[AdaWorldAPI]

C3c of D-EPI-COT. It builds on C3b (scent, #1842, merged) and is independent of C3 (#1843).

What it does

nars::scent_learn::ScentLearner uses investigation outcomes to learn which C3b leads to test first.

• It reorders leads only within a status group, so a contested lead stays ahead of a settled one (PZ8).
• Its output is an order. It holds no ledger or edge and writes no truth, Epi5 bits, CE64 bits or ledger entry (PZ4).
• Each outcome source trains it once (PZ6).

Numerics: no runtime rounding, no float roundtrip

The runtime adds no kernel; it uses the merged ndarray::hpc::codec9::EpiLane9 and ndarray::simd::lut9.

• Delta. A delta is computed in code space as M = |r|·x / 255 with integer truncation, which is odd and deterministic. Its sign comes from r.
• Zero rule. A bit mask sets the sign bit: P = present AND (positive OR M == 0). A delta of −0.49 therefore becomes −0 and then +0, never "never touched" (PZ2). lut9 applies the same rule to its output.
• Rate. The learning rate is one lut9 lookup of an odd table. Only symmetric rules are prepared; asymmetric and biased rules are refused (PZ3). Weights are added and saturated as integers and written back as SM9 codewords.
• Rounding. The only rounding happens once, offline, when the rate table is built. It follows TFLM LUTPopulateInt8: x·η, then TfLiteRound (ties away from zero), then saturation. data/tflm_lut_step_golden.txt was generated by docs/research/epistemic-cot/oracles/tflm_lut_step_oracle.cc, compiled against AdaWorldAPI/tflite-micro 5f64e4a. The table matches it on all 256 int8 codes for 4 rates.
• Spec binding. A prepared table carries its ModelSpec (schema, feature scale, rate), and a learner refuses a table built for another spec (PZ7). This binding is transient and in memory; it is not the EXO-152 PZL-4/G1 binding.

Falsifiers

There are 14 tests plus an ignored cost probe. Each of 10 disables turns a test red:

• cross-group ranking;
• no replay guard;
• no spec check;
• admitting an asymmetric rule;
• an absent feature written as touched;
• ties-to-even rounding;
• no weight write;
• dropping M == 0 from the zero rule;
• a float roundtrip with rounding for the delta;
• a rank that ignores the score.

The zero-rule disable stayed green twice before its test bound it: first no fixture had a zero delta, then the fixture's residual was positive. The test now uses a negative residual, which is the −0.49 case.

Measured

PZ5 trains on 3-decoy worlds and evaluates on 600 held-out 7-decoy worlds with the known link on the other side (standard error about 0.1):

policy	informative fixture	uninformative twin
learned	1.00	4.59
scent	4.51	4.51
round-robin	4.55	4.55
random	4.47	4.47

Earlier 60-world runs read 3.77 on the twin; that was sampling noise. Cost, on one host, release build, 8 leads (reported, not gated):

operation	time
rank	706 ns
learn	425 ns
scent	3.0 µs

NSL plans (AdaWorldAPI/neural-structured-learning #1–#8)

The NSL plans are docs-only and predate lut9. Two parts apply here:

• 05a: tables are certified offline and a query is a plain lookup. This PR follows that, with TFLM's ties-away rule deciding rational ties.
• Scratchpad §5: the TFLM rounding contract and the Codec code spaces are kept.

No lance-graph-nsl crate is added. The NSL scratchpad gates lance-graph code on a frozen NSL oracle (H7) and the G01–G12 neighbor fixtures (H1), and neither exists yet.

Gates

• Planner lib: 499 tests pass.
• fmt and clippy --lib are clean.

Not in this PR

• The informative fixture is separable, so 1.00 shows a separating feature is found, not performance on overlapping evidence.
• The training label is "hydrated". The c3-plan §5 label, "the ablation discriminated" (from the ledger's sensitivity), is the next step.
• No HIRO or Odoo outcomes are used.
• Bits 50..52 are untouched.

Owner check and claim: EXO-50 CC:session_01XAqKMPWTwovjYJEUMuGPb4/D-EPI-COT-C3C-OWNER-CHECK-20261028-01.

Draft; not for automatic merge.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XAqKMPWTwovjYJEUMuGPb4

[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: bcf2e4cd-765a-4597-9ad8-e7bec1a402d1

• 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 2026-10-28T21:51:20.076653Z	9d8fe7b	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.