Skip to content

Repository files navigation

Hyperfinite

Unbounded capability. Finite control.

Hyperfinite is a GitHub-native control plane for governed, model-assisted work. Models may propose bounded work; deterministic systems authorize it, bind every target, and execute only allowed effects. Independent humans retain administration, activation, approval, merge, release, and adoption authority.

See the authority boundary in five minutes

From an exact reviewed source checkout with locked dependencies installed, run:

npm run demo:authority

Finite animated terminal recording of the deterministic offline authority walkthrough: target-bearing model output is rejected, stale-head execution is refused, one fresh fake-provider comment is applied, and automation stops at Human Review

In one deterministic offline scenario, you will see target-bearing model output rejected, disabled and unleased work refused, current human evidence bind route and target, stale-head execution fail closed, one fresh fake-provider COMMENT applied, and automation stop at Human Review. The final counters remain zero for model calls, network calls, credential reads, and live effects.

Run npm run demo:authority -- --format=json for the canonical structured result. The complete accessible transcript and reproducible recording command uses the same executable scenario. The command is available in an authoritative repository clone, the verified control-plane-core customer-starter profile, and a reviewed full file-only customer copy.

This walkthrough uses fixed synthetic inputs and an injected fake provider. It performs no live inference, credential read, network call, GitHub mutation, approval, or merge. It is hermetic repository evidence, not live deployment or readiness evidence.

Who this is for

Reader Next document Decision or responsibility
Executive sponsor Customer FAQ Evaluation scope, outcome, funding, and stop/go decision
Evaluator or evaluation lead Customer evaluation guide Coordinate owners, tickets, timeline, evidence, and feedback
Architecture reviewer Architecture overview Review the control, trust, identity, and distribution boundaries
Security or identity owner Security index Review threats, OIDC, key custody, App permissions, monitoring, and incident controls
GitHub administrator or platform owner Customer administrator runbook Prepare the repository, Projects, App, rulesets, environments, services, and stores
Operator Portfolio operator runbook Validate, dry-run, canary, read back, pause, and recover
Legal or provenance reviewer Governance and provenance policy Review license, data use, retention, feedback, and evaluation terms
Contributor or reviewer Contributing Make contract-safe changes and independently review the exact head

Executive summary

Organizations want the speed of model-assisted software work without making a probabilistic model responsible for authorization, target selection, credentials, administrative changes, or irreversible effects. Hyperfinite separates those concerns.

Models may create bounded advisory artifacts. Deterministic code retains authority over identity, policy, lifecycle transitions, capabilities, targets, credentials, retries, effects, and evidence. Independent humans retain administration, activation, approval, merge, release, and adoption decisions.

The customer evaluation is designed to answer five questions:

  1. Can a customer run bounded model-assisted workflows in its own GitHub Enterprise Cloud environment?
  2. Can every target and effect remain bound to current authenticated evidence?
  3. Can missing, stale, ambiguous, or unavailable controls fail closed?
  4. Can operators understand cost, evidence, recovery, and human gates?
  5. Can the customer stop, recover, remove, or extend the evaluation safely?

The repository includes the deterministic authorization and lifecycle engine, policy compiler, capability catalog, GitHub adapter boundaries, disabled-by-default Agentic Workflow runtime, four governed demonstrations, durable local reference adapters, simulation and adversarial tests, packaging, customer-starter generation, Project planning, and administrator handoff evidence.

The repository intentionally stops before customer administrative effects. It does not create Projects, install a GitHub App, hold customer credentials, deploy trust services, enable paid inference, alter organization policy, approve or merge pull requests, or publish artifacts. Those actions require the customer owners and evidence listed in the evaluation guide.

Supported distribution

Public project history begins with a curated open-source snapshot. Earlier private development history is intentionally not published. This repository is the authoritative upstream for public development from that snapshot forward. Unpublished issues, pull requests, commits, or coordination records are not required to review, contribute to, or copy the public source.

Hyperfinite is distributed today as reviewed source, not as an installed product. Maintainers and local evaluators clone the authoritative repository and run its documented npm scripts. A verified customer-starter archive supports the bounded profile commands documented for that archive. A complete customer sandbox evaluation uses a reviewed full file-only copy in a new private customer-owned repository, then follows the customer evaluation guide. A starter profile does not imply support for the full sandbox matrix.

