From 5773942c8ffe41352104d673ff0eef949df10c71 Mon Sep 17 00:00:00 2001 From: Georg Bauer Date: Sat, 3 Oct 2026 19:10:48 +0200 Subject: [PATCH] Name the parity oracle per model family Co-Authored-By: Claude Opus 5.5 --- AGENTS.md | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/AGENTS.md b/AGENTS.md index d32dc54..7caa4f7 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -3,8 +3,11 @@ - Prefer simple, idiomatic Rust; reuse existing code and dependencies before adding abstractions or crates. - Keep changes focused, handle errors explicitly, and add the smallest useful test for non-trivial behavior. - Preserve `rustfmt` output and keep Clippy warning-free. -- Treat DS4 as the behavioral oracle. Model execution, token processing, context accounting, and KV-cache behavior must remain identical to DS4. Differences are bugs unless they are unavoidable Rust/platform requirements and the user explicitly approves them before implementation; when uncertain, preserve DS4 behavior. -- Keep all application, model-loading, inference-graph, session, and orchestration code in Rust. Never copy, vendor, compile, or call DS4's C implementation; use it only as a behavioral and performance reference. Native code is limited to the Objective-C Metal bridge and `.metal` kernels required to execute GPU work. +- Each model family has one behavioral oracle. Model execution, token processing, sampling/RNG, context accounting, and KV-cache/session behavior must remain identical to it. Differences are bugs unless they are unavoidable Rust/platform requirements and the user explicitly approves them before implementation; when uncertain, preserve the oracle's behavior. + - DeepSeek and GLM models: DS4. + - Qwen 3.8 Flash Next: the pinned MTPLX revision. Follow its current execution route, not only its numerical results. +- Keep family-specific behavior in that family's code. Before changing a helper shared across families, trace every caller and keep the other families' oracle behavior exact, with regression controls. +- Keep all application, model-loading, inference-graph, session, and orchestration code in Rust. Never copy, vendor, compile, or call DS4's C implementation or MTPLX/MLX Python; use them only as behavioral and performance references. Native code is limited to the Objective-C Metal bridge and `.metal` kernels required to execute GPU work. - This is not a GitHub project. Use direct `git` commands for version control and the `tea` CLI for forge operations; do not use GitHub tools or workflows. - Issues are managed with the command "tea" run from the project directory.