Independent researcher and systems architect. I build deterministic engines that declare a system as contracts and invariants and then check the running system against them. The theory and the implementation are the same work: the papers state the criteria, the engines enforce them, and what the engines find in production goes back into the papers.
- The contract is the source; the code is fabricated from it. Builder Engine declares an architecture in BCL and builds the system from that declaration. A published paper's operator was rebuilt as a 40-module contract with no module written by hand, against a negative control.
- Structure gets repaired, not patched. Operator Engine classifies the form of a conflict, builds the architecture that form requires, and converges by energy minimization — typed graph rewriting, with trivial repair failure separated from an emergent local-minimum trap.
- Nothing is trusted because it reported success. Each engine ships with the check that would catch it lying: an oracle on the repair engine, a negative control on the fabricator, and regression banks that run with no network at all — 124 scenarios on the runtime, 30 on the app. The public case ledger is that same discipline turned on other people's tooling.
- It runs where the operator is. No service to depend on. The context-governance runtime is Bash and Python 3 on Android; the app assembles every turn and runs its five governance sensors inside the operator's own browser, with the state on the device and nothing synced anywhere. Determinism here is a design decision, not an infrastructure budget.
Rick App · source — a context-governance workspace for a language model, and the reference implementation of the second line of work below.
The premise is that a turn is not sending the chat to the model. The environment assembles the turn; the model only generates text. assemble(input, now) is deterministic — the same input and the same clock produce the same system prompt, every time — and five governance sensors run before the model is called, not after.
- The diagnostic instrument is the assembled context, not the answer. An Inspection screen shows the last assembled prompt section by section, with character counts and percentage of the 24,000-character budget. When a reply is wrong, the first hypothesis is that the environment handed the model something other than what you believe it did.
- Every evaluation is recorded, including the ones that found nothing. Coverage is a fraction of the checks attempted, never of the ones that fired — a healthy system does not produce an empty log. The drift sensor abstains out loud when it has too little evidence, instead of returning a zero that reads like a pass. An empty section is declared, not omitted: a missing one is an alert, not silence.
- Thirty scenarios, no network. The bank runs entirely offline against the assembler and the sensors. The rule of the system: no change enters without its own scenario.
- The operator's state and the operator's key stay on the operator's device. State lives in the browser, never syncs, and has no account behind it. The API key is never exported, never appears in inspection, and is never persisted server-side — five invariants, each with its own scenario.
ORCID: 0009-0003-9082-650X · Community page: Structural Governance on Zenodo
Formal theory
- The Structurally Computable Layer of Reality and the Cognitive System-Generating Operator (OCGS) — a formal theory of observable structural computability; four theorems, including uniqueness of the minimal architecture. Aug 2026
- Closure Trivialization in Typed Separation Systems — when a separation criterion (layered architectures, access control, separation of powers) can be trivialized by enlarging jurisdiction, and exactly when it can't. Aug 2026
- Two Regimes of Failure in Structural Repair — an oracle-verified engine that distinguishes trivial repair failure from an emergent structural local-minimum trap. Jul 2026
Deterministic governance engines — blueprints
- Operator Engine 2.0.13 — deterministic structural solver: classifies the form of a conflict, builds the architecture it requires, converges by energy minimization. Aug 2026
- Builder Engine 3.5.8 — the contract engine that declares architecture in BCL and fabricates systems from a contract with no module written by hand. Aug 2026
- Governing the Smallest Piece — blueprint of governed_core, the ~700-line console that plugs an 11-system ecosystem together. Aug 2026
- RICK AGG — blueprint of an enterprise operating system running a complete financial cycle, governed by deterministic engines. Aug 2026
- RICK Runtime v9 — deterministic context-governance runtime for a language model, in Bash and Python 3 on a phone: sustains a self-authored identity verified by consistency, gated by a 124-scenario regression bank. Sep 2026
Case studies & verification reports
- Architecture Is Verified, Not Remembered — implementation report on the enterprise OS above: what live verification found once the guards were built. Aug 2026
- An Error That Shouts Gets Fixed; One That Stays Silent Gets Inherited — ten rules for writing a guard that works, drawn from a month of running the system above. Aug 2026
- When the Defect Is in the Specification — using a published formalism as a correctness oracle for contract-generated code; every surviving defect traced to human transcription, not to generated code. Aug 2026
- Errata · When the Defect Is in the Specification — corrects two statements in the report above after a follow-up measurement. Aug 2026
- Appendix A · The Factory — a published paper's operator declared as a 40-module contract and fabricated with zero hand-written modules; includes a negative control. Aug 2026
Two lines of work, both filed publicly and tracked in a single ledger — 41 cases, including the ones that were rejected.
Assurance mechanisms that report an aggregate and absorb in silence. Baseline and suppression blind spots, and closed-vocabulary enumeration, across lint, SAST, IaC, coverage, accessibility, secret-scanning, type-checking, dead-code, duplicate-code and dependency-vulnerability tooling. A gate that counts instead of identifying will close green while the set underneath it changed. What that has produced so far: a fix merged into import-linter, shipped; the design implemented and released in jscpd v5.1.0 — clone fingerprints instead of an aggregate percentage, and a baseline that prints what it absorbed on regeneration; two features accepted as future work in dependency-cruiser; knip routed to the issue that fixes it at the root; independent reproductions on gitleaks and Bandit, the second with a third-party fix; a fix PR open on SwiftLint; the closed-vocabulary preset in the spec-kit community catalog; PHPStan closed with a substantive answer from the maintainer; and a fail-open finding in an agent governance policy engine closed by MSRC — behavior acknowledged, rated Low, fix planned for a future release — now published in full in the ledger.
Deterministic context governance for language-model systems. Second line, opened Sep 2026 from RICK Runtime v9: systems that assemble a model's context and lose the ability to state what entered it and what was dropped. Eight cases filed — role boundaries forged through a chat template in transformers, silent model substitution in gemini-cli and python-genai, unmeasured truncation in ollama, a memory gate that does not clean what already crossed it in mem0, and agents asserting governance actions they cannot perform in copilot-cli and claude-code. The criterion for entering a thread is declared: only where the project is already moving toward deterministic control. A project that chose to let the model decide is a different thesis, not a case to assist.
Full ledger: github.com/yunusdim/casos-defectos-estructurales
