W0 (D-DCR-0): the replay budget, measured — the mask half dominates 3.8× - #1118
Merged
Conversation
Harness: lance-graph-contract/examples/dcr_w0_replay_budget (release, deterministic LCG — a probe for a replay plan is itself replayable). Corpus magnitudes are READ from dismech-causality-v3-v1 §11, not re-derived: 2,449 oracle edges / 534 diseases, 4,076 restraint, 361 unknown. Measured: - step (NarsTruth::revision + EvidenceMask::intersection): 14,285 steps/ms at a 4096-wide candidate set (70 ns/step). - branching shrink 1.66x per evidence item, flat 10^3..10^5 — labelled FIXTURE-SET: it measures the mechanism, never the corpus's real discriminative power (that needs D-CV3-0..2's frozen TSVs). - KILL check did NOT fire: scan 0.906 ms vs decision 0.008 ms, crossover ~25 chains, corpus ~98x above it => W5 stays live on cost (still HELD on the operator rung 5-9 ruling, a different gate). - kernel split, the ALU wave's actual question: revision 24.0 ns vs 4096-bit intersect 90.6 ns => MASK dominates 3.8x. The 64x64 tile is aimed at the half that costs; the whole oracle arm replays in 2.74 ms, so BUY is stated at >10x that in one budget. Correctly aimed, correctly deferred. Also ⊘ CORRECTION: the plan's §0 audit was under-cited — it swept the sibling repos (F-RLR-11) but not this repo's own contract module list or plans index, missing contract::dismech_evidence (which already carries the hidden-mediator oracle + epistemic-restraint control populations) and dismech-causality-v3-v1 §11 (the held-out benchmark, measured three ways). Nothing is contradicted; the cost was scoping, and W1-W3 now consume those arms instead of inventing a corpus. Plan §3a/§3b, EPIPHANIES prepend, STATUS_BOARD D-DCR-0 Shipped + D-DCR-0a. Gates: contract 1278/1278, clippy -D warnings clean, fmt clean, supersession index regenerated last.
|
Warning Review limit reachedNext included review available in 24 minutes. View limit detailsLimit details: You’ve used the included review currently available. Your 63 included PR review attempts over the past 7 days set your current allowance at 1 review per hour. Your organization has reached its usage spending cap. Adjust your spending cap in the billing tab. Review configuration: ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (4)
Comment |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_bdd39c45-4bb3-4413-8a1c-b5d33ee0990e) |
![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: da07d0ac4a
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".
…masks, both KILL scales All three codex findings were valid, and a fourth surfaced while fixing them. None was a bug in the code under test; every one was a defect in the probe's FIXTURE: P1 the eval timed NarsTruth::revision (f32, contract-side) where the plan defines it as NarsTables lookup + CausalEdge64 revision. Disclosing the substitution did not make it the promised measurement. Probe MOVED to lance-graph-planner, where causal-edge is reachable, and now times NarsTables::revise + CausalEdge64:: forward: 34.7 ns/step (28,818 steps/ms). P1 the mask fixture was Bits(Vec<u64>) — a heap alloc inside every timed intersection — while impl EvidenceMask for [u64; N] already ships (revision.rs:70) and IS the p64 shape ([u64;64] = 4096 bits). Both are now timed and the delta reported: 61.5 ns alloc-free vs 73.4 ns allocating (~16%). P2 the KILL gate ran at 2,449 chains where the plan pre-registers 10^5. Both scales now run at one candidate width; it fires at NEITHER (10^5: 13.88 ms scan vs 0.007 ms decision; 2,449: 0.340 ms). Crossover ~53 chains. + self-caught: dense_mask(rng, 1) sets NO bits (x % 1 == 0), so the frontier decision was scored against an empty live set. all_ones() is now its own constructor; dense_mask asserts one_in >= 2. Net: MASK dominates by 1.77x, NOT the 3.8x v1 reported. The direction survives (a 64x64 tile is aimed at the half that costs) but the margin is under half the first claim, so the ALU case is materially weaker than v1 said; BUY stands at >10x the 1.36 ms oracle arm. Records are append-only: plan §3b kept and superseded by §3c with the reason; STATUS_BOARD D-DCR-0 superseded in place; EPIPHANIES prepend E-EVERY-DEFECT-IN-A-MEASUREMENT-WAS-IN-ITS-FIXTURE-NOT-ITS-CODE-1 (23888 -> 23935 lines, verified non-truncating). Gates: planner 370/370, my example clippy -D warnings clean, fmt clean. NOTE: workspace --all-targets clippy is red on probe_r2il_* examples from #1051 — pre-existing, untouched here, and rust-test.yml documents that a workspace -D warnings step would be red.
