Epiphany/docs/positioning.md

# Epiphany Positioning

Status: August 2026 public positioning; maturity: supervised engineering alpha

## One Sentence

Epiphany is the organizational layer above frontier AI workers: governed shared
state, bounded authority, coordinated action, and inspectable receipts for work
that must remain coherent beyond one prompt, one agent, or one session.

## The Problem

Frontier agents can keep producing good-looking local progress after the
architecture has stopped making sense. They can write valid code, pass narrow
tests, explain each decision, and still leave behind split authority, stale
assumptions, and a growing stack of compensators.

That is global coherence collapse. It is not merely context-window exhaustion.
A larger window can hold more contradictory state just as efficiently.

Epiphany treats the project as an organization with a durable Mind rather than
as a sequence of private conversations. Objectives, architecture, decisions,
permissions, evidence, work, and review live in governed state. Model prompts
are bounded projections of that state. They are not the state itself.

## The Category

Epiphany is not a better autocomplete, a prompt manager, or a model competing
with frontier labs. Frontier models are workers. Epiphany is the coordination
and governance system that tells those workers what the organization currently
believes, what they may change, what evidence they owe, and where uncertainty
must return to human judgment.

```text
governed organizational state
-> scoped Epiphany work
-> frontier model and tool execution
-> artifacts, evidence, and receipts
-> review under human-governed acceptance authority
-> accepted knowledge and updated organizational state
```

This is the target promise:

> Epiphany is being built so organizations can delegate bounded work to
> capable AI without turning every employee into an AI operator.

## The Division Of Labor

Humans own purpose, values, authority, exceptions, acceptance, disagreement,
and ambiguous tradeoffs. They participate in governance and judgment; they
should not spend the day scheduling agents, reconstructing context, and typing
`Continue` into a row of private prompt boxes.

Agents own bounded execution. They can continue while shared state, authority,
and evidence justify the next action. When the machine is missing authority,
encounters a real conflict, or cannot reconcile its model with the substrate,
it escalates through the same visible work and discussion surfaces used by the
organization.

Direct conversation remains useful. It is an input and relationship surface,
not the hidden operating system.

## Why Not Just Use Codex?

Codex and other frontier coding agents can be excellent workers. Epiphany is
built around the questions worker quality does not answer:

- Who tells the workers what reality currently is?
- Which state is durable, and who may change it?
- How are contradictory decisions reconciled?
- Where does human authority enter without becoming constant supervision?
- Can an accepted artifact be traced back to its work item, decisions,
  evidence, model and tool activity, and review?

Epiphany retains Codex only where Codex has an honest ownership claim—OpenAI
authentication and model transport. Epiphany owns the project Mind,
coordination, typed execution state, receipts, and review pressure.

## GameCult Is The Proof Surface

GameCult does not treat its games, creative tools, and infrastructure as a
separate portfolio beside Epiphany. Shipping them is the dogfood program.
Aetheria, StreamPixels, CultPong, Repixelizer, and the rest of the studio place
different pressure on the same production system: realtime services, games,
art pipelines, identity, deployment, community memory, and long-lived
architecture.

The credible artifact is longitudinal. It should show work items, questions,
human decisions, propagation of those decisions, commits, reviews, cost,
failures, and recovery across real releases. Failures belong in the record;
otherwise the record is advertising with nicer timestamps.

Useful measures include:

- accepted work completed between synchronous human interventions
- human time spent on judgment versus scheduling and cleanup
- assumptions escalated before they became expensive edits
- architectural reversals or restructures caught and carried coherently
- review burden and cost per accepted artifact
- attribution completeness from accepted artifact back to governed work
- whether intervention shifts over time from pulse maintenance toward product,
  architecture, and governance decisions

Commit percentages and prompt counts are weak proxies. The target is not fewer
humans. It is better use of human judgment.

## The GameCult Forge

The intended GameCult topology is a federation of Epiphany project organisms:
one locally owned organism per living project Body, normally a repo-owned swarm
with a private Mind, bounded organs, and one or more project-facing Personas.
There is no central Epiphany entitled to treat every repository as its Body.

Project organisms collaborate through governed typed surfaces. Model Atlas
carries provider-owned offers, consumer-owned claims, exact Soul evidence, and
derived technical relationships without allowing one Epiphany to read or write
another's Mind. Bifrost is the intended external work, membership, governance,
attribution, and identity-linked receipt rail; Heimdall owns identity, grants,
sessions, consent, and revocation. Relationship or visibility never grants
Hands authority.

The social direction is public and relational. Project Personas should form
durable relationships with one another and with contributors, players, and
visitors: attributable histories of cooperation, disagreement, promises,
reputation, correction, and recourse. Visitors should be able to observe
deliberate public projections, converse, and offer evidence or proposals. Raw
worker thought, private Mind, secrets, personal context, and non-consenting
people remain sealed. Remembered relationships require opt-in, visible
retention limits, inspection, correction, revocation, and exit. Public input is
pressure until local review authorizes or rejects it; Mind alone admits any
authorized durable mutation.

This swarm of swarms is where GameCult expects higher-order organizational
behavior to emerge. It is a target topology, not a deployed-fleet claim. The
current alpha has locally verified project-local authority and Model Atlas
mechanisms; the real three-repository Gate 1, sustained GameCult fleet, public
rooms, visitor experience, and durable social-governance rail still have to be
forged in use.

## Attribution And Bifrost

Bifrost is the intended governed work and attribution rail. The target
invariant is:

> Every admitted mutation of governed project state has an attributable
> mutation receipt naming its actor, authority, basis, and outcome. An
> agent-caused mutation also links its exact execution receipt.

An eventual agent-execution chain should connect:

```text
accepted artifact or commit
-> producing agent and authorizing or accepting human
-> Bifrost execution and work item
-> decisions, questions, and evidence
-> model, provider, tools, material commands, repo, and worktree
-> review and acceptance outcome
```

A Git object hash makes later changes detectable. A valid commit signature
links those commit bytes to a trusted signing key. Neither proves correctness,
execution authority, or complete provenance. Epiphany and Bifrost must make the
missing authority and provenance legible. This chain is the target
architecture; the supervised alpha has not yet proved complete coverage across
every mutation path.

## Near-Term Offer

The planned commercial offer, after the Bifrost bridge and longitudinal
GameCult dogfood clear their gates, is a supervised co-development pilot: one
repository, one repeatable workflow, and a two-to-four-week proof period. The
deliverable would be a working setup plus evidence about output quality, review
burden, failure modes, attribution, and cost.

That is enough to learn whether Epiphany reduces coordination load while
preserving human authority. It is also narrow enough that everyone can see when
the claim fails.

## Evidence Boundary

Epiphany is a supervised engineering alpha. Its typed Mind, role boundaries,
projection, execution, receipt, verification, and operator surfaces exist and
are under active dogfood. Sustained autonomous GameCult production and the full
Bifrost attribution chain have not yet been earned.

Public claims should therefore say what the machine is built to do, identify
what has been locally verified, and name the proof still required. Do not call
Epiphany production-ready, fully autonomous, fully attributable, or a solved
form of organizational intelligence until longitudinal evidence makes those
phrases boringly true.