From 053f9a153e46e7f37e1a7b43e871a20bc0d910a1 Mon Sep 17 00:00:00 2001 From: Cody Kickertz Date: Wed, 16 Sep 2026 10:20:24 -0500 Subject: [PATCH] docs(repo): state host role instead of machine identity This repository is published; the machine it happens to build on is not part of its public contract. Three CLAUDE.md sites named a private host-operations repo and its GPU by model number, and the CI pipeline header justified its concurrency pin with a specific core count and co-resident workload. Each site now states the role that is actually true of the software: machine topology and GPU mode selection are owned by the host-operations environment, and research mode is mutually exclusive with inference mode on a single-GPU host. The concurrency pin keeps its reason, minus the core count: cargo's default parallelism is the host's total core count, so an unpinned run sizes itself to the whole machine rather than to its share of it, and contends with whatever else is running there. A dedicated machine does not retire the pin, because concurrent runs and co-resident services still share those cores. No generated block was touched; the kanon:auto markers in CLAUDE.md and README.md are untouched. --- .kanon-ci.toml | 10 +++++++--- CLAUDE.md | 6 +++--- 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/.kanon-ci.toml b/.kanon-ci.toml index 5e3d154..6ab81f7 100644 --- a/.kanon-ci.toml +++ b/.kanon-ci.toml @@ -5,9 +5,13 @@ # Forge CI pipeline for forkwright/zetesis. # # Mirrors the hardcoded default Rust gate (fmt → check → clippy → nextest -# → lint) but pins per-stage build + test concurrency to 8. Without this -# cap, cargo defaults to num_cpus (64 on this Threadripper) and parallel -# rustc + nextest processes can contend with other fleet work. +# → lint) but pins per-stage build + test concurrency to 8. The pin is a +# co-residency cap: cargo's default parallelism is the host's total core +# count, so an unpinned run sizes itself to the entire machine rather than +# to its share of it and contends with whatever else is running there. +# A dedicated machine does not retire the pin, because concurrent runs and +# co-resident services still share those cores; only an executor that +# confines each run to its own CPU budget does. # # Keep in sync with `crates/archeion/src/ci_config.rs::default_rust_gate` # when the upstream default changes — only the `--jobs` / `--test-threads` diff --git a/CLAUDE.md b/CLAUDE.md index bd2dc95..d7f5510 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -1,6 +1,6 @@ @@ -39,7 +39,7 @@ Zetesis has three planned fleet consumers (aletheia, dioptron, akroasis) plus li - Free-tier APIs have aggressive rate limits; `sylloge` tracks free-tier quotas separately from paid spend. - Deep research can blow $10+ in token costs per query if orchestrated against Anthropic/OpenAI. Default backend is local logismos. -- `menos gpu research` mode (Phase 6, in coordination with menos-ops) is exclusive with `menos gpu inference` on the W7900; operator picks. +- GPU research mode (Phase 6) is mutually exclusive with GPU inference mode on a single-GPU host; the operator picks which one holds the device. - Downstream use must comply with the scoped license map above; commercial client-contract work must not depend on zetesis. ## Related @@ -52,7 +52,7 @@ Zetesis has three planned fleet consumers (aletheia, dioptron, akroasis) plus li | logismos | Self-hosted orchestration backend | | koina + fjall | Cache + budget ledger persistence | | heurēma | Future semantic rerank of Tier 0 results | -| menos-ops | `menos gpu research` mode owner | +| host-operations environment | Owns GPU research/inference mode selection | | hermeneus (inside aletheia) | Sibling primitive: hermeneus unifies LLM providers, zetesis unifies research providers |