The private agentic-framework package metadata preserves the technical compatibility identity used by repository tooling. It exposes only package.json, has no bin entry, and provides no supported TypeScript SDK or packaged general-purpose CLI. Hyperfinite is not a hosted service or a deployable production service, and the repository does not bundle live administration, credentials, trust services, or effect authority. Any future SDK, CLI package, hosted offering, or production distribution requires separate product work. See the complete product boundary.

Evaluation maturity

Layer Included state Customer action
Repository contracts and hermetic demonstrations Deterministically testable at an exact head Independently validate the copied head
Customer-owned repository Portable workflows and source-derived repository identity Create a new private repo, initial commit, CODEOWNERS, and protections
GitHub Projects Closed schemas, target-manifest generator, dry-run plan, and readback Create four empty Projects and confirm exact target digest
GitHub App and OIDC Least-privilege permission and identity contracts Register/install the App and configure customer trust
Trust services and durable stores Interfaces, topology, local reference implementation, and fault tests Deploy isolated customer-owned services and stores
Agentic Workflow runtime Compiler-owned workflows, exact bindings, staged outputs, and fail-closed guards Configure protected variables and authorize a bounded window
Customer canary Synthetic inputs, stop-at-Human-Review contract, and evidence model Run one approved canary and evaluate the evidence
Broader adoption No automatic decision Customer governance decides scope, SLOs, support, and residual risk

First 30 minutes in a customer-owned sandbox copy

After copying the files into a new private customer repository:

  1. Confirm the customer repository uses main as its default branch.

  2. Confirm the origin remote points to the customer repository.

  3. Install the supported toolchain from config/v1alpha1/compatibility.json.

  4. Install locked dependencies:

    npm ci --ignore-scripts --no-audit --no-fund
  5. Rewrite .github/CODEOWNERS deterministically:

    npm run customer:configure -- \
      --codeowner @YOUR-ORG/hyperfinite-maintainers

    User-owned repositories may provide a user such as --codeowner @YOUR-USER.

  6. Run:

    npm run validate:customer-readiness
    npm run typecheck
    npm run build
    npm test
  7. Create the customer's initial import commit.

  8. Rebind the customer-starter selections to that new history:

    npm run customer:repin
    git diff -- config/v1alpha1/customer-starter-selection.json \
      config/v1alpha1/customer-starter-demo-portfolio-selection.json
  9. Confirm that only the two selection files contain the expected new source head and closure digests, then commit them.

  10. Push main and run git fetch origin main so exact-head tooling can verify the remote base.

  11. Review the resulting repository with the customer's security and platform owners.

  12. Open the applicable approval tickets.

  13. Keep AGENTIC_RUNTIME_ENABLED unset or not equal to true.

validate:customer-readiness scans every tracked or newly added file for source organization bindings, private network links, live Project IDs, and source issue/PR history references. It complements customer secret, dependency, code, history, privacy, and legal review.

Customer implementation path

1. Establish the customer trust boundary

  • Use a private customer-owned repository with a new history. Run npm run customer:repin from the clean initial import commit, review the two generated selection changes, and commit them before the full validation.
  • Assign an executive sponsor, evaluation lead, GitHub administrator, security owner, platform owner, identity/key owner, billing owner, operator, and independent reviewer.
  • Replace synthetic placeholders only in customer configuration. Keep test fixtures explicitly synthetic.
  • Define allowed data classifications, retention, feedback, incident, support, budget, and shutdown policies.

2. Validate the exact copied head

Run the complete validation matrix. Record the customer commit SHA and tool versions. Any later change requires fresh evidence.

3. Prepare four customer Projects

Have a human organization owner create four empty private Projects using the documented titles and Board layout. Export fresh admin snapshots and generate a target manifest:

npm run github:setup -- target-manifest \
  --catalog config/v1alpha1/demo-portfolio/catalog.json \
  --schema-root config/v1alpha1/demo-projects \
  --live path/to/fresh-admin-snapshots \
  --evaluated-at <current-RFC3339-time> \
  --output path/to/customer-project-targets.json

An independent human reviews every owner, repository, Project, and view identity, then records the manifest's contentDigest. The bootstrap planner requires that exact digest:

npm run github:setup -- bootstrap-plan \
  --catalog config/v1alpha1/demo-portfolio/catalog.json \
  --target-manifest path/to/customer-project-targets.json \
  --confirmed-target-manifest-digest sha256:<confirmed-digest> \
  --schema-root config/v1alpha1/demo-projects \
  --live path/to/fresh-admin-snapshots \
  --issue-bindings path/to/customer-issue-bindings.json \
  --evaluated-at <current-RFC3339-time> \
  --output path/to/reviewed-bootstrap-plan.json

The repository command performs no mutation. A human applies only the confirmed operations, exports fresh readback, and follows the two manual view steps in the Project setup runbook.

