[AdaWorldAPI]

Summary

Lab probe crates/tract-harvest-probe (its own workspace, excluded; no native crate depends on it). It reads an ONNX graph as a graph and resolves each node through the DT64 ON × WHEN × DO schema. Tract is used only as a cold typing frontend.

• Graph. Every ONNX node is an Identity position and every value a Property position, on one 64×64 DR#3 tile. on and do_ are Rows relations, and dep = do_ ∘ read_by is computed with dr_math_binding_probe::relcomp::compose. Ordered arguments are kept beside the relations, because ONNX operands are positional.
• Registry. Each op signature (domain, op_type, dtype, attribute rules, opsets) maps to one immutable lab classid (0x0F01..0x0F51).
  • WHEN is a three-valued (K3) guard over Tract's facts and the constant operands; UNKNOWN never counts as TRUE.
  • DO is an existing lance-graph-onnx / ndarray function; no new kernels.
  • An unregistered signature is refused.
• Plan. The plan resolves once, schedules from dep, and is sealed in a FrameStore frame (source digest, registry version, coordinates). run checks the frame and then routes each node to its DO.

This replaces the fixed admission recipes from the first version of this PR (admit.rs removed), following the correction "We want the graph … resolve to polyfill and predicate methods in DT64 'on when do'". Plan §0 has the full design.

Results (against ORT with ORT_DISABLE_ALL, the literal semantics)

• Tiny fixture: bit-exact, including the 256-input div3 sweep. Where(c, X, 1) now runs and returns 17: there is no rewrite to refuse, because Where and ReduceSum are separate methods.
• Jina v5 small, layer-0 input RMSNorm (13 nodes through the graph): 9,259 of 10,240 outputs exact, max 4 ULP. ReduceMean is the only Bounded method. Every row whose mean is bit-exact is bit-exact on all of its outputs.
• Full model coverage (7,620 nodes): 3,751 resolve with WHEN TRUE, including all 113 RMSNorms. 169 are registered but WHEN is not TRUE: Tract's typed stage labels only 2,546 of the 7,620 outputs, so those facts are UNKNOWN and refused. 3,700 have no registration; the largest groups are Unsqueeze, Shape, Gather, Concat and MatMul.
• Refusals, each with its own test:
  • impossible broadcast → When { False }
  • Erf and com.example:Abs → Unregistered (the domain is part of the signature)
  • Pow exponent other than 2 → When { False }
  • f16 or f64 cast → Unregistered
  • shapes removed → When { Unknown }
  • new source or registry generation → Frame(SourceChanged / MapChanged)
  • int32 overflow → refused at run time
• Disable runs: 7, each red on exactly the named test. They cover the broadcast guard, the Pow guard, attribute rules, the domain key, the frame check, UNKNOWN-as-TRUE, and schedule from source order.
• Cost (reported, not gated): resolve 7–12 µs. The hot plan takes 3.4 µs on the tiny graph (Tract 3.6 µs) and 14.9 µs on the Jina slice (Tract 113 µs).

Open

• Graphs with more than 64 positions need several tiles.
• Facts for the values Tract leaves unlabelled.
• Registrations for the shape-plumbing ops and f32 MatMul.
• Constants are cloned on every run.
• The classids are lab-local; minting them in OGAR is a separate decision.

Test plan

• cargo fmt, cargo clippy --release --all-targets -D warnings
• cargo test --release --no-fail-fast -- --include-ignored: 20 passed (Jina tests need python3 tools/fetch_jina_slice.py; CC-BY-NC weights are never committed)
• 7 disable runs, each red on the named test
• examples/stages (full-model coverage) and examples/cost

🤖 Generated with Claude Code

https://claude.ai/code/session_01XAqKMMjKnig6MXTd41Lf2v

[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: 619033d5-0716-4c94-81aa-f6919eb641b7

• 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

_{Comment @coderabbitai help to get the list of available commands.}

[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 2026-10-30T07:28:24.775070Z	5da7678	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: 5da7678ce8

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".