Status date: 2026-08-23
The repository implements the locked, local, simulation-backed protected-action vertical slice. This is an engineering completion boundary, not a claim of production network protection or public-launch readiness.
| Capability | Current state | Evidence boundary |
|---|---|---|
| Identity and enrollment | Durable deployment/operator identity, operator-scoped one-time enrollment, node lifecycle, scoped credentials, and an agent-side retry-safe key/proof bootstrap that retrieves an approved matching certificate | Operator approval remains intentional; platform-keystore enrollment, service packaging, and status visibility remain external work |
| Event and projection spine | Immutable idempotent events, SQLite projections, replay cursors, retention, signed hash-chained audit | Local relational store; no hosted replication claim |
| Decision plane | Deterministic posture, findings, policy compilation, signed expiring plans, secure-action gate | Decisions use available evidence and remain fail-closed on missing or stale inputs |
| Agent observation | Enrolled mTLS agent, signed durable queue, continuous Linux network/route/DNS collection, opt-in socket metadata, Tailscale read-only CLI probe, and Headscale BYOK read | Host observations are DETECTED; no packet/application payload, provider mutation, or process attribution is claimed |
| Enforcement transaction | Validate, snapshot, apply, verify, commit or roll back; durable local transaction state; non-mutating antiflock-agent plan verify inspection of signed plan identity, target, expiry, layout, parameters, and declared capabilities |
The executable verifier always reports executable: false; mandatory mesh, route, DNS, durable nftables ownership, and independently verified recovery drivers are absent, so production host mutation remains disabled |
| Secure Action SDK | Request binding, callback isolation, hold/block/allow, durable single-use grant consumption, mandatory execution-start audit, evidence-provenance enforcement | Reference TypeScript SDK and Aether demonstration only; simulation callbacks require an explicit per-call test opt-in |
| Android Guard | Pure Kotlin/JVM fail-closed state machine, platform ports, recording adapters, deterministic reference app | No APK, VpnService, real packet transport, or real-device validation |
| Third-Eye dashboard | Authenticated same-origin Core proxy, live projections/stream, locked views, and explicit enrolled-agent/flow/mesh/Nano setup cards | Private/local operator surface; Core credentials never enter browser JavaScript |
| Scrambler | Deterministic bounded simulation and explanation | Execution is disabled and no host/provider state is changed |
| Nano watchdog | Pinned Nano v0.1 compiler/runtime, audited immutable program admission, atomically advanced durable per-program SQLite cursor, and deterministic typed-finding proposal API | Proposal-only; Nano has no I/O or mutation capability and antiflock-nano actions still require consent under protected posture |
| Public surface | Verified-owned-asset provider contract plus deterministic Shodan-style, broker-registry, and paste-reference fixtures | Offline digest-only REPORTED/SIMULATION evidence; no live third-party scraping or raw public data retrieval |
| Vehicle appearance | Android coarse-feature, session-local HMAC correlation and aggregate-only output | In-memory reference only; no frames, faces, plates, VIN/OCR, embeddings, make/model, exact location, or cross-session tracking |
| Coffee-shop flow | Live Core + durable SQLite + simulator executes HOLD through verified recovery to ALLOW with five audit events |
All network and recovery verification is explicitly labeled simulation |
The sole release command is:
npm run verifyIt must pass from a clean checkout with locked dependencies and no undocumented
host tools. The final verified commit and evidence report are recorded in the
LOOP-18 run artifacts under _loopstate/LOOP-18/.
The strict acceptance harness has ten gates and reports
reference_vertical_slice_gates_passed. A release candidate requires 10 / 10;
the presence of a file alone is not sufficient for runnable Core, web, SDK, and
coffee-shop gates. The SDK gate includes a live Core/SQLite restart test that
proves hold-to-allow re-evaluation, exactly-once callback execution, durable
lifecycle audit, idempotent replay, and changed-content conflict rejection.
The following are intentionally outside this reference-release boundary. They are OPEN, and must not be inferred from a green verification run:
- production VPN or packet transport;
- a validated Android always-on/lockdown kill switch;
- unattended nftables or other privileged host mutation;
- real Tailscale, Headscale, WireGuard, DNS, captive-portal, or roaming tests;
- proof of interception, tracking, or active surveillance;
- imported or redistributed field-intelligence datasets;
- live Shodan, broker-registry, or paste-site retrieval;
- synthetic-profile publishing, location/device spoofing, or exposed honeypots;
- production Scrambler activation;
- public use of the
AntiFlockworking name; - external penetration testing, privacy review, legal review, or release signing.
These are separate OPEN gates because falsely claiming them would weaken the project's safety and evidence contracts. The integration-level acceptance criteria are in open decisions.