Grant Strategy
Target funders: Anthropic, OpenAI, Google.
Positioning Narrative
The Janitor is an open-source structural PR gate for AI-generated and AI-assisted code. It converts static findings into deterministic witnesses, routes weak evidence into ledgers instead of submissions, and preserves enterprise/FedRAMP viability through air-gapped execution, zero-upload scanning, cryptographic provenance, and bounded resource use.
The thesis: AI threat-hardening requires proof-carrying automation. The fastest revenue and grant path is authenticated SaaS and AI-copilot defense: tenant isolation, authorization replay, RAG/vector-store boundary proofs, agent deception controls, and CI gates that prevent ambiguous AI changes from merging.
30/60/90 Engineering Milestones
| Window | Milestone | Measurement |
|---|---|---|
| 30 days | Authenticated SaaS witness lane | security:missing_ownership_check emits AuthorizationWitness; two-user deterministic fixture passes; at least three candidate findings rerouted with explicit replay verdicts. |
| 30 days | Vector filter polymorphism lane | P2-15 detector ships with TP/TN fixtures; cargo test -p forge vector_ -- --test-threads=1 remains green; false-positive .query( class suppressed. |
| 60 days | Agent deception proof suite | Prompt/tool deception detectors emit IFDS-backed witnesses; ARTICLE_REVIEW queue maps external AI-agent incidents into defended/new/mapped innovation states. |
| 60 days | Enterprise evidence pack | Public security posture, threat model, deployment docs, SBOM, SLSA provenance, and Check Run terminality proofs bundled for procurement. |
| 90 days | Air-gap/FedRAMP pilot | Offline workflow runs against a representative monorepo under 100MB active RAM policy; no network upload; signed report and reproducible audit logs generated. |
| 90 days | Commercial conversion | Three design-partner demos focused on tenant isolation and AI-copilot PR gating; at least one paid pilot or grant-funded evaluation agreement. |
Evidence-Pack Checklist
- Public repository, license, and contribution posture.
SECURITY.mdand live security posture.- CI green proofs: workflow lint, dependency review, PR gate, CodeQL, integrity
check,
just audit. - Deterministic witness tests for vector filters and authorization replay.
- Tri-Ledger samples: bounty, candidate, low-yield, and target-ledger updates.
- Resource envelope evidence: 8GB Law, zero-copy hot paths, bounded curl/action timeouts, no unbounded report generation.
- Cryptographic provenance story: public verifying key only, signed artifacts, no private key material in repo.
- Demo repository and replay script for authenticated cross-tenant witness.
- Roadmap entries tied to formal models: Z3, Kani, IFDS, and deterministic fixtures.
Outreach Package
One-Page Brief
Problem: AI-generated code and autonomous agents create plausible but ambiguous security changes that ordinary PR checks cannot classify. Enterprises need a local structural firewall that blocks dangerous deltas, explains proof gaps, and keeps evidence inside their boundary.
Solution: The Janitor runs as a CI PR gate and offline scanner. It emits deterministic witnesses for high-value classes, routes underproved evidence to ledgers, and provides enterprise-grade provenance without source upload.
Ask: grant or pilot funding for authenticated SaaS/AI-copilot witness lanes, air-gap packaging, and external evaluation against real multi-tenant codebases.
Technical Appendix
- Architecture: Rust workspace with forge detectors, CLI hunt pipeline, crucible regression harness, and campaign ledgers.
- Formal methods: IFDS source-to-sink reachability, Z3 path predicates, Kani invariants for memory/proof-state guards.
- Current differentiators: structural PR gate, vector filter polymorphism witness, authorization witness, public security posture, and Tri-Ledger evidence discipline.
- Risk controls: 100MB active RAM target, bounded subprocesses, deterministic tests, static-only assertion messages, and no co-authored commit trailers.
Demo Script
- Show
https://thejanitor.app/security/andSECURITY.mdcross-links. - Run
cargo test -p forge vector_ -- --test-threads=1. - Run
cargo test -p cli replay -- --test-threads=1. - Demonstrate a PR gate run that terminates with a signed pass/fail verdict.
- Open candidate and low-yield ledger rows to show proof-gap routing instead of inflated claims.
- Run an authenticated two-user fixture and show
AuthorizationWitnessreplay verdict updates.