An architectural index.

What would it take to make every AI-touched code change reviewable before merge?

A gate that evaluates each AI-touched PR against the codebase's history, contracts, dependencies, tests, conventions, and security constraints can produce an inspectable chain of evidence before the change reaches main.

Self-deployable Next.js / TypeScript / Tailwind template. Reference implementation.

Public repository. Self-deployable. github.com/modarresi1913/Evoguard ↗

Scope narrowed from "AI code review platform" to "context-aware PR gate." Honest narrowing.

The original broader scope could not be supported by the construction. The narrowing was the failure — and the revision.

Scope narrowed. README rewritten to match what the construction actually does.

Prototype. Self-deployable. No production-readiness claim beyond what the repository demonstrates.

A benchmark of failure cases — AI-touched PRs the gate should have caught but missed.

What would an operating-system layer for biological neural networks look like — and what would have to be true for the question to be falsified?

Biological neural networks exhibit properties no current silicon substrate replicates. A shared OS layer would make systematic experimentation possible.

Versioned specification (v0.1, stable) + L0 reference emulator. CI green across Python 3.9–3.12. MIT-licensed.

Public specification, public emulator, public CI. No biological result is claimed. github.com/modarresi1913/OrganoidOS ↗

The broader claim — that biological substrates will prove practically useful — is not settled.

The construction cannot yet demonstrate any biological result. This is not hidden — it is the current state.

Specification versioned. Conformance levels defined. L0 is honest about what L0 means.

Specification v0.1 stable. L0 reference emulator. No biological experiment claimed.

v1.0 target: multi-vendor conformance — at least two independent L1 implementations.

Stateless AI treats two users with identical prompts identically, even when their interaction states differ. Can context estimation improve interaction quality — and where does it actively make things worse?

Estimating probabilistic interaction states from observable signals, then adapting the response policy, can improve interaction quality. Explicitly falsifiable.

Experimental architecture. Browser demo. v0.1.0. Apache 2.0. Every number labelled ESTIMATE, never DIAGNOSIS.

Live demo ↗ · Repository ↗ · Tests reported passing. Not a validated result.

The architecture must surface cases where its own hypothesis fails. The benchmark is designed to find them, not to confirm success.

Pending. The benchmark is designed to surface failure — when it does, the failure will be published as data, not hidden.

None yet. v0.1.0 is the first public release.

Experiment. Live demo. Falsifiable. Not validated.

Publish the failure cases the benchmark surfaces — separately from the cases where adaptation helped.

A single trade touches markets, maritime logistics, sanctions screening, contract documentation, letters of credit, and cargo insurance. Can a coordinated agent swarm preserve traceability across the full workflow?

Six specialized agents — Market Analyst, Logistics Coordinator, Documentation Expert, Compliance Officer, Finance Advisor, Operations Manager — collaborating across the trade lifecycle can preserve traceable, contestable decisions end-to-end.

Production-grade Next.js 16 / TypeScript / Prisma codebase. Six-agent architecture. MIT-licensed.

Public repository. Whether deployed in any specific environment is not claimed. github.com/modarresi1913/hanna-ai ↗

A multi-agent system cannot be safer than the models and data it is built on. Traceability is necessary but not sufficient.

Not yet encountered in deployment — because deployment is not claimed. The construction has not yet met real workflow resistance.

None yet. The architecture is as originally specified.

Prototype. Production-grade codebase. No deployment claimed.

A documented trade-decision replay — a single end-to-end trade reconstructed after the fact, with every agent decision traceable and contestable.

What must an architecture make explicit if it is to remain revisable by those it affects?

Platforms, protocols, institutions, and AI systems can be treated as things that construct, constrain, distribute, and transform practical possibility — and a six-dimension audit kernel can make that shaping inspectable.

Open-source philosophical program, v2. Possibility-graph model, six-dimension audit kernel, ONTacture method, falsification conditions. MIT-licensed.

github.com/modarresi1913/onturgy ↗ · Read the manuscript →

Four empirical passes. Each pass was a place where reality disagreed with an earlier version.

v1 failed in specific, documented ways. The failures are incorporated into v2 — not hidden.

v2 incorporates progressive revisions from four empirical passes. Incorporated resistances, not validations.

Revised. v2. Published specifically to be challenged.

External critique. v3 will depend on what the critique surfaces.

A clinical claim can be net-supported yet highly fragile — meaning it looks settled only because its weak points have not yet been tested. How do you make the cracks visible?

A QBAF-style bipolar argument graph applied to clinical evidence review can compute a Fragility Index (F) that quantifies how provisional each part of the reasoning actually is.

Zero-dependency, single-file research tool (~515 lines). MIT-licensed. v0.1.

The repository explicitly labels itself a research prototype; it is not a clinical decision tool. github.com/modarresi1913/kintsugi-clinical-fragility ↗

A fragility score is itself a claim, and itself fragile. The construction does not escape this — it surfaces it.

The fragility score cannot be more certain than the arguments it scores. This is a structural limit, not a bug.

None yet. v0.1 is the first public release.

Prototype. Research tool. Not a clinical decision tool.

A small set of clinical claims scored end-to-end, with the scoring itself opened to contestation.

Median-optimised models flatten exactly the part of the output distribution where bold creativity lives. Can a routing model be trained to go there on purpose?

A routing-style model in the spirit of OpenRouter, but aimed at bold, unconventional creativity rather than median-quality output, can surface what median-optimised systems flatten. Falsifiable.

In early construction. No public weights, demo, benchmark, or API.

None yet. This entry exists specifically to make the absence of evidence inspectable.

Not yet encountered — the construction has not reached the stage where resistance is possible.

Not yet possible. The hypothesis has not yet met a benchmark.

None yet. The hypothesis is published before the construction exists.

Hypothesis. In early construction. Not yet visible.

A reference implementation. Then a benchmark. Then — honestly — whatever the benchmark surfaces.