01 · Measure
Record original and compiled bytes, model-neutral estimated tokens, reduction percentage, source coverage, and duplicates removed.
Context CompilerDeterministic · source-linked · model-neutral
Give agents less context. Keep the evidence path.
Turn bounded raw documents into a task-specific Context Pack with deduplicated passages, source references, stable hashes, and transparent before-and-after measurements.
The compiler is deterministic middleware, not a model wrapper. It reduces what reaches the model while preserving an inspectable relationship to the supplied sources.
02 · Trace
Every included passage points to a source ID, passage position, and content hash. The original document is never stored.
03 · Constrain
Set a context budget and explicit task. Passages that do not fit are excluded visibly rather than silently overflowing the prompt.
Declared boundary
What this release does not claim
It does not promise a fixed percentage reduction, exact provider token counts, complete source coverage, factual verification, or hallucination prevention. Those are workload-specific outcomes to measure. Use MPS when consequential claims need a separate provenance review.
Machine integration
A bounded contract for human and autonomous buyers.
API-key clients or autonomous x402 buyers call the Context Pack API with a task, a token budget, and 1–8 textual sources. The response is returned directly; the service retains only a privacy-safe outcome ledger, not source text or the compiled pack.
Implementation guides
Reproduce the method before trusting the result.
System relationship
The portable evidence object between Maha's controls.
MPS can review consequential claims in a Context Pack; the Enterprise MCP Gateway can restrict which tools receive it; and the commercial ledger can measure where compact, source-linked context is actually used.