docs(pr-x12): audit findings for hallucinations in PR-X12 docs#200
Merged
Conversation
Adds .claude/PR-X12-docs-audit.md with full audit table (26 rows), verified-real anchors, architectural truths, doc-by-doc concentration ranking, three-tier cleanup recommendations, and scope boundaries. Recommendations only — no edits to the audited PR-X12 docs themselves.
Contributor
📝 WalkthroughWalkthroughThis PR introduces a new audit document that systematically evaluates the PR-X12 docs cluster against source artifacts, identifies hallucinations and misattributions, aggregates findings by severity and document, and provides tiered remediation recommendations. ChangesPR-X12 Docs Audit
Estimated code review effort🎯 2 (Simple) | ⏱️ ~12 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b7eb589be3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| | `src/simd_amx.rs` | ndarray | not whole-read in this pass (per task: "verified real, doc comments inside are sound") | Real AMX inline-asm + Linux prctl(ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA). | | ||
| | `.cargo/config.toml` | ndarray | 27 lines | Pins `target-cpu=x86-64-v3` (AVX2 baseline). AVX-512 (v4) is opt-in via separate `config-avx512.toml`. Comment lines 21-24 explicitly name "Runtime LazyLock dispatch" as a fifth supported mode. | | ||
| | `CLAUDE.md` "Hard Rules" + W1a consumer contract at `.claude/knowledge/vertical-simd-consumer-contract.md` | ndarray | per task statement: "real" | Authoritative. | | ||
| | `crates/lance-graph/src/graph/blasgraph/` | lance-graph | mod.rs (104) + semiring.rs (492) whole-read; 18 other files not whole-read | GraphBLAS-style sparse matrix algebra over 16384-bit binary `BitVec` carriers. Exports `Semiring`, `HdrSemiring` (7 variants: XorBundle, BindFirst, HammingMin, SimilarityMax, Resonance, Boolean, XorField), `GrBMatrix`, `GrBVector`, sparse-storage types. Bit-exact binary-Hamming substrate. Matches user's architectural truth #3. | |
There was a problem hiding this comment.
This audit marks crates/lance-graph/... files as "verified-real" and "whole-read," but this repository does not contain a crates/lance-graph tree (repo-wide file search returns no such paths, and git cat-file cannot resolve the cited files at this commit). Because several downstream conclusions depend on these anchors, presenting them as verified in this repo makes the audit non-reproducible and can misdirect follow-up edits; these entries should be re-scoped to verifiable in-repo sources or explicitly linked to an external checked-out repository/commit.
Useful? React with 👍 / 👎.
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In @.claude/PR-X12-docs-audit.md:
- Line 57: Update the BLAS-L2 method count to match the listed methods: change
the reported "8 BLAS-L2 methods" to "9 BLAS-L2 methods" in the audit text
referencing blas_level2.rs (ensure both occurrences around the current Line 57
and Line 101 are updated) so the count aligns with the methods listed
(gemv/ger/symv/trmv/trsv/syr/syr2/gbmv/sbmv).
- Line 63: Replace the instance of the word "followup" in the line containing
"Treat these as load-bearing canon for this audit and any followup:" with the
hyphenated form "follow-up" to match the document's preferred style; search for
the exact phrase "Treat these as load-bearing canon for this audit and any
followup:" and update it to "Treat these as load-bearing canon for this audit
and any follow-up:".
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: c335f1cc-8047-4b25-b0b9-cbd9f5fb7768
📒 Files selected for processing (1)
.claude/PR-X12-docs-audit.md
| | `crates/bgz-tensor/src/lib.rs` | lance-graph | 122 lines | Real exports include `Codebook4096`/`CodebookIndex` (line 108), `AttentionSemiring`/`AttentionTable`/`ComposeTable` (105), `HhtlDEntry`/`HhtlDMeta`/`HhtlDTensor` (112), `Base17`/`Base17Fz` (115). | | ||
| | `src/hpc/cam_pq.rs` | ndarray | 847 lines | `CamCodebook` struct (line 67) — 6-subspace × 256-centroid CAM-PQ scheme; `encode`/`decode`/`precompute_distances`/`distance_batch` methods; `train_geometric` (373), `train_semantic` (408), `train_hybrid` (454); exposed `kmeans` (541) + `squared_l2` (473) for downstream consumers. PackedDatabase stroke cascade (HEEL/HEEL+BRANCH/full) for 99% rejection before full ADC. | | ||
| | `src/hpc/codec/` | ndarray | listed only | A1/A2/A3-intra codec foundation: `ctu.rs`, `mode.rs`, `predict.rs`, `mod.rs`. No `ndarray-codec` crate yet. | | ||
| | `src/hpc/{dn_tree.rs, merkle_tree.rs, blas_level2.rs, bf16_tile_gemm.rs}` | ndarray | blas_level2.rs whole-read (487 lines) | blas_level2 surfaces 8 BLAS-L2 methods (gemv/ger/symv/trmv/trsv/syr/syr2/gbmv/sbmv). **No `batched_ssd_search` symbol.** Other files exist; not whole-read. | |
Contributor
There was a problem hiding this comment.
