Deterministic, trace-first execution for language model workflows.
If it influenced an output, it must be written down.

RATATOSKR is a local-first execution enforcer for language-model workflows.

It does not train models.
It does not host models.
It does not retrieve knowledge to “improve answers.”
It does not attempt to be intelligent.

Instead, RATATOSKR refuses to execute unless all inputs, context, memory references, engine configuration, and outputs are made explicit, persisted, and inspectable.

If a model produces an output without explicitly citing the material it used, execution fails.

Language models are treated as replaceable engines.
Memory and context live outside the model.
Every run produces durable artifacts that can be replayed, audited, and explained.

RATATOSKR is infrastructure — not an assistant.

If something influenced an output, RATATOSKR requires it to be written down.
If something changed, RATATOSKR records it.
If it is not visible on disk, it did not happen.

There is no hidden state.
There is no implicit memory.
There is no silent retrieval.
There is no background execution.

RATATOSKR is RAG-adjacent by nature.

It may assemble external text (documents, excerpts, notes) to condition generation. However, its purpose is not to improve answer quality.

Its purpose is to enforce execution semantics.

Where most RAG pipelines treat provenance as optional, RATATOSKR makes provenance mandatory.

RATATOSKR does not ask:

“What information should we retrieve?”

It asks:

“What information actually influenced this result — and where is the proof?”

Each execution creates a trace directory containing:

• the frozen task specification (input.yaml)
• the fully assembled prompt (prompt.txt)
• resolved context artifacts (documents + extracted chunks)
• explicit memory references
• engine declaration
• raw model output
• a material usage ledger proving which inputs were actually used
• trace metadata and lifecycle state

These artifacts are append-only, durable, and replayable.

RATATOSKR is designed to work in concert, not competition, with:

• FUR — durable AI conversation memory
• Yggdrasil-CLI — project and codebase flattening
• downstream engines (local or remote) treated as pure executors

RATATOSKR defines execution truth.
Other systems may consume it.

• v0.3.0 — Citation-enforced execution
• Message-level material atoms
• Mandatory post-inference validation
• Material usage ledger (auditable grounding)
• Engine abstraction intentionally minimal

Future versions will extend capability without breaking the execution contract.

RATATOSKR intentionally does not:

• optimize for scale or throughput
• provide SaaS or hosted inference
• perform federated or decentralized compute
• hide complexity behind convenience
• infer intent or “help” beyond what is specified

Any feature that violates these constraints requires a new scope.

Licensed under the Apache License, Version 2.0.

RATATOSKR is the paper trail your language model cannot escape.