execution_guard
Fail-closed. External writes and spend default OFF.
GhostOS
GhostOS is the governed operating system behind Ghostlines.
It observes state, retains context, routes work, executes inside bounded authority, verifies outcomes, and escalates consequential decisions to a human. It refuses to self-execute strategic or capital decisions.
4
Public artifacts
Human-gated
Writes
Unattended
Operations with an audit trail
Executive overview
Not a chatbot. Not a conventional operating system.
Ghostlines builds governed AI operating systems capable of observing, researching, reasoning, coordinating, monitoring, and executing bounded real-world work across business functions.
GhostOS is the local-first operating layer behind that. It keeps persistent state, runs deterministic decision and underwriting engines, schedules workflows, integrates external systems, and coordinates bounded unattended AI workers. Consequential actions pass through a fail-closed three-state permission model — auto-allow, hard-deny, or human-escalate. It never self-executes a strategic or capital decision.
Architecture
The governed operating loop.
This is the platform ceiling. Every recommendation is tagged by evidence source. Nothing is blended. The last consequential node is a person.
Persist state / evidenceKeep the trail. Rank, including when the honest answer is REJECT.Ranked, Not Curated
Worked example — capital-allocation workflow (one GhostOS workflow, not the platform ceiling): Discovery → Underwriting → Investment Report → Human decision. Deterministic engines score and tag. Model-driven workers draft, research, and execute bounded tasks. External writes default OFF.
Governance
Fail-closed writes. Three states. No pending.
Three-state permissions
AUTO-ALLOW / HARD-DENY / HUMAN-ESCALATE. Pending is not allowed.
Escalation queue
Durable queue, not an interactive prompt that stalls overnight work.
Human final say
GhostOS never self-executes a strategic or capital decision. This is not a production deployment inside any employer’s systems.
Real operating capabilities
What GhostOS actually does today.
Implemented now. Planned architecture is labeled separately and is not sold as live.
Governed unattended operation
Overnight and scheduled missions run with an audit trail. Typed outcomes. The process exiting is not treated as the work being done.
Persistent state
Missions, leases, escalations, and evidence live in GhostOS-owned state — not in a vendor chat memory.
Deterministic engines
Underwriting and decision engines tag evidence by source. Ranked opportunity and capital-allocation reports, including REJECT. A person still makes the final call.
Scheduled workflows & integrations
Bounded workflows against real repositories and public sources (including SEC EDGAR ingest on the Company Intelligence artifact). External writes default OFF.
Landlock capability controls
Filesystem and network Landlock controls on bounded workers. Command-string permission matching is in production use. Complete interpreter and network sandboxing is not claimed.
What it refuses to automate
Strategic and capital decisions stay human. Independent review is exercised, not automatic standing production behavior. Complete interpreter and network sandboxing is not claimed.
CROSS-MODEL VERIFICATION
Trust No Model. Verify Everything.
GhostOS dynamically assigns local and cloud AI providers across builder, independent-review, and adversarial-review roles based on suitability, availability, independence, and cost — while deterministic validation, evidence binding, and Chairman authority remain outside the model layer.
FLAGSHIP CASE
No model gets the last word.
Claude built it. OpenAI Codex audited it — and found defects Claude had missed. Then Grok audited Codex, confirming some findings, rejecting or refining others, and uncovering additional defects neither model had surfaced.
That pattern became a governing principle inside GhostOS: the builder does not certify the builder, the reviewer is not presumed correct, and model confidence never outranks evidence.
The objective is not model consensus. It is verified mission state.
Public artifacts
Clone and check.
Same four public repositories as the homepage, ordered by current technical and evidentiary strength. Law Engine is not presented as the strongest artifact today, and is not current legal-service capability.
Enterprise Agent Lab
MCP server, deny-by-default writes, named human-approval override, append-only audit log, dual native Python and LangGraph implementations.
Reference implementation, not a production deployment at a named enterprise.
Company Intelligence Pipeline
SEC EDGAR ingest → normalize → validate → SQL analytics for five public companies: Apple, Microsoft, Alphabet, Amazon, NVIDIA. Last published GitHub snapshot: 2,826 facts / 135,371 values / 32 tests / 6 SQL analytics. Public, Apache-2.0.
Five public companies, not a market-data platform. Chart: public docs/revenue_trajectory.html only.
Law Engine
Provenance-tracked UCC learning and legal-reasoning architecture spanning Articles 1, 2, 3 and 9, with jurisdiction-specific source corpora. Public, Apache-2.0. A real, growing automated test suite. Precedent-conflict layer is a two-case prototype.
Educational / technical demonstration. Not legal advice and not current legal-service capability.
AI-Native Content Operating System
Public content-ops architecture. Systems thinking, not a measured client case. Example metrics in the repository are illustrative.
Architecture document. Metrics are illustrative, not Ghostlines client outcomes.
GhostOS engine tests number in the thousands at the controller/engines layer — order of magnitude, not a hero KPI.
Case studies
Public on GitHub.
Reliability case study
Typed outcomes, three-state permissions, durable escalation. Unattended operations are real and bounded.
Public on GitHub →Law Engine UX / task-first iteration
Task-first product iteration on the learning surface.
Public on GitHub →Law Engine product / architecture
Product and architecture decisions for a provenance-tracked legal-learning system.
Public on GitHub →Precedent conflict mapping
“Same law, different outcome.” Two judicial opinions, one flagship question. Two-case prototype.
Public on GitHub →Operating rules
Five constraints every decision answers to.
Business outcomes over engineering elegance
A passing test suite is not business progress on its own.
Evidence over model confidence
No estimate becomes fact because it was stated confidently.
Engineering tied to business metrics
Name the metric a subsystem is expected to move before building it.
Mandatory experiment kill conditions
Every project carries an explicit kill condition before capital or labor is committed.
Minimize required labor for durable value
Raw output is not the target. Systems should compound useful work while keeping a human on consequential command.
Related Ghostlines services
Qualification and coordination — not GhostOS-operated fulfillment.
Merchant processing
Ghostlines qualifies the opportunity, coordinates statement review, and connects qualified businesses with a specialist payment-processing partner.
Ghostlines is not a payment processor.
Merchant Services →Asset protection
Ghostlines provides white-glove intake and coordination. Legal analysis, entity formation, and trust implementation are handled by a qualified, specialist trustee law firm.
Asset Protection →Briefing
Bring one workflow. We will show the control path.
Same enterprise form as the company homepage.