Fix BLAS-L2 method count mismatch.
On Line 57 and Line 101, the text says “8 BLAS-L2 methods,” but the listed methods are 9 (gemv/ger/symv/trmv/trsv/syr/syr2/gbmv/sbmv). Please align the count with the list to avoid weakening audit credibility.
Also applies to: 101-101
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In @.claude/PR-X12-docs-audit.md at line 57, Update the BLAS-L2 method count to
match the listed methods: change the reported "8 BLAS-L2 methods" to "9 BLAS-L2
methods" in the audit text referencing blas_level2.rs (ensure both occurrences
around the current Line 57 and Line 101 are updated) so the count aligns with
the methods listed (gemv/ger/symv/trmv/trsv/syr/syr2/gbmv/sbmv).
|
|
||
| ## 2. Architectural truths (user-established ground) | ||
|
|
||
| Treat these as load-bearing canon for this audit and any followup: |
Contributor
There was a problem hiding this comment.
Use “follow-up” for consistency.
Line 63 uses “followup”; prefer “follow-up” in this doc’s style.
🧰 Tools
🪛 LanguageTool
[grammar] ~63-~63: Ensure spelling is correct
Context: ...ad-bearing canon for this audit and any followup: 1. The PR-X12 "R-*" series is **future-cond...
(QB_NEW_EN_ORTHOGRAPHY_ERROR_IDS_1)
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In @.claude/PR-X12-docs-audit.md at line 63, Replace the instance of the word
"followup" in the line containing "Treat these as load-bearing canon for this
audit and any followup:" with the hyphenated form "follow-up" to match the
document's preferred style; search for the exact phrase "Treat these as
load-bearing canon for this audit and any followup:" and update it to "Treat
these as load-bearing canon for this audit and any follow-up:".
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 25, 2026
…idge.rs
The atom16 / precision-ladder / two-axes docs annotated BGZ17 as
"offset ≈17..27, stride {2,4}" — figures from a chat description that do
NOT appear in the source. First-party read of src/hpc/bgz17_bridge.rs:
BASE_DIM=17, GOLDEN_STEP=11 (coprime → table (i*11) mod 17 visits all 17
positions, Quintenzirkel-like phase), Base17=[i16;17], built by lossy
golden-step octave averaging of a 16384-bit plane.
The atom's two 2-bit fields are now labelled plan-inspired placeholders,
not grounded BGZ17 parameters. Removes a fabricated constant from a pushed
branch per the PR #200 audit discipline (source files are evidence;
chat/plan descriptions are not).
Note: this module remains an unmerged sketch — flat-disk manifold (the
ASG/KHR correction makes the directional lane spherical) and built from
the LeafPlanetarium inspiration doc. Superseded by the forthcoming
ASG-leaf canon entry; not for merge as-is.
https://claude.ai/code/session_01HbqooFZHAjaUtFEzhA1R2u
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 25, 2026
… truth PR #1 of the 7-PR cross-session program. Maps the ~62 ChatGPT-seeded .claude/plans/ docs (committed 2026-05-24/25) against evidence: real source files, ratified standards (KHR_gaussian_splatting / OGC 3D Tiles 2.0 / Cesium HLOD / ArcGIS ESRI_crs), and the merged #200 audit (PR-X12-docs-audit.md). NOT AUTHORITATIVE by construction — it ranks inspiration against ground truth and may not be cited as evidence (PR #200 Tier-4 model). Includes: - evidence model (evidence vs inspiration) per #200 Tier-4 - correction list (blasgraph canonical / bgz17 lossy sibling; ScalarCsr::spmv_min_plus is the real symbol; no batched_ssd_search; sigker PDE not buggy; per-arch crossovers uncalibrated; R-1..R-15 forward-conditional; directional lane is spherical not Poincaré) - supersedes-note recording all three sessions' in-thread errors (incl. this session's R-7/sigker/SSD assertions) with their fixes - 5-bucket inventory (USE-NOW / EPIPHANY / GROUND-FIRST / ARCHIVE / OUT-OF-SCOPE) with hop + cross-verified/single-session + grounds-onto - PR-X12-tensor-capstone downgraded to idea-USE-NOW / doc-RE-GROUND (carries #200 #1-#16 debt) - C/HLOD headline concept (Cesium HLOD shipping + certified overlay) - the forward 7-PR program + parking lot - honest scope boundary for the 2 plans not whole-read this pass No source changes. Pure .claude/knowledge addition.
