English · 한국어 · 日本語 · 简体中文 · 繁體中文 · Español · Português (BR) · Deutsch · Français · Русский · Italiano · Bahasa Indonesia · Türkçe

Code-based automation and agent-orchestrated systems: an agent platform with a fail-closed permission layer, local retrieval pipelines with evaluation gates, and deterministic simulations that replay bit-for-bit.

Every number on this page is a measurement, not an estimate. The command that produced it lives in the repo it describes, and the repo re-checks it before publishing.

The interesting part is not "built with AI". It is what sits between the model and the repository.

• Two coding agents, one deterministic referee. agent-relay is an external driver that runs two agent CLIs against each other until a gate passes. The gate decides "done" — never an agent's opinion of its own work. 12 unit tests, passing with PATH reduced to the system directories, so the driver's refusals are provable without either agent installed.
• Deterministic gates run before the model is ever called. Purity checkers, not vibes: fatal-funnel's sim gate bans 11 non-deterministic APIs; reasonforge's read-only test enforces 6 forbidden patterns. Retrieval quality is a build gate too — a golden set plus thresholds that fail the run when they regress (rag-eval-demo).
• UI claims come from probes and screenshots, not from memory. hollowmere ships shot, relic-qa and mobile-qa harnesses; fatal-funnel has 35 headless harnesses among the 40 scripts in tools/ and 4 fixed screenshot cameras. This lesson was learned the expensive way: a tooltip that "works" can be dead behind one input-filter flag; a synthetic input probe catches that, reading the code does not.
• Assets are licence-gated before they are generated. assetforge refuses to generate from any model absent from its registry. Downstream, all 27 published snapshots carry a CREDITS.md with a 4-tier redistribution table — so anyone forking knows exactly what they must remove.
• Honest labels, enforced by a schema. In euan-portfolio every published metric is typed kind: z.enum(['measured', 'estimated']). A number cannot be published without declaring which one it is. That rule is why this page has no round marketing figures on it.
• Publication is a pipeline, not a copy-paste. Each public repo here is produced by a repeatable sanitizing publish whose claims file re-executes every quantity in the README and refuses to create a commit if one drifts. Across the 27 snapshots that is 629 test/spec files.

Not all 27 repos — the ones that show a distinct capability.

Browser games (open the link and play)

• fatal-funnel — deterministic tactical sim, 40 weapons, 12 hand-authored missions · live
• hollowmere — compact action RPG: parry, relics, two zones. Vertical slice at v0.5.1 · live
• neko-shift — small complete idle game about staffing ratios, 18 test files · live
• ashglass-reliquary — isometric ARPG slice. Shares its deterministic simulation core with hollowmere; combat rules, encounters and content are its own. · live
• moonshard-warden — three-wave isometric arena: read the telegraph, chain the mooncut, dash the committed strike. 43 test files and the strictest asset ledger of the set — 100 per-file SHA-256 rows, all CC0. · live

Engine games (Godot / Unity — vertical slices and prototypes, none commercially released)

• voidclad — capital-ship artillery in space: shell travel time, inertial lead, armour-angle penetration, on a deterministic 30 Hz sim (27 test files)
• cairnfall — solo boss-raid ARPG, ten three-phase bosses; the "the raid never carries you" claim is a measured post-fight stat, not an enforced invariant — if an ally lands the killing blow, the run report says so (33 test files)
• todak — desktop pixel pets that live along the bottom of your screen (36 test files)
• hordecaller (30) · ragtail (29) · driftfolk + its Unity port — the port's point is numeric parity with the original · gatewarden · emberline
• tidewrack — directed-autonomy battle royale, Unity. Listed separately because this snapshot is a verification skeleton, not a playable slice: the deterministic sim and its golden-file harness are the shipped part, and 11 of 15 numbered soak assertions are still pending.

AI and agent tooling

• baton — the agent platform: 63-tool MCP surface, approval broker, 16 orchestrator packages, 90 Swift sources
• reasonforge — read-only reasoning harness, 8 MCP tools, 0 dependencies
• agent-relay — gate-terminated loop between two coding agents
• assetforge — local, zero-cost asset generation behind a fail-closed licence gate
• content-qa-pipeline — generate → judge → compose → publish, where "not sure" parks the run instead of shipping it

Data, backend and domain tooling

• officegrid — multi-tenant asset administration where the tenant boundary lives in PostgreSQL: 82 row-level-security policies across 11 migration files, with 6 pgTAP test files aimed at them — not application-layer filtering
• marketing-knowledge-rag · rag-eval-demo — retrieval with the evaluation gate as the point
• crm-mcp-agent — agent + governance kernel over an all-synthetic CDP
• ad-compliance-linter — first-pass linter for Korean financial/insurance ad copy. All 5 shipped rulesets are marked "verified": false in the data itself and the report cannot print without its disclaimer: it flags copy for human review, it does not certify compliance
• loopsmith — play the instruments or write the patterns; the same song, edited losslessly from both sides (27 test files)
• euan-portfolio — bilingual static site where every number declares measured vs estimated · live

All 27 repos are sanitized snapshots: code under Apache-2.0, assets governed per-repo by a 4-tier CREDITS.md. Games are personal builds and vertical slices — playable, gated, and honestly labelled, not commercial releases.

Contact: open an issue or discussion on any repo here.