[AdaWorldAPI]

Lab only. No production change. dt64, Quack, mask-risc and ndarray are used unchanged. No Gather, no new identity table, no witness store, no per-object allocation. The full write-up is in .claude/probes/2026-11-03-d-global-where/REPORT.md.

Falsifier: 21/21 (probe/ → falsify)

Two guards were disabled to prove they are load-bearing; both checks then went red:

• F4: join by local index instead of the DR route.
• F6: leave the presence bit set.

What the checks establish:

• Routing (F1–F5, F18): production DR routing (LocalMap, MapTable::{map, map_at, locate}, Frame) is exact and bidirectional. It reconciles two selections that hold the same global id at different local positions.
  • Joining the two selections by local index gives 32; the oracle and the DR-aligned path both give 14.
  • A cross-selection relation evaluates correctly across two DR256 tiles.
• Presence change (F6, F7): deleting an edge leaves the identity resolvable and turns the answer false. Deleting an unrelated edge changes nothing.
• XOR consistency (F8, F9): the forward plane XOR the transposed reverse plane is 0. A single flipped bit is found at exactly its edge.
• Frames (F10, F11):
  • An old frame replays its old layout.
  • A foreign frame, a future version and an ambiguous id are all refused.
• Two methods, one selection (F12, F13): where_dr3::LocalSelection projects bit-identically for two methods. Its SchemaChanged on a second method is therefore a coupling, not a correctness need. Scope refuses on its own.
• No allocation (F14, F15): the folds allocate 0 bytes in 0 calls.
• Boolean vs bag (F17): Boolean and bag answers differ (156 vs 1095), and both equal their oracles.

DuckDB 2.0.0 parity gate (benchmark gate preserved here)

Setup:

• 8 workloads × {DR64, DR256, multi-tile 16×DR256} × {sparse, medium, dense, skewed}.
• DuckDB uses its normal optimizer at threads 1 and 4, with PREPARE/EXECUTE and 201 warm runs. The Python call floor is subtracted.
• Our side is single-threaded.

Answers: 336/336 agree (duckdb/compare.py exits 1 on any mismatch).

Warm, best arm: 12/0/0 W/P/L for every workload. This does not establish general superiority.

• DuckDB's fixed cost per query is about 600 µs: 703 µs for a filtered count over 65,536 rows, 12,106 µs over 10M rows. That cost dominates most warm rows at these sizes (domain ≤ 4,096 nodes).
• Rows where the gap is work avoided rather than overhead:

workload	ours	DuckDB (t=1)	why
W4 3-hop, multi_dense	131 µs	1,500,724 µs	—
W4 3-hop, multi_skewed	89 µs	57,066 µs	DuckDB builds 3,726,779 join rows for a 1,164 answer (EXPLAIN ANALYZE in results/explain/)
W3, multi_dense	7.7 µs	29,097 µs	—
W6 bag count, multi_dense	119 µs	22,726 µs	—

Losses (all recorded): every loss is caused by addressing; none is caused by the fold.

• Cold (prep + first query) on every multi_* dataset: 8.7–520 ms ours against 1.9–35 ms DuckDB. DuckDB's load (40–117 ms) is excluded from its cold figure.
• W7 per-call routing: 650–970 µs, against 5 µs when the routing is done once and reused.
• In both cases the cost is MapTable::locate, which reads every listed tile per id.
• Updates on multi_dense take 4.8 ms. That beats DuckDB's DELETE but is over the workspace's 1 ms gate.

Plan choice is by measured cost:

• W3: csr wins everywhere.
• W4: csr wins on small or sparse data; selection-first frontier masks win on medium, dense and skewed data.
• Tile-first never won.

Remaining parity gap: scale ≥ 10M rows is unmeasured on our side.

Verdict

• ALREADY WORKS: DR routing, frames, presence change, XOR consistency, Quack WHERE, allocation-free folds.
• WIRE NOW (lab-proven): route a selection into the relation frame once per frame and reuse it.
• OWNER DECISION:
  • May one selection serve several MethodRefs under one Scope (EXO-46)?
  • Is a global → tile index wanted beside locate's ambiguity scan?
  • Production has no relation-presence type outside Cube.
• PROBE: ≥ 10M-row scale; building the plane via Cube::new_tiled.
• PARK: EXO-73 recursion, EXO-108 Bag, physical prefetch.
• SKIP: duplicate witnesses, object allocation, a second routing store, mask transfers between programs, new Gather operators.

Files (all under .claude/probes/2026-11-03-d-global-where/)

• SPEC.md, REPORT.md
• probe/: its own workspace, binaries falsify and bench
• duckdb/: bench_duckdb.py, compare.py
• results/: logs, JSON, compare tables, EXPLAIN ANALYZE output
• Board entry .claude/board/entries/2026-11-03-d-global-where-p0.md, with the index regenerated.

Not included: benchmark commit a84464205 (D-DR-FOLD), #1941 and #1942. No merge requested.

---

Generated by Claude Code

[coderabbitai]

Important

• Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 1c79fb02-6387-46db-90d9-0351f76da891

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn

CodeRabbit Verify

• Run CodeRabbit Verify to test the changes in this pull request.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-03T18:27:05.549204Z	65c89b3	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 65c89b3264

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".