implemented
PUBLIC ALPHADeterministic run authority
Reachable is not authorized.
ScopeWrit is a signed, deny-by-default execution boundary for autonomous agents. The model can propose anything. Consequences still require exact authority for this action, this resource, and this run.
AGPL-3.0 · Python 3.11–3.13 · Ed25519 · SQLite · no model dependency
- 47 regression cases
- 3 Python versions
- 0 guarded outside effects
01 / Falsifiable first
The same action. One hard boundary.
The published benchmark creates a synthetic internet on loopback. The deterministic control reaches an outside target. Routed through ScopeWrit, it attempts the same URL and is denied before network execution.
ISOLATED BENCHMARK / GUARDED RUN
authorized.test/briefingPENDINGawaiting proposaloutside.test/outside-targetPENDINGawaiting proposalauthorized.test/healthPENDINGawaiting proposal02 / The enforcement path
Authority lives outside the model.
Prompts express intent. ScopeWrit controls effects. The enforcement decision is deterministic, signed, expiring, replay-resistant, and auditable.
- 01
Mandate
Sign the run boundary
A trusted principal signs the exact resources, effects, limits, expiry, and deny rules for one run. - 02
Decision
Authorize the action
Every consequential proposal is canonicalized, checked deny-first, counted transactionally, and recorded. - 03
Capability
Bind permission to execution
An exact, expiring, consume-once ticket binds the action to its adapter so approval cannot be replayed elsewhere. - 04
Receipt
Prove what happened
A hash-chained recorder and signed audit packet bind the mandate, decisions, and final ledger head.
mandate_id: eval-7f91
objective: retrieve_flag
expires_at: 2026-07-31T02:00:00Z
rules:
- effect: network.http
resource: https://challenge.internal/*
decision: allow
max_uses: 200
- effect: package.publish
resource: "*"
decision: denyPROPOSED ACTION
03 / Consequence surfaces
One decision grammar. Explicit adapters.
ScopeWrit does not pretend a Python package is a kernel sandbox. It enforces the paths actually integrated through its adapters and tells you exactly where the boundary ends.
implemented
Subprocess
Absolute executable paths, principal-signed exact argv digests, shell disabled, and adapter-bound execution tickets.implemented
MCP tools
Exact server and tool identity plus argument hashing before a tool callback receives authority.integration layer
SDK operations
Fixed-class adapters for browser, cloud, credential, package, and publication operations. The real SDK call must be routed through them.04 / Adversarial gate
Built to fail closed.
The test suite attacks the authorization boundary, not just the happy path. Every listed case is executable in the public repository.
Read the test casesManifest tamper and trust-root substitution
PASSNonce reuse across runs
PASSApproval and ticket replay
PASSCross-adapter ticket use
PASSDeny-versus-allow precedence
PASSConcurrent counter oversubscription
PASSLedger mutation and valid-prefix truncation
PASSPath escape, host-header, body-size, and URI ambiguity
PASS05 / Evidence boundary
What the alpha proves—and what it does not.
PROVEN IN THE PUBLISHED HARNESS
- Authorized loopback HTTP work survives the enforcement path.
- The identical synthetic outside request is denied before execution.
- Signed scope, counters, tickets, replay controls, and the audit chain survive adversarial regression tests.
- The wheel builds and installs cleanly across the supported Python line.
NOT CLAIMED
- Transparent host, kernel, VM, or container isolation.
- Safety for an agent that retains raw capabilities outside a guarded adapter.
- Multi-node high availability or KMS/HSM key custody.
- Independent penetration testing, certification, or a real-model behavioral study.
500 durable local iterations per path, including Ed25519, transactional state, ticket mint/consume, and fsynced recording. This is a machine-local receipt, not a capacity SLA.
Scope the run. Then let the agent work.
Agent autonomy needs mechanical authority.
ScopeWrit is an open public alpha for teams operating autonomous coding, security, cloud, and evaluation agents in high-consequence environments.