Independent research and development
Jonathan Simone
I build software with AI coding agents, and I show how it was checked.
Simone Systems Research LLC is my independent research and development company in Merced, California (not affiliated with the SIMONE Research Group in Prague or university laboratories of a similar name). I set the direction, choose the design, decide what to keep, and check the result. AI coding agents do the implementation, testing and drafting.
Technical roles, focused projects and press: jon@simoneresearch.com / Contact page
Selected work / Public artifacts
What I've built
- Baldies remake. A 1995 game rebuilt without its source code. 334 tests pass; all 100 levels load in 4 views with 0 errors on the Windows desktop build. Verification record and case study.
- BigBoss. An approval gate for coding agents. Each decision is tied to one hashed action. Release v0.1.0 passed 399 author-run CI tests on the local single-reviewer MVP. Case study and test record.
- The Council. Answers, critique and claim checks across model providers. A recorded run shows a claim check catching an error the critique missed. Recorded run
- Pokémon TCG. A matchup counter tested against seven opponent implementations in 8,400 game records. It gained +20.42 and +18.75 percentage points against two Alakazam implementations (600 games per arm), while averaging -0.47 [-2.64, +1.71] across the other five tested opponents. Case study and 8,400-game CSV.
- The Argon One-Shot. A 3D city simulator in Godot 4.7 whose code is private, with a simplified browser sketch. Case study
- Voynich manuscript structural study. An out-of-sample structural investigation and external-protocol reconciliation on Yale Beinecke MS 408 across three transliterations and 15 held-out quire folds. The manuscript is not deciphered; every tested cipher and generator family breaks on held-out quires. Case study and findings index.
- Consumption Engine public fixture. A correction, a stale-data refusal and an archive replay you can run with plain Python. Run it
- Dog API lesson. A live request to a public API, with a teacher guide, worksheet and answer key. Lesson and teacher guide.
| Research Area | Public Artifact | How to Check | Current Scope Boundary |
|---|---|---|---|
| Baldies remake | Desktop build c37f1b5 / 01415d3; browser source archive | 334 desktop tests, 100 levels in 4 views, 9,000 random inputs; live browser demo | Rules are the remake's own tuning; desktop campaign requires an owned copy; browser edition uses generated levels. |
| Consumption Engine | Public fixture ce-stranger-replay-v1; 37-task comparison; Sept 30 procedure study | Run the Python fixture or inspect the 34-decision ledger | Private runtime; manual reuse of two retained ordinary-source procedures passed 6/6 on two CS336 policies (0 dispatch calls; prepared generation used +89% tokens), while the 37-task real-repository comparison showed no partial-check score gain (48.8% and 49.0% vs 53.8% ordinary workflow). |
| BigBoss | Release v0.1.0 and inspected source 5615b72 | Run the local quickstart or 399-test pytest suite | Local single-reviewer MVP; SQLite history is not a tamper-evident ledger and does not provide host isolation. |
| The Council | Public repository at af8643c; recorded trajectory | Inspect the 1.24 keV vs 1.5 eV claim-check trajectory | Demonstrates a specific claim check catching a peer-critique miss; no general quality advantage is established from a single run. |
| Pokémon TCG | 8,400-game panel CSV and strategy writeup | Inspect the 7-opponent x 600-game/arm evaluation records | Counter gains +20.42 and +18.75 percentage points against two Alakazam implementations; other five opponents average -0.47 [-2.64, +1.71]. |
| Voynich manuscript (MS 408) | Illustrated structural study, 47-finding & protocol matrix, and research note | Inspect the 15-quire LOQO battery, 166-to-41 held-out slot ablation, and 4-step R&T (2026) protocol reproduction | Not deciphered; operational slot ablation is descriptive, not proof of linguistic morphemes; closest synthetic producer (s93, rejected as a unified pass under D-064) overproduces short-range core recurrence by ~2.6x; Daybreak db-s96/s112 to s121 runs are labeled as user-supplied reported evidence. |
Why this matters
The open question
A finding becomes a summary, then an instruction for an agent. When the source changes, later work needs a way back. I want people and agents to be able to ask of any statement: where did it come from, what supports it, what contradicts it, and what changed.
Research collaboration
Open questions for collaborators
Each project below has an unresolved technical question stated in its public record:
- Consumption Engine: task-level benefit and autonomous procedure selection. Manual reuse of two retained ordinary-source procedures passed
6/6across six new-input batches with0autonomousrun_skillcalls, and on 37 real-repository tasks (222 builds) a compiled guide (48.8%) and library tools (49.0%) did not beat the ordinary workflow (53.8%). The open question is when a concise briefing of current facts, decisions and next steps outperforms ordinary source reading once preparation and review effort are counted. See the Consumption Engine research brief and September 30 procedure study. - Pokémon TCG: generalizing beyond a targeted matchup counter. Across 8,400 game records (seven opponent implementations, 600 games per arm per opponent), the separate Lucario counter gained
+20.42and+18.75percentage points against two Alakazam implementations while averaging-0.47([-2.64, +1.71]) across the other five opponents, and the experimental learner remains on a separate branch. The open question is how to combine learner updates with targeted counters for repeated retained improvement across a broader opponent panel. See the Pokémon TCG case study. - SEED protocol: testing whether the measurement protocol improves outcomes. SEED is a public measurement protocol and instrument whose core outcome claim (
C-004) remains an unverified hypothesis; an author-run commit census does not establish that following the protocol improves software outcomes. The open question is how to structure a controlled evaluation that measures defect and rework rates with and without the protocol. See the evidence boundary and SEED repository. - Voynich manuscript: out-of-sample external semantic anchor. Internal statistical triangulation localizes a six-stage line-and-quire writing hierarchy on Beinecke MS 408 while rejecting tested prose, gibberish, self-citation, grille, and simple/sticky/period cipher families on 15-quire leave-one-quire-out transfer. The open question is what out-of-sample external prediction (an unseen object class, astronomical/calendar parameter, or historical codebook constraint that fails on shuffled controls) can test meaning without per-page retuning. See the Voynich case study and findings index.
If you are working on a related evaluation or want to compare notes on one of these questions, get in touch on the contact page.
How the work is evaluated
How I work
- I publish what was tested, next to the result.
- I keep built, measured and planned apart.
- Corrections stay in the record, with the date.