PRISM v2.1.1 released #11
Ronkupper
announced in
Announcements
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
PRISM v2.1.1 is out — a release-hygiene PATCH over v2.1.0. Surfaces sourced from the post-v2.0 shakedown and a Cowork irregular-run diagnostic on 2026-05-01.
Highlights
New: SP-15 — Triangulation integrity. Codifies the falsifier-grade triangulation premise as a Standing Principle, with two corollaries the framework had been carrying implicitly: single-vendor multi-agent fan-out is parallel execution, not falsifier-grade triangulation (equivalence dispatch requires distinct vendors); and self-triangulation — when the orchestration vendor is also one of the triangulated execution vendors — carries no asymmetric weighting in convergence.
§1.1 framing-line tightening.
(orchestration on Claude; execution on any vendor)→(orchestration on Claude; execution on selected vendor per Vendor Selection). Compressed-attention readers had been latching onto "any vendor" as a category statement and missing the §4.5 default-stance correction 900 lines downstream.§3.2.3 / §3.2.1 Prompt-digest semantics rewrite. Field-semantics bullet now leads with purpose ("detects wrong-prompt / wrong-attachment delivery at dispatch boundaries") before mechanism, and adds the explicit anti-pattern callout: Generating the digest at return time provides zero integrity check — there is nothing to compare against. The Envelope-template token is inverted to
[orchestration-generated at dispatch; copy verbatim; never recomputed]so the counter-signal sits inside the template itself.Calibration items worth reporting back
If you run a multi-vendor PRISM audit where Claude orchestrates and also participates as one of the triangulated vendors, SP-15's self-triangulation corollary is the load-bearing claim. Convergence should feel mechanical; if it doesn't, that's worth a Discussions note.
The §1.1 framing change should be invisible to operators who already understand Vendor Selection. The new wording is the load-bearing surface for compressed-attention reads and for vendors who are loading the framework with limited context budget.
Files & provenance
PRISM.mdPRISM_v2_1_1.mdBacklog and what's coming next: see
PRISM_backlog.mdonmain.Beta Was this translation helpful? Give feedback.
All reactions