SLM-128 (LDI3-01): complete structured objectives — ambiguous/unobserved mass, locality tethers, role weighting

[Tyler-R-Kendrick]

Completes the LDI3-01 structured-objective library already on main (from #359) with the acceptance criteria it did not yet cover, reconciling in the remaining features from the parallel SLM-128 work (my closed #355). All additions are backward compatible — new config fields default off, new params default None — so #359's behavior and its tests are unchanged (14 existing + 3 new = 17 green; full tests/test_harnesses/preference = 186 green).

This is the "merge the best results / resolve conflicts" reconciliation: rather than a competing duplicate, it takes the best of both implementations into the one on main.

What #359 was missing (now added)

• Ambiguous/unobserved actions — _prepare extracts them from the ObjectiveView, validates them inside the legal set, enforces disjoint good/bad/ambiguous/unobserved partitions, and reports their legal mass separately (the issue requires the four masses reported separately and never mislabeled). LDI3-01: structured local-preference objectives (SLM-128) #359 previously ignored these partitions.
• Reference/locality tethers — new non_target_tether / target_tether / target_grace config + _locality_tether, composed onto every objective in structured_decision_loss and metered independently of the barrier so the target anchor never double-counts (an explicit acceptance criterion LDI3-01: structured local-preference objectives (SLM-128) #359 lacked entirely).
• Semantic role weighting — wires the previously-dead default_role_weight config field and adds an optional good_role_weights so structural tokens can be down-weighted in the barrier anchor; reports mean_role_weight.

Tests

3 added: partition-mass reporting + legal-set validation; tether composition + separate metering + reference requirement; role-weight down-weighting + honored default_role_weight. ruff + python -m scripts.repo_policy clean.

Honesty

Additive completion only. No model update, no matrix run, no quality claim; existing objective math untouched.

🤖 Generated with Claude Code

https://claude.ai/code/session_01U5Fy8PnMTV54p675y9T3B8

---

Generated by Claude Code

Summary by CodeRabbit

• New Features

  • Added support for tracking ambiguous and unobserved legal actions separately in objective metrics.
  • Added configurable role weighting for preferred actions in barrier-based objectives.
  • Added optional locality tethering for target and non-target actions using reference logits.
  • Added configuration controls for tether strength and target grace periods.

• Validation

  • Improved validation for action partitions and locality-tether configuration.
  • Added metrics for partition legal mass, role weights, and locality loss.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
slm-training	Ready	Preview, Comment	Jul 18, 2026 1:25pm

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 95af42e5-ca7f-40c5-add4-e6abee29f4f9

📥 Commits

Reviewing files that changed from the base of the PR and between 21b2611 and 37c2923.

📒 Files selected for processing (2)

• src/slm_training/harnesses/preference/structured_objectives.py
• tests/test_harnesses/preference/test_structured_objectives.py

---

📝 Walkthrough

Walkthrough

Structured preference objectives now preserve ambiguous and unobserved legal partitions in metrics, support per-action barrier role weights, and optionally compose target or non-target locality tether losses against reference logits.

Changes

Structured objective extensions

Layer / File(s)	Summary
Legal partition accounting src/slm_training/harnesses/preference/structured_objectives.py, tests/test_harnesses/preference/test_structured_objectives.py	Preparation validates disjoint legal partitions and reports ambiguous/unobserved legal masses and counts, with tests covering normalization and invalid IDs.
Role-weighted barrier src/slm_training/harnesses/preference/structured_objectives.py, tests/test_harnesses/preference/test_structured_objectives.py	The TAB barrier applies per-good role weights or the default role weight and reports mean_role_weight; tests verify reduced barrier loss.
Locality tether composition src/slm_training/harnesses/preference/structured_objectives.py, tests/test_harnesses/preference/test_structured_objectives.py	New tether configuration fields enable target and non-target MSE losses against reference logits, with separate metrics and missing-reference validation.

Estimated code review effort: 3 (Moderate) | ~20 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 15.38% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title is concise and accurately summarizes the main structured-objective additions.
Linked Issues check	✅ Passed	The changes align with #355 by adding ambiguous/unobserved handling, locality tethers, role weighting, and validation/tests.
Out of Scope Changes check	✅ Passed	No unrelated code changes are evident beyond the structured-objective and test updates described.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/slm-training-issues-d8l0ba

---

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.}