…d heat, and the reduction is why Operator question, answered with the W0 harness instead of architecture. Three objects are called 'masking' here and they are not the same kind: - ogar_r2il::CallMask is [u64; 3] (192 bits, lazy word-tests, no alloc) => needs nothing; a slice-API call would cost more than the work. - the replay candidate set is [u64; 64] (4096 bits) => the real candidate, and W0 measures it dominating the promised step ~1.8x. - ndarray::hpc::jitson is a JSON-config -> Cranelift native SCAN kernel compiler, not a masking library. Relevant to W5's frontier decision, never to the mask half. Measured at 4096 bits: scalar [u64;64] 65.2 ns vs ndarray simd_int_ops::mask_and (U64x8) 60.4 ns = 1.08x, a DEAD HEAT. The decomposition says why: SIMD and 11.1 ns + scalar popcount 56.7 ns, so the POPCOUNT is 5.1x the AND. Borrowing the AND changes nothing because the AND was never the cost. The primitive that would pay does not exist: a fused mask_and_popcount(&[u64], &[u64]) -> u32 keeping the AND in registers and reducing with VPOPCNTDQ. ndarray has popcnt on its AVX-512 typed wrapper but no fused slice API, and the workspace SIMD invariant means it must be added IN ndarray, never hand-rolled here. Surfaced, not filed — upstream asks are the operator's call. Consequence: the deferred p64 64x64 wave's TARGET moves (fuse the reduction, don't accelerate the op); its BUY threshold is unchanged. Probe extended with the ndarray arm + decomposition; a >1.15x/<0.87x band replaces the earlier faster/slower label, which called 1.00x 'FASTER'. Gates: planner 370/370, example clippy -D warnings clean, fmt clean.
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.
W0 of
dismech-causal-replay-v1(#1117). One committed, re-runnable probe —lance-graph-contract/examples/dcr_w0_replay_budget— plus the plan/board record. No production code, no new carrier: the probe's wide bitset is an example-local fixture implementing the existingEvidenceMasktrait (the trait exists to be implemented;F-RLR-2is about the production path).The corpus scale is read, not invented
dismech-causality-v3-v1.md§3a/§11 already measured the supervision corpus three independent ways: 2,449 oracle edges over 534 diseases, 4,076 restraint rows, 361 unknown rows. W0 reports against those magnitudes.Measured
1. Step throughput (
NarsTruth::revision+EvidenceMask::intersection): 29,184 steps/ms at a 64-wide candidate set, 14,285 steps/ms at 4,096-wide (70 ns/step).2. Branching shrink: 1.53×–1.66× per evidence item, flat from 10³ to 10⁵ candidates — explicitly labelled fixture-set (2/3 support, 1/10 refute density). It measures the mechanism's cost and scaling, never the corpus's real discriminative power; that needs the frozen oracle/restraint TSVs (D-CV3-0..2, consumer-side).
3. KILL check — did NOT fire. Full scan of the oracle arm (2,449 chains, len 4) = 0.906 ms; one frontier decision over 64 observations = 0.008 ms, ~100× cheaper, crossover at ~25 chains. The corpus sits ~98× above crossover ⇒ W5 stays live on cost grounds (it remains HELD on the operator rung 5–9 table — a different gate). The plan pre-registered the opposite outcome as a real possibility; it was checked, not assumed.
4. Kernel split — the ALU wave's actual question.
revisionalone 24.0 ns vs a 4096-bitintersection+count at 90.6 ns ⇒ MASK dominates by 3.8×. The deferred p64 64×64 tile accelerates the half that actually costs, so the "4096 bits = one node's budget" shape argument survives measurement — which is not the usual outcome.5. ALU BUY threshold (what W0 owed). The whole oracle arm replays in 2.74 ms at chain length 16 ⇒ BUY only when a workload sustains >10× that in one budget (≈143,000 steps/ms). Correctly aimed, correctly deferred.
W0 also states what it did not measure, so nobody cites it as if it had: real per-evidence discriminative power,
CausalEdge64's packed step (planner-side, one dependency layer out of this zero-dep crate), and loco dispatch cost (covered byogar-dismech's own tests; duplicating them here would be a second truth).⊘ Correction — #1117's §0 audit was under-cited
It swept the sibling repos (F-RLR-11) but not this repo's contract module list or plans index, missing
contract::dismech_evidence(686 LOC, shipped — which already carries the hidden-mediator oracle and the epistemic-restraint control, a ready-made two-sided falsifier) anddismech-causality-v3-v1.md§11 (the held-out benchmark). Nothing in the plan is contradicted — "no baked slab exists" still holds. The cost was scoping: W1–W3 would have invented a falsifier corpus that §11 already specifies. The join is now the plan's §3a: D-CV3's benchmark IS D-DCR's falsifier; D-DCR is the engine it grades. A sibling-repo sweep does not discharge the same-repo sweep — rubicon §F's own ⊘ made the identical mistake five days earlier.Boards in the same commit: plan §3a/§3b,
EPIPHANIESprepend (verified non-destructive: +61 lines, prior head entry intact),STATUS_BOARDD-DCR-0 Shipped + new D-DCR-0a; supersession index regenerated last.Gates: contract 1278/1278,
clippy --all-targets --no-deps -D warningsclean, fmt clean.🤖 Generated with Claude Code
https://claude.ai/code/session_01PFnYKqw6d7TTiB9cT8eFdK
Generated by Claude Code