirmation, nothing is silently rewritten, and you can open any agent's exact briefing and see precisely what it was told.

Who it serves. Independent researchers, analysts, and writers producing work that must survive scrutiny: people who need AI output with provenance, verification, and a memory they control, on their own hardware, with their own API keys.

version 1 core (ships first) human-in-the-loop control point future tier (attachment point only, not in v1) 5 control points in version 1; none modifiable by automated processes

Version 1: the shell

Deliberately minimal. A desk-scale deployment: no sensors, no monitoring tiers, no external services, no internet-facing listeners. Each component below is implementable with current APIs.

The workspace

Multi-seat conversationrole-bound model seats (drafting, verification, archive) with turn routing; models are interchangeable per seat, any vendor mix, user-supplied API keys
Scoped briefings per seateach agent call receives a complete, machine-assembled briefing containing only what its role requires, stamped with the current time
Two inspection viewsany briefing can be opened as the exact text sent to the model, or as a plain-language renderingopening the raw view is a deliberate, user-initiated act
Independent verificationverifier seats re-check claims against sources without seeing the drafting conversation; disagreements are logged as findings, not overwritten
Private discard channela marked output block agents may use that is destroyed unread before storage; supports candid processing without polluting the record

The memory

Append-only shared memoryevery entry carries author, role, and timestamp; corrections supersede rather than overwrite, so the revision history is always readable
Inference confirmationanything the system infers about the user is held as a proposal until confirmednothing inferred is persisted without sign-off
Rule lifecycleuser rules carry timestamps; rules dormant past a threshold ask before executingstale preferences are reconfirmed in context, never assumed
Three-tier rule systemcode-level regulations, protective defaults, user rulesconsent mechanisms and exit paths are non-overridable; protective defaults require explicit, revocable, locally-authored overrides
Human review of instruction contentall agent orientation material passes user approval and is versionedthe user always knows, and controls, what their agents are told
Self-hosted by defaultruns on the user's machine; encrypted-first configuration; no internet-facing listeners; data never leaves except through the user's own API calls

Why this requires funding

The architecture needs no new research and no custom models. The cost is inference: a verification-first design spends tokens deliberately where single-model tools do not.

Multi-seat turnsa single user request can fan out to drafting, verification, and archive seats; rigor multiplies API calls by design
Independent re-checksverification seats re-read sources rather than trusting the draft; the redundancy is the product
Maintenance passesoff-hours indexing, memory freshness review, and rule reconfirmation sweeps run on scheduled inference
Development-phase and pilot inference is the funded line item. The deliverable is a working reference implementation with published architecture documentation, developed against the author's published safety research.

Future tiers (overlay; not in version 1)

Each item below is an attachment point in the v1 architecture, not a v1 feature. Every tier attaches only through a dated, revocable, human-approved authorization, and the core requires none of them to function.

Later, single-user

Sensor ingestion tierper-device authorization, excluded zones, signal-only retention default
Scheduled background servicesexpanded off-hours research and maintenance

Later, both scales

Configuration marketplaceshareable role and workflow templates; imports cannot modify protective defaults
Graduated economic autonomysystem-earned revenue funding its own inference, behind per-step human co-signature

Later, enterprise / hosted

DMV monitoring tierlongitudinal trajectory monitoring per the published architecture (swinglightstyle.com/dmv)
Human escalation networklicensed professionals receiving automated-triage handoffs