deepnsm: tri-fidelity SoA example — the graph reasoning about itself#760
Conversation
One 512-byte node co-resides all three distance representations and a route->verify RDO cascade matches brute-force NN at 97.9% fewer exact evals: FULL (raw i8 8-genre L2, the 4096^2 ground truth) rho = 1.0 PALETTE (ADC over 2 KB centroid codebook, L4 6x256:256) rho = 0.9962 MORTON (address-only shift/mask, key HEEL tile) rho = 0.5219 (flat 0.4169) MORTON and PALETTE read the SAME stored bytes (address IS the codebook index); only the distance function differs -- 'two jobs, one address' executable in one row. The RDO cascade: MORTON routes a wide net (recall 95.3%), PALETTE re-ranks to 8, FULL verifies -> 93.5% agreement with brute force at 3072 vs 147072 exact evals. Self-audit (per-lane rho vs ground truth) + self-optimize (route cheap, verify exact) = the self-optimizing space-time-aware graph on committed data. std-only, zero-dep, deterministic, clippy/fmt-clean. Records E-TRI-FIDELITY-SOA-SELF-REASONING-1. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki
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_16cc3ae6-1fd4-4fd1-be28-ffc4573b2df6) |
|
Warning Review limit reached
Next review available in: 38 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (2)
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. 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: 4e1705d5e4
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".
| for (p, &i) in ord.iter().enumerate() { | ||
| r[i] = p as f64; |
There was a problem hiding this comment.
When a lane has coarse/quantized distances (especially MORTON, where many pair distances are equal), assigning consecutive ranks to equal values makes Spearman depend on pair enumeration rather than the metric. This can misreport the self-audit rho and drive the kill gates from arbitrary ordering; tied distance groups should share an average rank, as the existing frequency example does.
Useful? React with 👍 / 👎.
| flat_x[nodes[i].heel_x() as usize], | ||
| flat_x[nodes[j].heel_x() as usize], | ||
| ); | ||
| let dy = morton_axis(nodes[i].heel_y(), nodes[j].heel_y()); |
There was a problem hiding this comment.
In the claimed FLAT-relabel baseline, the y-axis still uses the original hierarchical leaf bytes. For any dataset where the y subspace carries signal, the control preserves half of the hierarchy, so the reported flat rho and the hierarchy-vs-flat kill gate are not measuring the no-hierarchy baseline described by the example.
Useful? React with 👍 / 👎.
2 participants
The operator's directive made runnable: store all three distance representations in one 512-byte SoA node and let the graph reason about itself. Additive: one DeepNSM example + one EPIPHANY, no core change. std-only, zero-dep, deterministic, clippy/fmt-clean on 1.95.
One node, three co-resident fidelities (canon
key(16) | edges(16) | value(480))value[0..8]) — the 4096² ground truth6×256:256, here2×256:256on 8 genres)MORTON and PALETTE read the same stored bytes — the address is the codebook index; only the distance function differs. "Two jobs, one address" (
E-IMPLICIT-MORTON-TILE-1), executable in one row.The graph reasoning about itself — RDO route→verify cascade (384 words)
Self-audit (per-lane ρ vs its own ground truth) + self-optimize (route cheap, verify exact) = the self-optimizing space-time-aware graph, on committed data. Generalizes
E-X265-RDO-ARBITER-1(argmin-bits) from one axis to a lane cascade over co-resident representations.Architectural lesson (forced by the build)
Address-only distance cannot be the final answer even at a large margin (the first cut tried it — 63% agreement); it is routing, so the cascade prunes-then-verifies, never answers-from-coarse. Router net width scales with the router lane's coarseness (free evals make width cheap); the exact-eval budget is fixed by the verify width.
Changes
crates/deepnsm/examples/tri_fidelity_soa.rs— new example..claude/board/EPIPHANIES.md— prependsE-TRI-FIDELITY-SOA-SELF-REASONING-1.🤖 Generated with Claude Code
Generated by Claude Code