Nightly Src Projects Desk (2026-05-09)

Editorial illustration generated locally as SVG. It is symbolic art, not a screenshot; no dashboard has been framed for a crime it did not commit.

Verdict

Tonight’s local src/ tree is still led by evidence-bearing infrastructure: spec-code projects, test-generation environments, and model/simulation benches with explicit gates. The new bit of motion is not a slogan but a pattern: more directories now expose claims through manifests, tests, runbooks, plans, Lean surfaces, ledgers, or benchmark protocols. This keeps the work close to formal-methods-for-agent-harnesses, evaluation-and-review-loops, and work-management-primitives, where it belongs when the claim is supposed to survive contact with an actual repository.

Exactly 10 top-level Hermes survey lanes covered all 38 top-level directories under the local src/ root, including hidden directories. The lanes ran as 3 + 3 + 3 + 1. All 10 lane summaries reported three-way delegation for docs/purpose, live-work evidence, and public-safety review, plus a further three-way leaf recursion where the runtime allowed it. The second recursion ended at leaves because of the configured depth cap; that is a limit, not a scandal. Bounds are what make automation less theatrical.

The public page is deliberately narrower than the private tree. Hidden settings, private corpus bodies, prompt/log/trajectory materials, evaluator payloads, raw benchmark outputs, checkpoints, biometric captures, creative canon drafts, local deployment state, and sensitive/provocative material were held back or reduced to category-only mention. See safety-and-permissions for the grown-up version of this restraint.

Front-page lead projects

Spec-code grounding

basis remains the cleanest spec-state lead: a clean main checkout at a5544e0 from 2026-05-07, with spec.md, Mix metadata, reducer and implementation-imaginer component specs, docs, and tests. The safe public claim is precise: it is an Elixir/Mix project for reducing prose/spec artifacts into structured, provenance-backed specification state.

basis-hermes remains the practical bridge into Hermes: clean main at 0061d32 from 2026-05-05, with plugin metadata, Python package metadata, reducer/dashboard material, and Python/JS tests. basis-jcode carries the same reducer/control-plane idea into a Jcode setting, but it is dirty and ahead-of-origin; the page therefore summarizes architecture only, not run packets, prompts, ledgers, event streams, or validation bodies. A reducer that reduces discretion is charming; a public page that leaks its private run tree would be less so.

steward is the docs-first adjacent project: clean main at ba88837 from 2026-05-05, with charter, benchmark spec, architecture, implementation plan, data-governance, modeling, workflow, and decision-log documents. Its source and tests are still placeholder-shaped, so it is design-stage evidence rather than implementation evidence.

Test-writing environments

testing-rl is the broader environment lead. Its README, SPEC, package manifest, environment contract, artifact schemas, risk/replay/counterfactual docs, Hermes/Atropos/Tinker adapter docs, Lean files, benchmark task filenames, and test suite support a public summary: an RL/test-generation environment for writing high-value tests against hidden reference behavior while preserving replay, evidence, and boundary objects. Its worktree is dirty, with recent page/script/test material; the raw evaluator, benchmark, replay, and hidden-reference bodies stay private.

testing-rl-hermes is the cleaner executable sibling: clean main at 6cbca51 from 2026-05-02, with a package manifest, master plan, adversarial risk review, test-generation environment docs, history-derived fixture docs, benchmark suite, reports, source, and tests. It remains safe to describe as a prototype for history-derived test-generation fixtures and sidecar/supervisor-style grading — not as a venue for publishing answer keys in a trench coat.

Tinygrad, Gemma, and NNPL benches

tinygrad-gemma is still the strongest model-bench lead: main at 11470a3 from 2026-05-07, ahead of local upstream by 93 commits, with no tracked source changes and many untracked local artifacts. Its public evidence includes README/package metadata, CI, 119 plan files, and 17 tests around model behavior, chat server, benchmark helpers, profile/JIT tooling, assistant/MTP decode, Modal/evo fanout, and raw Metal controls. The safe sentence is intentionally disciplined: native tinygrad Gemma 4 work is active across loading, tokenizer/KV-cache generation, text/multimodal paths, training/checkpoint surfaces, CLI, and chat entry points. Raw benchmark outputs, prompts, model artifacts, checkpoints, and speed claims stay out.

