An independent check that has to pass before an automated system files, denies, or pays.

Two rules. Every material fact the decision relies on has to trace to a source that actually contains it. Every legal number it uses has to be the one in force on the date the decision covers. If either fails, the action stops, and the stop says what is missing or what is present but out of date. The check follows the published rule, so anyone can recompute a stop without trusting me. The law stays the authority, not the checker.

I'm starting with public benefits, where the rules are written down and a maintained rules engine already exists.

Redtape is a published benchmark, on the Prime Intellect Environments Hub, that scores whether a model knows when a benefits question cannot be answered from the facts it was given. Answer keys come from PolicyEngine and are fixed when the tasks are generated, so no language model grades anything and every score can be reproduced from a cache. Its original headline result is retracted. Before re-running, I committed both possible corrections and the rule for choosing between them. The re-run, plus a disclosed follow-up analysis, showed the design could not test its own central claim, and the retraction sits at the top of the repository and the Hub listing. The newest model I tested applied next year's benefit amounts to this year's cases, which is the failure the check above is built to stop.

Three bug reports filed against benchmarks I did not build, each one a case where a correct answer was graded as wrong:

• prime-envs #825: the verifier passes test input through a single command-line argument, so on Linux any input over about 128 KB fails to start and a correct solution is recorded as failing.
• prime-envs #826: the verdict depends on how many cores the grading machine has and how busy it is, so the same solution can pass or fail from run to run.
• human-eval PR #23, comment: HumanEval problem 116's tests pass in negative numbers that its own specification excludes.

PolicyEngine #9374. I reported two places where the rules engine my benchmark depends on diverged from the law. Half of that report was my mistake: my probe had hardcoded one state. I retracted that half publicly on the issue. The other half stands, and the maintainers invited a fix, which I'm preparing.

Glassbox: Every project below independently reinvented the same four disciplines: an LLM proposes, deterministic code decides, a human confirms before anything consequential happens, and the system's own limitations get disclosed, not hidden. Glassbox is what's left after distilling that pattern out of 70+ separate builds: an importable Python library for all four pieces, a machine-checkable specification (glassbox-spec.json) defining exactly what "glassbox" means, and a conformance checker that statically analyzes a real codebase against that spec. It never imports or executes the code, it just proves the guarantees hold.

It's tested against itself honestly, not just demoed. Run for real against four unmodified sibling projects in this portfolio, it scores each 56 to 68 out of 100, never a rigged 100, because none of them were built against this spec. And one piece of the library (bounded repair/retry on failed structured extraction) is disclosed as genuinely new: a portfolio-wide search found no project that re-prompts Claude on a bad response. Every one just fails loud. That gap, and the real file/line citations for everything else, are documented in docs/REFERENCE_IMPLEMENTATIONS.md.

LogChain keeps the record of each check: what was decided, from which facts, under which version of the rules, in an append-only log where any later edit is detectable. It is a Rust tool with a second, independent verifier in Python, and the two are checked against each other on every test run. Writing its format specification exposed a root collision in my own Merkle tree (the same class as CVE-2012-2459) that both implementations had agreed on. I shipped the fix as format version 2, built on RFC 6962, as a clean break that voids every version 1 root.

The pattern above was distilled from these. More than eighty builds, mostly defense and national-security tools, built between June and August 2026. Several of the standalone web demos were later merged into Arbor, Analyst's Desk, and Cleared Facility Suite. Live demos run on a free tier and can take about a minute to wake up.

awesome-defense-tech: a curated list of every public API, open standard, and policy document used across this portfolio. 35+ free APIs (CISA KEV, USASpending.gov, Space-Track, USGS, EIA, WHO DON, congress.gov, CourtListener, SEC EDGAR, and more), 30+ frameworks (MITRE ATT&CK, DISARM, NIST AI RMF, DoD AI Ethical Principles, CNSA 2.0, NIST PQC, CMMC, Purdue Model, ICD 203), and 49 worked examples. If you're building in this space, start here.

Stack: Python · FastAPI · Claude (Anthropic) · Click · Rich · SQLAlchemy · pandas · numpy · NetworkX · ChromaDB · Z3 · Jinja2 · Render · TypeScript · React · Next.js · D3.js · Chart.js · Rust

Approach: Claude extracts and synthesizes; deterministic Python rules make every decision. No LLM randomness in compliance, scoring, or attribution paths.