TPT Rosetta

Status: experimental research project. Rosetta has no released API, no stable schema, and no measured results yet. Nothing in this repository should be read as a claim that it saves tokens, cost, or time.

TPT Rosetta investigates whether AI task requests can be represented more compactly and precisely than ordinary natural-language prompts without reducing task success, correctness, or important nuance.

The project will provide:

• a canonical, model-independent semantic representation of task intent, constraints, permissions, context references, and acceptance criteria;
• encoders that render that representation as ordinary English, concise English, structured text, and experimental symbolic notation;
• a benchmark harness that scores task outcomes, constraint violations, and cost per successful task rather than token length alone.

The governing documents are:

• spec.txt — the initial design specification (v0.1.0 proposal).
• todo.md — the phased roadmap and current task list.

Principles

• Measure, do not assume. Shorter strings, invented symbols, and fewer apparent words are not evidence of fewer model tokens or better results.
• Negative and inconclusive findings are reported, not hidden.
• Requests are representations, not grants of authority. Encoding a request must never broaden its permissions, and references to tools or resources are identifiers, not authorisation.
• Treat prompts, retrieved documents, repository content, and model output as untrusted data.
• Rosetta is usable as a standalone library and CLI. Integrations with other TPT repositories are optional and must preserve a safe fallback to the original prompt.

Scope boundaries

Rosetta owns the semantic task schema, notation experiments, encoders, decoders, validators, and representation-quality benchmarks.

Rosetta does not own:

• TPT Weave — repository indexing, context retrieval, and context selection.
• TPT Code Command Center — request interception, provider and model routing, credentials, and workflow orchestration.
• TPT-UIR — the typed computational intermediate representation. Rosetta is not a generic computation language.
• TPT Script — general-purpose scripting.
• TPT Telos — formal verification. Rosetta may reference it for properties with a precise mapping and supported proof semantics. An ordinary LLM judgment is never labelled a formal proof.

The integration boundaries above are proposals. Check the implementation of any adjacent repository before relying on a feature from it.

Non-goals

• Replacing human languages in general communication.
• Requiring a custom tokenizer, vocabulary, or trained model.
• Guaranteeing fewer tokens for every model or task.
• Executing instructions or modifying repositories on its own.

Claims

No universal token-saving or efficiency claim is made. Any efficiency claim will be limited to the models, tasks, settings, and evaluation conditions that were actually tested, and will report regressions.

Status of this repository

The Rust workspace builds and its offline tests pass. It has four crates:

• crates/rosetta-core: canonical task model, validation, canonicalisation, diff.
• crates/rosetta-encode: DSL parser and renderer, four encodings, fallback.
• crates/rosetta-bench: corpus loader, run planning, token estimates, metrics.
• crates/rosetta-cli: the rosetta command.

Try it with the example task:

cargo run -q -p rosetta-cli -- validate examples/split-parser.rosetta
cargo run -q -p rosetta-cli -- encode examples/split-parser.rosetta --to concise
cargo run -q -p rosetta-cli -- corpus check fixtures/corpus/v1

Documentation lives in docs/: notation, architecture, benchmark methodology, security and integrations.

No model has been run against the benchmark corpus, so the repository has no measured results. See todo.md for what remains.

Licence

Licensed under either of the MIT Licence (LICENSE-MIT) or the Apache Licence, Version 2.0 (LICENSE-APACHE), at your option.