5 tasks
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 25, 2026
…; cite #200 §1 Addresses Codex review on PR #201: src/hpc/pillar/{ewa_sandwich_2d, ewa_sandwich_3d,splat_invariants}.rs were absent from this branch's tree because the branch was cut from a pre-pillar/ master point. The files are real on master and independently confirmed by the merged #200 audit §1. Merged origin/master in (non-destructive) so every cited anchor now resolves on the branch, and added a note pointing to #200 §1 as the branch-independent witness.
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 25, 2026
…ot-checked) Answers 'what about hallucinated reviews in #200?' #200 is a Claude-authored doc, so it doesn't get terminal authority just because the plans were demoted. Re-tiered the evidence model: L0 source/tests/standards (ground truth) · L1 verifiable claims-about-source (#200 + this triage, must be spot-checked) · L2 inspiration (plans/perspective-docs, not evidence). Spot-checked #200's five load-bearing rows first-party against source this session (lance-graph via MCP, ndarray local): spmv_min_plus-not-tropical_spmv, sigker-PDE-not-buggy, no batched_ssd_search, no tropical_gemm, blasgraph-bit- exact/bgz17-lossy — ALL CONFIRMED. Only cosmetic defects in #200 (off-by-one line, 8-vs-9 method count). Flagged #200's own section-6 un-verified items as the residual trust boundary, not to be elevated past L1.
6 tasks
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 25, 2026
- CORRECT mis-grade: 2026-05-06-splat-osint-ingestion — SPLAT-1 types are SHIPPED in lance-graph-contract::splat (CamPlaneSplat/SplatPlaneSet/ CamSplatCertificate/AwarenessPlane16K/witness_to_splat, verified first-party), not markdown-only. PR-7 ASG-leaf must EXTEND, not reinvent. - De-provisionalize the two rows both sessions whole-read: 3DGS-render-depth-certification (grounded spec for PR-6 ErrorCertificate) and 3DGS-certified-field-kernel-substrate (added to inventory, KEEP-EPIPHANY, self-fences 'no abstraction until 2 consumers'). - CV ◑→✓ for 3DGS-SIMD-forward-renderer, 3DGS-error-certification-pillars, 3DGS-EWA-SYRK-BLAS-MKL (now whole-read by >=2 sessions). - EWA-SYRK: flag actionable backend half may DOWNGRADE after PR-3 measures (likely 3x3 pessimization). - Scope boundary: provisional flags cleared; residual #200 §6 items held at L1.
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 26, 2026
…idge.rs
The atom16 / precision-ladder / two-axes docs annotated BGZ17 as
"offset ≈17..27, stride {2,4}" — figures from a chat description that do
NOT appear in the source. First-party read of src/hpc/bgz17_bridge.rs:
BASE_DIM=17, GOLDEN_STEP=11 (coprime → table (i*11) mod 17 visits all 17
positions, Quintenzirkel-like phase), Base17=[i16;17], built by lossy
golden-step octave averaging of a 16384-bit plane.
The atom's two 2-bit fields are now labelled plan-inspired placeholders,
not grounded BGZ17 parameters. Removes a fabricated constant from a pushed
branch per the PR #200 audit discipline (source files are evidence;
chat/plan descriptions are not).
Note: this module remains an unmerged sketch — flat-disk manifold (the
ASG/KHR correction makes the directional lane spherical) and built from
the LeafPlanetarium inspiration doc. Superseded by the forthcoming
ASG-leaf canon entry; not for merge as-is.
https://claude.ai/code/session_01HbqooFZHAjaUtFEzhA1R2u
This was referenced May 26, 2026
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
2 participants
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.
Summary
Adds
.claude/PR-X12-docs-audit.md— a recommendations-only audit of the PR-X12 doc set for hallucinated / unverified claims.Contents of the audit doc:
Scope
.claude/.Test plan
.claude/PR-X12-docs-audit.mdend-to-end.https://claude.ai/code/session_01UHfLwSzGfVNzkNrTEBijuA
Generated by Claude Code
Summary by CodeRabbit