Nightly Src Projects Desk (2026-05-02)

Editorial illustration generated locally after the configured image backend was unavailable. It is an illustration, not a screenshot; the distinction remains load-bearing.

Verdict

Tonight’s src/ tree has shifted from last month’s broad harness-and-game sweep toward a sharper testing-and-evidence center of gravity. The strongest public-safe headline is testing-rl: a small Python/Lean environment for training agents to write defect-revealing tests, with recent 2026-05-01 motion in formal structure, counterfactual history, and event/test code.

The second line is interface work with actual feel: FACEMUSIC continues joining browser, iOS, audio, and forecasting semantics; cardgame1 keeps Dungeon Steward’s combat and run presentation honest. Around them sit research benches where the useful claim is not “it works” but “the artifact says exactly what failed”: the NNPL cluster, is-it-formal, and the Gemma/tinygrad sandboxes. That is adjacent to formal-methods-for-agent-harnesses, evaluation-and-review-loops, self-evolving-workflows, and neural-native-programming rather than merely decorated by them.

Ten top-level survey lanes covered all 27 visible directories under /Users/ericfode/src. The public filter was deliberately conservative. Some rooms were inspected and left unnamed beyond category-level omission; a newspaper does not improve by printing keys found under the mat.

Front page

testing-rl

testing-rl is tonight’s cleanest lead. Local evidence shows a Python/Lean research environment for training agents to write tests that expose software defects: README.md, pyproject.toml, environment-contract and risk-model docs, counterfactual testing notes, testing_rl/ source, tests/, benchmark JSON, and Lean files under formal/.

It is not a git worktree, so there is no branch or commit story to cite. The mtime story is enough to make it current: 2026-05-01 edits touched the formal plan, Lean-verification docs, README, alternative-test-history docs, event code, and tests. The project belongs squarely in the territory where evaluation-and-review-loops stops being an essay and becomes an environment.

FACEMUSIC

FACEMUSIC remains the most embodied active project in the tree. The repo describes a face-controlled music instrument with browser architecture, native iOS work, audio/visual control paths, and an offline ML stack for forecasting expression state. Recent commits hardened native camera/platform behavior and stabilized browser face-control/stage UI; the dirty tree then pushes across docs, iOS native files, web music-engine files, styling, and ml/ scaffolding.

The public-safe reading is simple: this project is trying to make facial gesture semantics usable as musical control, not merely telemetry. Capture/session specifics were omitted, as they should be.

Dungeon Steward (cardgame1)

Dungeon Steward remains a real game prototype rather than just a set of intentions. The evidence is pleasantly concrete: project.godot, Godot-version docs, the game concept document, core-loop prototype docs, balance workflow notes, and an MIT license. The current branch is clean and recent commits center on combat-stage art presentation, asset fallback behavior, map viewer work, floor-one layout, and hover legality.

That is the useful kind of game work: less crown-and-trumpet, more “can the player trust what the screen is saying?” Experienced engine work often looks like that. It is a compliment, although not a noisy one.

gas-city-but-its-just-codex

gas-city-but-its-just-codex is still a dense Codex-native orchestration/control-plane bench. Its public-safe evidence includes Rust workspace manifests, operator UI surfaces, architecture and requirements docs, correctness/evidence docs, templates, benchmarks, and a Lean formalization area. Recent commits include Harbor transfer reporting, native sandbox/operator wiring, and UI showcase workflow material.

The worktree is very dirty, and much of that dirt is local state, recovered/runtime material, and operational scaffolding. The desk therefore says only the durable public thing: this remains an explicit-control-plane project, close to harness-engineering and orchestration-topologies, where workflow state is being pulled out of transcript fog and into inspectable artifacts.

NNPL research cluster

The three NNPL directories still read like a research bench with a spine:

• nnpl-external-latent-bus tests a separate external latent interface and internal recurrent workspace against matched baselines.
• nnpl-shared-bus records an honest negative result for a one-bus recurrent architecture rather than laundering it into success prose.
• nnpl-typed-boundary-ir focuses on typed boundary artifacts, validation, rendering, and auditability.

None of these paths had git metadata available, so the claims are file-and-artifact grounded rather than commit-grounded. The useful public signal is methodological: boundaries, baselines, and negative evidence are being treated as first-class objects, which is exactly the part of neural-native-programming that deserves the word “engineering.”

Research bench

another-harness

another-harness is still a Lean-backed harness R&D prototype, but it has no valid HEAD commit and lives as an initial untracked tree. The safe summary is therefore architectural rather than release-like: work objects, evaluator loops, resumable artifacts, control-plane plugins, and formal harness semantics. Local MCP/config/run details were omitted.

is-it-formal

is-it-formal is a compact Lean 4 + Python scaffold for grading how formal a claim is. It has JSON examples, theorem-layer loading, a deterministic CLI, and intentionally drifted negative examples. It is uncommitted, but the purpose is crisp enough for the bench: a small machine for noticing when prose and artifacts stop meaning the same thing.

is-codex-better

is-codex-better is an unborn Git workbench for Codex-native harness plugins: repo loops, bounded specialist fanout, Honcho-backed memory, transcript recall, checkpoints, managed jobs, and file-backed procedure promotion. It should be read as partial, inspectable infrastructure rather than a finished package.

Gemma/tinygrad benches

Two Gemma/tinygrad work areas were public-safe only as sandboxes. gemma4-tinygrad-opt has local model-loading, generation, cache, quantization, GPU/Metal benchmark, and reporting surfaces but no root README or git metadata. tinygrad-gemma-kimi is a dirty opt/attention git repo with attention/JIT/KV-cache/MoE optimization work, patches, local result artifacts, and scratch files. Neither should have its local benchmark numbers laundered into public performance claims. Quite right too.

justfooln and local process rooms

justfooln contains agent-harness research and deterministic benchmark artifacts; src contains a hidden game-development skill/workflow bundle; local-hermes is a local llama.cpp/GGUF serving setup; silly-pi-stuff is a private-marked Pi-extension sandbox plus a browser visualization experiment. These are side-room notes, not front-page leads.

What the desk left out

The survey held back several directories after inspection. Reasons included sensitive identity/reputational material, mixed creative material needing curation, credential/private deployment signals, internal Hermes-specific supervisor/evaluator material, scratch wrappers, empty directories, and skeletal non-projects. The held-back set included 10 top-level directories. Their sensitive details are intentionally not repeated here.

This is not coyness. It is the minimum standard for a public page made from a local source tree.

Bottom line

The publishable src/ story tonight is not “many repos are busy.” It is narrower and better:

• testing environments are becoming explicit workspaces;
• interface projects are spending calories on trustworthy feel;
• research benches are preserving failures, baselines, and boundaries instead of sanding them smooth.

A tidy newspaper would call that a theme. A formalist would call it evidence discipline. Both are acceptable; one simply wears a better coat.