HiveMind Runtime

One governed vertical, end to end

HiveMind is the runtime that turns a small model into a trustworthy system: deterministic rules first, bounded neural proposals second, independent verification always.

Architecture

The vertical

A task travels one path. Each boundary owns exactly one thing, and none of them trusts the previous stage.

Hard boundaries

The model owns only inference-private state. It cannot authorize memory, learning or continuity.
Hyphae owns durable knowledge, evidence bindings and recovery. Retrieved text is advisory; verified source bindings are authoritative.
The runtime owns lifecycle, timeouts, context assembly and typed receipts. Verification is independent of the model that proposed the answer.

01 TaskEnvelope

02 ELIZA executive

03 Low-bit navigator

04 Hyphae retrieval + proof

05 Verifier

Contribution + receipt

ABSTAIN + receipt

Execution modes

Four ways to run, one contract

The same governed loop runs across privacy postures. Credentials resolve at runtime from the OS keychain — never from prompts, weights or config files.

ONLINE

Local inference + Hyphae Cloud. Fresh knowledge with tenant-scoped authorization.

DEGRADED_OFFLINE

Local inference + verified offline cache. Useful answers with proofs, no network.

PRIVATE_LOCAL

Local inference + local or private Hyphae. Nothing leaves the machine. Air-gap friendly.

EPHEMERAL

No personal-memory reads or writes. A session that leaves no trace by design.

Virtualized context

Not an infinite prompt — something better

MARS does not promise infinite context. It provides resumable access to durable memory: a small working set, iterative retrieval, explicit citations and checkpointed task state.

16K

Normal working set

Policy, task state, recent dialogue, top-k evidence, output reserve — everything a bounded step needs.

32K

Escalated working set

For genuinely wide steps, with explicit budget accounting and no silent truncation.

128K

Exceptional audit path

Full physical capacity, reserved for audits. Context quality is measured by evidence recall — not advertised token length.

128K is physical capacity, not a promise of comprehension. Long tasks checkpoint verified claims and source IDs — they do not grow the prompt forever.

Teacher lane

Frontier models — only when it counts

Larger GPU models act as teachers and escalation services. Their outputs cross the same verifier and the same policy boundary as everything else.

01 Escalate the hard 5% instead of paying frontier prices for the routine 95%.
02 Teacher claims remain hypotheses until verified against their sources.
03 No teacher writes authoritative memory directly. Ever.

Local-first API

A versioned contract on loopback

The backend binds only to 127.0.0.1 and speaks a stable OpenAPI 3.1 contract. Any professional frontend integrates against these eight endpoints — user-facing messages in English and Spanish.

GET /v1/health Liveness of the local runtime
GET /v1/runtime Runtime identity: artifact SHA-256, posture, languages
POST /v1/knowledge/inspect Preview a document before ingesting — digest, size, warnings
POST /v1/knowledge/commit Ingest a document into the local knowledge plane
GET /v1/knowledge/sources Inventory of committed sources with content digests
DELETE /v1/knowledge/sources/{digest} Remove one exact content-addressed source
POST /v1/chat Governed question → decision, answer, sources, channels, receipt id
GET /v1/receipts/{id} Fetch the canonical receipt for any response

ask — with a receipt back

$ curl -s http://127.0.0.1:8765/v1/chat \

-H "Content-Type: application/json" \

-d '{"request_id":"42","prompt":"…","locale":"en"}'

{"decision":"SUPPORTED", "receipt_id":"1d0f…"}

Loopback-only by design: external origins are rejected, secrets never appear in the contract, and a rejected verification returns a typed error — not a hallucination.

Intelligence you can audit.

MARS, HiveMind and Hyphae are being built in the open by Celiums Solutions — local-first, fail-closed, and receipted end to end.

Explore the technology See the use cases