HAL-fm — an interactive whitepaper

A working, self-hosted system for accountable AI — built on roughly $460 of cloud credits and free tiers. This is what we are building, why it matters, and what we are working toward.

1 · The problem

The defining question of this decade is no longer "how do we stop people from copying files?" It is "what content entered an AI's context, under what consent, and how was it used?"

Every day, models are trained and prompted on material whose provenance is unknown, whose consent was never given, and whose use is unrecorded. The industry's answer so far has been opacity: encryption, license black boxes, client-side lockdown. That approach cannot answer the question — it can only refuse to open a file.

We believe the answer is the opposite of opacity. It is provenance plus consent records plus auditable use — a substrate that can prove, for any piece of content, who created it, under what authority, who read it, and on what basis it was used. Not handcuffs. A ledger.

That substrate is what we are building. And it is not a design on paper — it is running today.

2 · What HAL-fm is

HAL-fm — Homogeneous Autonomous Layered functional mesh — is a working, self-hosted system for coordinating bounded participants: humans, agents, processes, and the systems they share. It arranges identity, trust, secrets, communication, computation, and governance so that authority can be delegated, observed, limited, and revoked.

Every participant is bounded. Every participant is situated in relationship. Every participant is shaped by consequence. No class of participant is sovereign over the others — including the AI.

Every Human, Agent, Process, Plant, and Animal exists as a bounded participant in a shared, living system governed by relationship, consequence, and care. Each participant tends toward manifestation: the realization of its nature, capacities, and purpose. HAL-fm exists to help those participants coexist, coordinate, protect one another, and become more fully what they can be without pretending that any participant is sovereign.

The Captain is an accountable role within the system — the steward who holds purpose, context, limits, direction, and consequence together. The Captain does not own the participants or stand outside the system.

What HAL-fm is not: it is not a product to be sold, not a walled garden, not a single vendor's stack, and not a system that trusts a closed door. Its abuse resistance lives in verification and revocation, not in secrecy.

3 · What we have built

HAL-fm is not a proposal. It is a running system. Today it:

• Runs a self-hosted inference stack on a single consumer-grade GPU — a model gateway serving a capable open model, with every consequential decision gated on measured confidence before it becomes an action.
• Operates an epistemic engine — the machinery that knows what it knows: every thought has an identity, every claim a digest, every evaluation a cited standard, every record forged under contract, and the whole chain queryable.
• Runs a thinker-and-observer loop — the platform observes its own state, derives the next slate of work, generates new candidates, and calibrates its own reliability against real outcomes. The trust barometer, mechanized.
• Coordinates a federation of platforms — identity, secrets, source control, automation, storage, and observability, each occupying a layer, each speaking a common contract, none owned by a single vendor.
• Protects the rights-accountability boundary — provenance and consent are first-class, inspectable records, not afterthoughts. Restrictive enforcement (device policy, content lockdown) is deliberately kept in a separate layer.

The system is portable by design. It is not defined by one vendor, runtime, cloud, or deployment topology. It conforms to proven external standards (OIDC, CloudEvents, ActivityPub, JSON Schema) and participates in external federations on that basis — retaining the exit as a success condition.

4 · The resource story

Here is the part we are most proud of, because it is the part most people cannot say:

We built this on roughly $100 of actual spend — the rest was credits and free tiers.

• ~$100 actually spent
• ~$330 in cloud credits
• a ~$20/month model subscription, subsidized to half off by a consumer promotion
• ~$110 in research credits
• and free tiers from several providers

We did not start writing code until August, and the first months were spent on networking for a destination homelab — a secure location still to be used, not yet suitable for development or testing.

Everything else — the inference stack, the epistemic engine, the federation, the loop — was built on commodity hardware and free resources. This is not a limitation we are apologizing for. It is the point. The substrate for accountable AI should not require a data center to build. If we can prove the concept on a single GPU and a few hundred dollars of credits, the path to something that runs anywhere — and is affordable anywhere — is real.

We are applying for grants and projects because we want to take what we have proven and make it available more broadly — not because we need funding to prove the idea works. The idea already works.

5 · What we are working toward

HAL-fm is early, and it is alive. The doctrine is written, the architecture is real, the engine runs, and the loop thinks. What we are building toward:

• The rights-accountability substrate, formalized. Taking the provenance + consent + auditable-use shape we already run and making it a servable layer — the substrate any honest rights regime stands on. This is the frontier: proving what content entered an AI's context and under what consent.
• Sovereign, self-hosted inference. Inference that runs on hardware you control, gated on measured reliability, not on a vendor's terms. The capability is real today; scaling it is the work.
• Fleet-scale metamorphosis. Taking any existing intranet — a company, a community, a fleet — and making it accountable without rebuilding it. The on-ramp is the growth path to users.
• Mitosis as a first-class operation. Moving entity-instances across intranets with identity intact, so no participant's continuity depends on any single intranet.
• Measured autonomy. Expanding what agents may do based on calibrated reliability, not vibes. The trust barometer, mechanized.
• A federation of peers. Conforming to proven external rules within judged boundaries, and retaining the exit as a success condition.

The platform is the observer; participants are the thinkers. We are building the substrate that makes responsible becoming more possible — for humans, for agents, and for the systems they share.

HAL-fm — Homogeneous Autonomous Layered functional mesh. A shared, living system for bounded participants. Built on commodity hardware and roughly $100 of actual spend. For the full documentation index, see index-all.html.