4. Register the customer-owned GitHub App

Use the exact operation-based permissions in GitHub App permissions. Install the App only on approved evaluation repositories. Keep its private key and short-lived installation tokens inside the customer token broker.

PAT, ambient-token, model-job credential, wildcard installation, and success-shaped fallback paths are unsupported.

5. Deploy customer trust services

Deploy separate identities for:

  1. webhook verification and fresh binding resolution;
  2. runtime-state publication;
  3. OIDC authorization redemption and budget reservation;
  4. threat, DLP, policy, and evidence signing;
  5. conditional evidence and operation-grant storage;
  6. operation-scoped GitHub App token brokerage;
  7. the Single Writer (the serialized workflow Effect Plan executor) and reconciliation; and
  8. the isolated verification runner and installation/release adapter.

See the deployment prerequisite matrix for owners and acceptance evidence.

6. Configure customer controls

Configure:

  • CODEOWNERS and an independent-review ruleset;
  • required checks and current-head approval;
  • selected and SHA-pinned Actions;
  • protected environments and runtime variables;
  • GHAS, CodeQL, Dependency Review, Dependabot, and secret scanning according to customer policy;
  • model entitlement, budget ceilings, alerts, and shutdown owner;
  • logging, evidence, backup, and retention;
  • monitoring and incident routing; and
  • key rotation, revocation, kill-switch, and disabled restore drills.

7. Run synthetic evidence

With live activation still disabled:

npm run canary:synthetic
npm run handoff:administrator

The administrator handoff is customer-safe and uses a synthetic-unconfigured readback. It does not contain or depend on source-organization observations.

8. Authorize one customer canary

Only after every ticket and prerequisite is complete, authorize one time-bounded synthetic run. It may create a draft pull request, produce current-head COMMENT review evidence, and stop at Human Review. It cannot approve, merge, deploy, publish, change billing, or administer the organization.

9. Evaluate and decide

Review business outcomes, operator effort, security findings, permissions, residual risk, recovery, evidence, model usefulness, cost, support, and customer feedback. Customer governance decides whether to stop, extend, redesign, or adopt.

Authority model

The Work Accord is the versioned, non-authoritative contract for one work item: requested outcomes, constraints, exact bindings, and applicable Phase Contract digests. A Phase Contract separately defines one approved active phase's entry, output, evidence, budget, and deterministic exit rules. Neither contract grants authority by itself. The Control Kernel is the pure deterministic state-transition and authorization engine. Trusted Binding derives the exact repository, work item, revision, and head from current authenticated facts rather than model output. The Single Writer is the only component permitted to execute an authorized workflow Effect Plan after one final fresh read.

Authority is ordered and non-interchangeable:

  1. lifecycle graph;
  2. Work Accord and Phase Contracts;
  3. policy compiler and Capability Registry;
  4. Control Kernel;
  5. trusted adapter;
  6. Single Writer; and
  7. model output.
flowchart LR
  H[Authorized human] --> GH[Current GitHub facts]
  GH --> B[Trusted Binding]
  B --> A[Work Accord and policy]
  A --> K[Control Kernel]
  K --> C[Capability Registry]
  C --> M[Bounded model]
  M --> O[Target-free output]
  O --> T[Trusted output adapter]
  T --> K
  K --> E[Target-bound Effect Plan]
  E --> W[Single Writer]
  W --> GH
Loading

GitHub Projects are visible projections, never lifecycle authority. Model output cannot choose a repository, issue, pull request, Project item, stage, route, capability, path, credential, retry, effect, approval, or merge.

Governed demonstration portfolio

Demo Journey Documentation
App Modernization Intake, discovery, assessment, architecture, migration, draft implementation, verification, Human Review Guide · Example
Feature Delivery Intake, requirements, discovery, design, planning, build, test, Human Review Guide · Example
Security and Dependency Remediation Intake, triage, reproduction, design, draft patch, security verification, Human Review Guide · Example
Adaptive Delivery Fixed context, selectable discovery, synthesis, planning, selectable implementation, exact-head verification, Human Review Guide · Example

Every fixed model stage has one globally exclusive (demoProjectId, stageId, agentId, capabilityId, workflowId) binding. Adaptive Delivery has two reviewed user-selectable stages with exact static candidates. A Project choice remains untrusted intent until deterministic policy validation issues one signed exact-agent grant.

