# Experimental release boundaries

[Hermes PR #116246](https://github.com/NousResearch/hermes-agent/pull/116246) motivates this project: ranking alone cannot replace text condensation. This page records limits or review evidence for the scoring, retention, storage, and compaction fixes.

This is a release candidate, not a production-certified replacement.

Tests use synthetic text and scores. HTTP tests use loopback servers. They do not measure provider quality, real billing, or comparative performance against private transcripts.

Cloudflare Clef is implemented and tested but has never been called live. Every Cloudflare credential available to this project returned HTTP 401 from Workers AI, and no account id is configured, so the Clef wire contract comes from Cloudflare's published documentation rather than an observed response. Treat the first real Clef call as unverified. Its unit tests use a synthetic transport plus a socket-capture check on the request shape, which proves what is sent and how a response is read, not that Cloudflare accepts it.

No local decision model other than the Laya default has been called live from this repository. No local engine server runs here, so the `local_model` slot selecting Laya or other pre-deterministic routing models is a contract claim from those projects' published documentation rather than a measurement, and no Cloudflare credential available to this project is authorized for Workers AI.

Selected conversation text is sent to the configured provider. This is not an automatic secret-redaction system. The local raw archive is not encrypted by this package. Use a private data directory.

Hermes uses bundled upstream LCM. DSH uses its native BasicCompactionEngine with a separate SQLite archive and summary links, not a full TypeScript port of Hermes LCM.

The byte-based request cap is conservative, not an exact provider tokenizer. Oversized candidates remain unscored. Provider failure leaves host condensation available without fabricated scores.

Further qualification is required for live providers, sustained DSH sessions, cancellation, restart restoration of calibration, independently measured recall, and randomized cross-language parity.

DSH summary nodes currently record generated summaries, including attempts the host may reject. They do not prove a surface replacement committed. Their predecessor chain is not a complete source DAG. Raw evidence is stored separately.

Publication does not activate this engine in a running agent. Package registry publication is separate from GitHub publication.