Accountable Surface

A diamond lattice with every other cell filled with fine lines, bulging outward as if seen through a lens, with a bright core in the one open cell at the center.A diamond lattice with every other cell filled with fine lines, bulging outward as if seen through a lens, with a bright core in the one open cell at the center.

workbench that allows only approved AI agent actions

Accountable Surface lets an AI agent take only the file, command, web, or browser action a person has approved. It checks the request and authorization, blocks or pauses when needed, verifies the outcome, rolls back reversible failures, and records decisions and outcomes in a journal. Persisted journals are hash-chained so later edits, deletions, or reordering are detected.

Inspect sourceRun or verifyRead evidence
Status
active
Access
install
Release
0.3.1 on PyPI; latest GitHub release v0.2.0
Checked
2026-09-26

Listed in the catalog under

Who it is for

Audiences

  • authorized security teams
  • evaluation engineers
  • agent-system engineers

Deployment contexts

  • owned or expressly authorized systems
  • review and CI lanes
  • controlled agent workflows

Inputs and outputs

Inputs

  • perceived interface structure
  • externally supplied action grants
  • bounded effector requests
  • rollback policy

Outputs

  • authorized effects or denials
  • post-action re-perception
  • rollback results
  • durable action journal

Run or evaluate

Entrypython -m pip install accountable-surface==0.3.1
VerificationFrom a source checkout at tag v0.3.1: python -m pip install -e ".[test]"; python examples/demo.py; python -m pytest

Inspect the public source

Architecture and relationships

Related records

  • Flywheelverified-inference platform and desktop client
  • EMETchecker that content still matches its source
  • behavior-transform.iolocal wrapper for file, web, and tool input and output

Accountable Machines is the companion theory route, not this product route.

Architecture, dependencies, and relationship notes

Dependencies

  • documented sibling source packages for enabled capabilities
  • coherence_membrane sibling source package for clean wheel import

Relation-backed claims

  • Flywheel includes Accountable Surface as a lane
    Flywheel 1.0.4 declares Accountable Surface 0.3.1 as an actuation lane with additional Coherence Membrane and Proof Surface source roots.

Current evidence

accountable-surface 0.3.1 on PyPI

Verified

How we know

PyPI serves accountable-surface 0.3.1, uploaded on 22 September 2026 (Pacific time). A clean virtual environment installed it from PyPI with its dependencies coherence-membrane 0.1.0 and proof-surface 0.1.0, and both packages imported. Tag v0.3.1 exists on GitHub with no GitHub release attached.

Accountable Surface v0.2.0

Verified

How we know

The v0.2.0 GitHub release was published on 17 September 2026 (Pacific time) and is the latest GitHub release. PyPI carries the newer 0.3.1.

Accountable Surface v0.1.0

Verified

How we know

The v0.1.0 GitHub release is public, not draft, and not prerelease. It targets commit 9ebc4519e92ce674ac9058e662e655842c69be42 and publishes accountable_surface-0.1.0-py3-none-any.whl SHA-256 9f08e87868a755141e515a7b0e99be67caee3770f33d8adc11506effa8ac1e52, accountable_surface-0.1.0.tar.gz SHA-256 ac655bf23dee3714c7ee59c87fae98c09a75dcaaf0244bbb641e04d3d05eaa9b, and accountable-surface-010-release-SHA256SUMS.txt SHA-256 a28e4e2b37baf1899137db681a2ea7387e48577eb452cff69e41605b71e59c84. The release notes require the repository's documented source setup because some capabilities need sibling source packages. A clean wheel install succeeded, then importing accountable_surface failed on missing coherence_membrane. This is release evidence, not self-contained package readiness or real Windows native-app acceptance.

Accountable Surface source 0.1.0

Verified

How we know

The public repository and v0.1.0 release notes identify source 0.1.0. The source setup includes sibling source dependencies; self-authored tests are not independent review.

Limitations and authorization boundaries

  • The model cannot supply its own authorization.
  • Real Windows native-application acceptance remains unverified; release validation used a fake UI Automation driver.
  • Self-authored tests do not establish independent adversarial review or production safety.
  • The v0.1.0 GitHub wheel was not a self-contained install. PyPI 0.3.1 installs its coherence-membrane and proof-surface dependencies from PyPI; the FastMCP server entry, accountable-surface-server, needs the [server] extra; the stdio entry accountable-surface-mcp, which Flywheel 1.0.4 launches, does not.

Authorization boundary. Every effect requires authority external to the model and remains limited to the granted scope.