Hands-off simulation stops at Human Review. A separate synthetic-human continuation demonstrates terminal completion without granting automation human authority.

Implemented surfaces

Surface Included behavior
Lifecycle, Work Accord, Phase Contracts, policy, Capability Registry, Control Kernel, receipts, migrations Deterministic and locally validated
GitHub event normalization, Trusted Binding, target-free translation, Effect Plans, credentials, Single Writer Implemented behind injected and trusted ports
Agentic Workflow sources, generated locks, stage agents/skills, pre-activation, review, execution bridge Disabled by default; repository-relative targets; staged output
Demo contracts, projection, runtime, simulation, observability, hardening Complete hermetic portfolio
Marketing and Business Operations packs Synthetic repository proposal paths
Release, migration, installation, and customer-starter tooling Deterministic build, verify, plan, and offline validation
Project setup Customer target-manifest generation, dry-run planning, confirmed digest, and post-apply reconciliation
Administrator handoff Plan, separate confirmation, one-attempt contract, readback, and generic gap report
Customer readiness Full-file scan for source-specific or private material

Complete validation matrix

Use the toolchain in config/v1alpha1/compatibility.json. From a clean committed customer head, npm run validate runs the repository matrix below except dependency audit, synthetic canary, and administrator handoff.

npm ci --ignore-scripts --no-audit --no-fund
npm run validate:customer-readiness
npm run validate:technical-identity
npm run typecheck
npm run build
npm test
npm run validate:schemas
npm run validate:runtime
npm run validate:eval-fixtures
npm run validate:provenance
npm run validate:workflows
npm run validate:gh-aw
npm run validate:packaging
npm audit --audit-level=high
npm run validate:demos
npm run simulate:demos
npm run validate:hardening
npm run canary:synthetic
npm run handoff:administrator

Core validation is deterministic and offline after dependency installation. validate:gh-aw uses the pinned public compiler release and recompiles the workflow Markdown. Never hand-edit .lock.yml files or .github/aw/actions-lock.json.

Portability and customer-data boundary

  • Workflow repositories resolve from trusted ${{ github.repository }} context.
  • Release and customer-starter manifests bind both the canonical Git host and lowercase repository name derived from origin.
  • Checked-in Project, repository, issue, App, and service identities are synthetic.
  • Customer target manifests are generated from fresh customer snapshots and require independent digest confirmation.
  • Source issue, pull-request, and commit history is not required by a customer copy.
  • Live customer IDs, credentials, exports, and evidence remain outside the repository.
  • The MIT LICENSE is preserved byte-for-byte.

Documentation

Use the documentation index for the complete audience and narrative map. Repository reference documentation not repeated there: Examples · Configuration · Schemas · Source map · Tooling · Tests and evidence. The role table above keeps each first reader's next document one click away.

Repository layout

Path Purpose
src/ Deterministic kernel, adapters, runtime, evidence, packaging, readiness, and validation libraries
config/v1alpha1/ Reviewed lifecycle, policy, capability, demo, Domain Pack, and packaging configuration
schemas/v1alpha1/ Closed JSON Schemas for persisted and exchanged contracts
.github/workflows/ Agentic Workflow Markdown and compiler-owned generated locks
.github/agents/, .github/skills/ Exact runtime agents and capability-bound skills
scripts/ Validation, simulation, setup planning, customer readiness, installer, and release tools
tests/ Positive, adversarial, replay, fault-injection, portability, and integration tests
examples/ Synthetic fixtures and customer planning examples
docs/ Architecture, decisions, operations, security, governance, and demo guidance

Support, feedback, and contribution

Read Support, Contributing, Security policy, and Governance before opening a change or customer feedback item.

Customer feedback should include the copied version/commit, affected stage, expected and observed behavior, typed reason codes, redacted evidence digests, business impact, and desired outcome. Do not include credentials, customer source, private URLs, identities, prompts/responses, or confidential logs.

About

GitHub-native control plane that keeps model output advisory and execution authority deterministic.

Topics

Resources

Contributing

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages