[AlexanderFengler]

Stacked on #1292: this branch contains its commits and shrinks to its own six once #1292 is merged. Review the rl commits only.

Purpose. #1292 lets a likelihood config declare which parameter shifts the lower edge of the response-time support ("ndt_edge_shift": {"param": ..., "scale": ...}, edge at t − scale·param). The RL layer built its distribution without that declaration, so an RLSSM whose decision process has non-decision-time variability would have floored its likelihood at t. No bundled RL backbone has such a parameter today; this closes the gap before ddm_st / ddm_normal_st (#1346, #1345) exist.

Design. The declaration is a property of the decision process, so it travels with the decision-process spec and the RL layer only forwards it:

• _build_ssm_spec_from_modelconfig copies the backbone's approx_differentiable declaration into the spec; register_ssm(..., ndt_edge_shift=None) accepts one for custom SSMs and validates it against list_params_ssm.
• get_rlssm_model_config passes it to RLSSMConfig, and raises a clear ValueError when the named parameter (or t) is computed per trial by the learning process, since the support edge cannot be resolved from a parameter the distribution never sees.
• RLSSMConfig gains the field (from_rlssm_dict reads the optional key; from_ssms_model takes it from the decision-process spec; a ModelConfig decision process hands over its own); validate() applies the same rules as Config, via a shared helper moved out of Config.validate (behaviour unchanged, tests/test_config.py untouched).
• RLSSM._make_model_distribution forwards it to make_distribution.
• aDDM is untouched (no ndt variability, separate config class).

Tests. 22 new cases in tests/rl/: the spec carries the backbone's declaration (and None without one); register_ssm stores it and rejects bad declarations; get_rlssm_model_config propagates it and rejects a computed parameter; RLSSMConfig validation, dict round trip, from_ssms_model, ModelConfig inheritance and explicit-override precedence; RLSSM forwards it to make_distribution; and end to end, the compiled RL log-likelihood is exactly LOGP_LB for rt ≤ t − st and finite above with a declared {"param": "st", "scale": 1.0}. Each test fails with its source change reverted.

Checks run. tests/rl (244 passed, 4 skipped), tests/test_config.py, tests/distribution_utils (slow included), fast suite (1274 passed); ruff check, ruff format --check, mypy src/hssm clean at every commit; pyrefly shows only main's two pre-existing errors. In-memory merges: clean against main and against #1293 on top of main + this head; no conflicts beyond those #1292 already has with #1344/#1345/#1346.

Not in this PR. Changelog entry (written on main after the merge wave, with #1292's); any ssm-simulators-level declaration of the support edge (under discussion separately).

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added configurable support-edge adjustments for response-time likelihoods. Models can specify a parameter and scale to set the edge below non-decision time; responses at or below the edge receive the likelihood floor, while missing responses remain unaffected.
  • Applied the adjustment to the full-DDM model and made it available for registered models and reinforcement-learning models.
• Documentation
  • Clarified how support-edge adjustments work and how the full-DDM model’s non-decision-time parameter affects its response-time support.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: eba11853-4e5f-4b93-b535-c0775ca78680

📥 Commits

Reviewing files that changed from the base of the PR and between cd26b0d and 406bf33.

📒 Files selected for processing (15)

• src/hssm/_types.py
• src/hssm/config.py
• src/hssm/distribution_utils/dist.py
• src/hssm/hssm.py
• src/hssm/likelihoods/blackbox.py
• src/hssm/modelconfig/full_ddm_config.py
• src/hssm/register.py
• src/hssm/rl/config.py
• src/hssm/rl/registry.py
• src/hssm/rl/rlssm.py
• tests/distribution_utils/test_distribution_utils.py
• tests/rl/test_registry.py
• tests/rl/test_rlssm.py
• tests/rl/test_rlssm_config.py
• tests/test_config.py

Included review availability: This review used your included allowance. Your plan provides up to 2 included reviews per hour; 1 remain after this review.

---

📝 Walkthrough

Walkthrough

Adds optional support-edge metadata that defines a response-time edge as t - scale * param. Standard and RLSSM distributions use that edge to determine when to floor log-probabilities. The full_ddm configuration declares st with a scale of 0.5.

Changes

NDT Support-Edge Flooring

Layer / File(s)	Summary
Define and validate support-edge declarations src/hssm/_types.py, src/hssm/config.py, src/hssm/modelconfig/full_ddm_config.py, src/hssm/likelihoods/blackbox.py, src/hssm/register.py, tests/test_config.py	Adds the NDTEdgeShift type and configuration fields. Validation requires t and the named parameter in list_params, and a finite, non-negative numeric scale. The full_ddm configuration declares st with scale 0.5.
Resolve and apply the response-time edge src/hssm/distribution_utils/dist.py, src/hssm/hssm.py, tests/distribution_utils/test_distribution_utils.py	Resolves the edge from t and the optional declaration, then applies the floor in both likelihood paths. HSSM passes the declaration to distribution construction. Tests cover edge boundaries, missing responses, and the full_ddm declaration.
Carry declarations through RLSSM configuration src/hssm/rl/config.py, src/hssm/rl/registry.py, tests/rl/test_registry.py, tests/rl/test_rlssm_config.py	Copies support-edge metadata through RL registry and configuration paths. Validation checks SSM parameters and rejects declarations that refer to parameters computed by the learning process.
Apply support-edge flooring in RLSSM src/hssm/rl/rlssm.py, tests/rl/test_rlssm.py	Passes the configured declaration into distribution construction. Tests check forwarding and compiled log-probabilities around the declared edge.

Priority: ➖ Normal

Estimated code review effort: 3 (Moderate) | ~25 minutes

Change: Feature

Sequence Diagram(s)

sequenceDiagram
  participant HSSM
  participant make_distribution
  participant HSSMDistribution
  HSSM->>make_distribution: pass ndt_edge_shift
  make_distribution->>HSSMDistribution: configure distribution
  HSSMDistribution->>HSSMDistribution: resolve edge from t and shift declaration
  HSSMDistribution->>HSSMDistribution: floor logp at or below edge

Loading

Suggested reviewers: krishnbera

Merge Risk: ⚪ Minimal · up to 406bf

No identified support-edge issue remains that should prevent merging after normal checks.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the RLSSM decision-process propagation of ndt_edge_shift, which is a central part of the stated objective and changeset.
Docstring Coverage	✅ Passed	Docstring coverage is 82.80% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 93 functions across 15 files.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autopilot · 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.}