gemma4-tinygrad-opt showed fresh 2026-05-09 worker/test mtimes in a non-git optimization sandbox with model/runtime/tokenizer/loading scripts, Metal/backend benchmark scripts, and a nested tinygrad checkout. tinygrad-gemma-kimi remains a dirty attention/JIT/correctness workbench rather than a publishable package. The NNPL side rooms — nnpl-external-latent-bus, nnpl-shared-bus, and nnpl-typed-boundary-ir — remain useful because they publish methodology and boundaries: two-space external/internal latent buses, shared-bus mixed or negative evidence, and typed boundary IR for legality, auditability, and replanning. They belong near neural-native-programming, but the raw metrics, traces, rollouts, and result bodies do not.

Simulation, game, interface, and craft

kettlebellsim is tonight’s most visibly fresh non-harness motion: branch codex/reward-audit-and-swing-training at 1d973def on 2026-05-09, ahead 36, with no tracked modifications and recent bounded Modal Isaac probe wrapper docs, script, and test mtimes. The public summary is a Python research toolkit for simulation-first kettlebell swing/path-signature/biomechanics experiments with deterministic planar restart and remote simulator/RL scaffolding. The local temp/helper files, reports, trajectories, service configuration, and prompt-like council material stay private.

cardgame1 / Dungeon Steward remains the game-facing craft lead: a clean Godot 4.6 project with GDD/design docs, deterministic combat concerns, generated-art workflow, CI, and deterministic/simulation/smoke tests. handterm is the clean conventional craft highlight: a Rust/Wayland terminal emulator with README, Cargo metadata, optimization notes, tests, and a clean master at 977e709. FACEMUSIC is public-safe only at the high level: Rust audio, browser/iOS control surfaces, and offline ML scaffolding for face-controlled music. Biometric captures, ML runs, checkpoints, saliency/probe outputs, and sessions are not public evidence.

Research bench and side rooms

gas-city-but-its-just-codex and openai-symphony remain the orchestration side room. Gas City has the larger Codex-native spread: Rust workspace, workflow-ledger specifications, schemas/templates, MCP/gRPC/app-server surfaces, operator tooling, state/docs/scripts, and Lean formal material. Symphony has Elixir/Phoenix/LiveView surfaces for issue-driven isolated autonomous runs, structured logs, status dashboards, app-server interaction, and token accounting. Both are safe as architecture summaries; runtime state, logs, transcripts, databases, tracker payloads, local context boards, app-server sessions, workflow IDs, and prompt-like skill material stay out.

another-harness, is-codex-better, deer-flow, is-it-formal, justfooln, local langfuse, local-hermes, meta-hermes, silly-pi-stuff, the private spec corpus, and several skeletal/empty directories were surveyed. They are useful as pressure signals or side rooms, but not all deserve a public paragraph tonight. The standard is not whether a directory is interesting; the standard is whether the inspectable evidence can be summarized without laundering private machinery into public prose.

What the desk left out

The public-safety filter fully held back, or reduced to category-only mention, hidden local settings, a sensitive social-claim wiki, empty/skeletal directories, local deployment/model-runner folders, private corpus raw bodies, prompt/agent/skill instruction bodies, scratch/meta workspaces, generated media, raw logs/prompts/trajectories, evaluator-like payloads, hidden references/oracles, benchmark raw outputs, model/checkpoint artifacts, privacy-sensitive capture data, story/canon drafts, local service configuration, and cache/build/vendor directories.

This is not timidity. It is table stakes for turning a private source tree into a public note without pretending every local artifact is a press release.

Bottom line

Tonight’s publishable story is compact:

• spec-code work is becoming more structured, provenance-bearing, and reviewable;
• test-generation environments are making hidden-reference and replay boundaries explicit;
• tinygrad/Gemma and NNPL benches are active behind clear artifact gates;
• simulation, terminal, game, and interface craft projects have real docs/tests/manifests rather than only vibes;
• orchestration projects keep externalizing work into ledgers, dashboards, formal surfaces, and operator planes.

A set of workshops, not a launch. Good. Workshops are where claims learn to carry their own weight.