docs: accuracy pass — de-stale x86, fix misleading claims, cut internal scratch - #54
Merged
Conversation
…al scratch Tier A (broken / misleading / stale, verified): - CONTRIBUTING, ARCHITECTURE: x86 AVX-512 backend has LANDED (was "policy skeleton / not supported / fails fast"); note it's opt-in via BACKENDS= - DEPLOY: mark the GHCR-container and SSH-deploy rows as sketches (no such workflow exists); EMBED_MODEL_URL now defaults to BitNet → geist-bitnet ships by default; name the geist-bitnet artifact - ROADMAP: prebuilt x86-64 binaries now ship (was "not yet"), reconcile w/ CHANGELOG - TERNARY_BITNET: status was "not yet measured / blocked on Pi / in progress" but the doc's own results show 17.4 t/s vs bitnet.cpp 8.2 — rewrite to "measured", cut the stale "what unblocks" list to the one open gap (2B-4T TQ2_0 model) - benchmark/BENCHMARK: fix docs/BENCHMARKING.md → BENCHMARKING.md (404) - TERNARY_BITNET: fix broken ../... link - agent.md: constrained decode ships (forced name + arg re-key); scope the open item to in-sampler grammar masking, not "generate-then-validate" - README: version 0.3.0 → 0.3.3 (badge/tag were already 0.3.3) Tier B (remove internal scratch from the public repo): - delete docs/LINUX_X86_PERF_PROFILE.md (599-line profiling diary with reverted experiments and a "not reproducible here" note) and docs/LINUX_X86_SPEC.md (pre-impl spec for the landed backend); rescue the final x86 numbers into a concise benchmark/BENCHMARK_X86.md and repoint headline_results.json - delete tests/migration-plan.md + its completed "Migration Status" section - remove .agent/AGENT.md; gitignore .agent/ and .agent_memory_unit/ Tier C (bloat / redundancy): - benchmark/README: surface the ternary (2x) and x86 results that the hub omitted; reconcile the stale "decode tie" with the spec-head 7.5 t/s - QUICKSTART: cut the single-file prose duplicated in README/DEPLOY, keep the geist_model_load_from_memory snippet + link to DEPLOY
- include/geist.h GEIST_VERSION_STRING 0.3.0 → 0.3.3 (matches tag/badge/README); version.c already derives from the macro, so `geist --version` now prints 0.3.3 - scripts/check-version.sh: assert include/geist.h == README badge/status/citation (source of truth is geist.h; a guard, not a generator) - ci.yml: run the check on every PR (new `version-consistency` job) - release.yml: new `version-guard` job — a `v*` tag must equal GEIST_VERSION_STRING and have a matching `## [x.y.z]` CHANGELOG section, else the release is blocked
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follow-up to the README overhaul: a correctness/conciseness sweep across the rest of the docs (three parallel review agents + manual verification).
🔴 A — Broken / misleading / stale (verified)
CONTRIBUTING.md("x86 not supported yet / fails fast") anddocs/ARCHITECTURE.md("policy skeleton, no compute path") now reflect the shipping AVX-512/VNNI backend (opt-in viaBACKENDS="cpu_x86 cpu_scalar").docs/DEPLOY.mddescribed infra that doesn't exist — GHCR container + SSH-deploy rows are now marked as sketches. Also:EMBED_MODEL_URLdefaults to BitNet (was framed opt-in), sogeist-bitnet-<platform>ships every release; artifact now named.ROADMAP.mdsaid prebuilt x86 binaries were "not yet" — they ship (reconciled with CHANGELOG).benchmark/TERNARY_BITNET.mdstatus header said "not yet measured / blocked on Pi / in progress" while the doc's own body reports 17.4 t/s vs bitnet.cpp 8.2. Rewrote to "measured", cut the stale unblock-list to the one real gap (2B-4T TQ2_0 model).docs/BENCHMARKING.md→BENCHMARKING.md(×2), and a../...link.docs/agent.md: constrained decode ships (forced name + arg re-key); scoped the open item to in-sampler grammar masking.README.md: version 0.3.0 → 0.3.3 (badge + git tag were already 0.3.3).🟠 B — Internal scratch removed from the public repo
docs/LINUX_X86_PERF_PROFILE.md(599-line profiling diary with reverted experiments and a "514 t/s… never reproducible here" note) anddocs/LINUX_X86_SPEC.md(pre-impl spec for the now-landed backend). Rescued the final x86 numbers into a concisebenchmark/BENCHMARK_X86.md; repointedheadline_results.json.tests/migration-plan.md+ its completed "Migration Status" section..agent/AGENT.md; gitignored.agent/and.agent_memory_unit/.🟡 C — Bloat / redundancy
benchmark/README.mdhub omitted the ternary (2×) and x86 wins entirely — now surfaced + linked. Reconciled the stale "decode tie" with the spec-head 7.5 t/s.docs/QUICKSTART.md: cut single-file prose duplicated in README/DEPLOY, kept the uniquegeist_model_load_from_memorysnippet + DEPLOY link.Net −1145 lines. All relative doc links verified to resolve; no references to the deleted files remain.
Deliberately deferred (need your call — not release-blocking)
geist.hGEIST_VERSION_STRINGis still"0.3.0"while the tag/badge/README say 0.3.3 — that's a code/release bump, not a doc edit.[Unreleased]was never cut for 0.3.2 / 0.3.3.geisten/geistlibslug (README release links, CHANGELOG, CITATION.cff) left as-is per your earlier "don't mass-normalize" call — works via redirect.