Deterministic floor, model ceiling
An agentic OS for
LLM-centered development.
Zunstack runs a fleet of projects — research, games, fiction, benchmarks — under one harness that gives every AI session memory, gates, and bookkeeping it can't skip.
the fleet — live topology · click a node
Doctrine
Deterministic floor, model ceiling.
The model decides; deterministic tooling enforces and records. Nothing that records or projects state ever calls a model — records are pure functions of the op log, so every projection reproduces.
Zunstack / engine
An agentic OS: harness crafting, memory, gates, bookkeeping.
| module | provenance | cadence | carried by |
|---|---|---|---|
| structures-index | R — regenerated | immediate | all (universal map) |
| harness-structure | P — authored | batch | all |
| oplog / op-reports | hybrid P+R | immediate | all |
| behavior · convergence | P — model-synthesized | batch | wiki-type |
| standing · docs-index | P — authored | batch | wiki-type |
| code_module | R — code-cycle | batch | code repos |
| trendline | R — op-series metrics | batch | planned |
| contract | P — tracked contract | own cycle | all |
Sessions forget. The fleet doesn't.
Claude-control
The Zunstack app — 99% of user interaction lives here.
the plane — real capture
One surface for every session — desktop, phone, village.
LLM-research
A database that compounds — not a pile of search results.
| layer | contents | state |
|---|---|---|
| raw | immutable captured sources | built |
| source pages + derivatives | distilled summaries · claim-level provenance (31 files) | built |
| concepts / tools | ~76 interlinked user-facing pages | built · growing |
| synthesis + reports | user-led insight · daily/weekly movement reports | next |
Knowledge that compounds while you sleep.
Self-writing LLM
Event sourcing applied to fiction.
A story world you can revert like code.
ARPG sandbox
A game development suite with a playable co-op ARPG inside.
live from the workshop — real capture
in-game — real capture
[ dungeon forge — scroll to load ]
the real generator + renderer, running live
Build system, identity system, and endgame share one geometry.
Bench / ML
A personal measurement instrument for model quality.
3,111
rows scored
2,882
loop points
405
failures caught
$40.11
total cost
| suite | signal | scoring |
|---|---|---|
| IFEval + IFBench | instruction adherence | deterministic checks |
| BBH (hard subset) | reasoning under thinking | exact match |
| GPQA-Diamond + HLE | raw capability | exact match |
| ARC-AGI-2 | abstract reasoning | grid equality |
| Aider-Polyglot | agentic coding | test suite pass |
| comms_recall (custom) | buried-detail survival | recall scoring — opus 0.85 · haiku 0.53 |
Vibes are not data. This is data.
Endgame
The projects aren't a portfolio. They're organs of one machine.
End-state Zunstack is an operating system for intent: you say what you want, and the fleet does the rest. Sessions spawn under harnesses built for their work. Choices get made from taste profiles — your decision-making, as data — instead of waiting on you. Quality is measured, never vibed. Prompts are recursively refined against bench scores. Knowledge compounds from the web while you sleep. And every layer feeds the next: the research informs the refinement, the refinement sharpens the sessions, the sessions build the tools that tighten the loop.
engine
rigor for semantic work
Code has compilers and tests; prompts, contracts, and instructions have nothing. The engine gives non-mathematical work the same discipline — gates, records, validation loops.
bench
testing harnesses, prompts, variance
You can't improve what you can't measure. Bench turns harness changes, prompt edits, and model variance into scored, confound-stamped data.
ML
recursive improvement
Bench deltas become training signal: prompts recursively refined against real scores instead of intuition. The optimization loop over the whole stack.
LLM-research
fresh knowledge, compounding
Models go stale and web searches don't accumulate. A vetted, revertible database that scans in loops and compounds — feeding the improvement cycle fresh insight 24/7.
claude-control
the surface + automated choices
A human approving every step is the bottleneck. Taste profiles — decision data derived from engineers — let runs make your choices without you in the room.
worldbuilding vault
storybuilding as a system
AI fiction fails on continuity — docs full of info don't stop contradictions. Event-sourced, delta-gated, reversible world state makes long-form story a solved data problem.
ARPG sandbox
the code-native proving ground
The other pole of the spectrum: where compiler + tests + golden corpus already make drift loud, the harness sheds machinery — proving the module system adapts to its feedback signal.