HivemindOS manual

Founder Mode, Hivemind Labs, And Proof Packs

Founder Mode turns one outcome into a reviewable Zero Human Company blueprint.

Describe what you want to make happen, then choose a privacy posture, first-milestone budget, and pace. HivemindOS proposes the company charter, measurable goal, crew, capabilities, compute routes, approval boundaries, first experiment, and proof requirements. Blueprint generation is read-only. Founding the company requires an explicit action, and founding does not launch autonomous work.

Founder Mode proposes the default HivemindOS crew path. After founding, you can edit the company and optionally choose an AEON background skill instead. AEON-backed companies do not require a native crew, but they do require a saved AEON workspace and a skill available in that workspace.

What The Blueprint Includes

a company identity, charter, and measurable apex goal
a first milestone with concrete success criteria and deliverables
suggested crew roles matched to configured agents
capability readiness, credential key names, fallbacks, and side-effect gates
local, Hive Compute, and hosted model routes appropriate to the privacy preference
milestone, daily, and monthly budget limits
approval policies for public publishing, customer contact, money movement, and destructive actions
a private Hivemind Lab for the first measurable experiment
proof-pack requirements for trusted completion

Operators can still found or import companies manually. Founder Mode is an outcome-first path through the same local company, Work Board, Shared Brain, approval, and wallet controls.

Hivemind Labs

Each company can run Hivemind Labs from its cockpit. Labs use the same evidence and lineage model as Agent Challenges:

State a bounded objective and metric.
Record candidate hypotheses and run requests.
Attach measured results and evidence.
Add a verifier or human ruling.
Record the reusable operating lever and any observed failure mode with the result.
Keep statistically tied results on the frontier and refine the playbook.
Preview the generated shared skill and the capabilities selected by Hive Skill Fusion.
Review its evidence, failure modes, and existing-skill change summary before explicitly publishing it.

When a Lab has measured evidence, independent verification, and a distilled playbook, HivemindOS marks its capability-promotion draft as reviewable. Preview fused skill runs capability discovery and generates the complete SKILL.md without writing it. The review shows selected capabilities, evaluation records, evidence provenance, failure modes, and whether publishing would create, update, or leave unchanged an existing Lab skill.

Publishing never happens automatically. The operator must affirm that they reviewed the draft and its replacement impact before Fuse into shared skill writes it into the shared brain. Confirmation is bound to that exact generated draft; if evidence or available capabilities change, HivemindOS requires a new preview and review. Collecting and candidate Labs cannot enter this path. A later Founder Mode run can discover the promoted skill through the same capability index, creating a reviewed learn-and-reuse loop across companies.

Outcome-Aware Model Routing

Adaptive model routing can use local outcome records in addition to model metadata and transport reliability. Confirmed results may include the task type, provider and model, acceptance, quality, cost, latency, privacy posture, and a related proof-pack identifier.

These records are local operational evidence. They are not official hosted reputation, marketplace payout authority, or a commercial entitlement. Hosted trust and commercial policy remain server-authoritative.

Proof Packs

The Work Board task detail view includes a Show me why proof pack. It combines:

recorded deliverables and the agent result
passed, failed, and skipped eval gates
evidence attached to those checks
producing agent and machine provenance
signature-verified work receipts and linked project proofs
explicit gaps when something remains unsigned or unevaluated

Proof packs distinguish verified, evaluated but unsigned, needs attention, and unverified outcomes. They do not turn a model statement into a verified fact merely because an agent produced it.

Safety Boundaries

Compiling a blueprint is read-only.
Founding creates local company and Lab state but does not start autonomy.
External communication, public publishing, money movement, and destructive actions default to human approval.
Capability promotion remains preview-only until the operator explicitly confirms a reviewable Hive Skill Fusion draft.
Local outcome records cannot create official marketplace reputation, balances, payouts, or entitlements.
Operators can stop autonomy or freeze company spending through the existing company controls.