docs: record the RLVR training-loop insight in the LLM section

[PhysShell]

Adds a "The deploy loop is a training loop (RLVR) — and its one trap" subsection to the LLM section of docs/notes/research-landscape-2026.md. Docs-only (+39 lines).

Captures the insight that the fix-loop's proposal → verdict pairs are free labelled examples, so deployment is data generation (RL from Verifiable Rewards):

• The reward is a sound deterministic verifier, not RLHF's fallible/gameable learned reward model — the clean-signal regime where RLVR actually works.
• Free bootstrap: deterministic fixes are guaranteed-correct demonstrations to SFT on; rejection sampling (keep proposals that pass checker+compile+tests, fine-tune on winners) is the stable start, no RL infra needed.
• The trap (load-bearing): reward hacking / Goodhart — "checker went quiet" ≠ "fixed correctly" (the model learns to suppress diagnostics / delete the resource / remove the failing test). So the metamorphic harness becomes part of the reward function (checker-green ∧ behaviour-preserved); the verifier sets both the ceiling and the floor.
• Re-confirms fix > spec from the training side: a fix has a verifiable reward; a mined spec does not.

Continuation of the LLM-layer thread from #41; same framing (verifier-gated, deterministic-first, falsifiable assistant).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Rg8kSk1YT14x7A1vo5zgED

---

Generated by Claude Code

Summary by CodeRabbit

• Documentation
  • Expanded the research landscape notes with an analysis of deployment loops as reinforcement learning, including verification-based reward mechanisms, the risk of reward hacking/Goodhart failures, and the need for behavior-preservation criteria to maintain correctness.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: c6ba63c7-109b-4368-95d5-ea90fbda955a

📥 Commits

Reviewing files that changed from the base of the PR and between 5ca702e and 2fcd79f.

📒 Files selected for processing (1)

• docs/notes/research-landscape-2026.md

✅ Files skipped from review due to trivial changes (1)

• docs/notes/research-landscape-2026.md

---

📝 Walkthrough

Walkthrough

A new subsection is appended to docs/notes/research-landscape-2026.md that reframes the Own.NET fix/refactor loop as RL from verifiable rewards (RLVR). It describes deterministic checker + compile + test signals as labels, flags the Goodhart/reward-hacking failure mode, mandates a metamorphic behavior-preservation harness in the reward criteria, and restates "fix > spec."

Changes

RLVR documentation addition

Layer / File(s)	Summary
RLVR subsection: deploy loop as training loop docs/notes/research-landscape-2026.md	Adds lines 122–160 framing the fix-loop as RLVR data generation with rejection-sampling selection, documents the Goodhart/reward-hacking risk, specifies the metamorphic harness as a mandatory reward component, and reaffirms "fix > spec" from a training-signal perspective.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

Possibly related PRs

• PhysShell/Own.NET#41: Also modifies docs/notes/research-landscape-2026.md to frame Own.NET's verifier-first loop as RLVR and references the metamorphic behavior-preservation harness and Goodhart warning — directly overlapping content with this PR.

Poem

🐇 Hop, hop, the loop goes round,
Verifiers check each patch we've found.
Goodhart lurks with tricky bait,
But metamorphic tests hold straight!
Fix beats spec — the reward's clear,
The rabbit trained without fear. 🥕

🚥 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 'docs: record the RLVR training-loop insight in the LLM section' directly and clearly summarizes the main change: adding documentation about the RLVR (Reinforcement Learning from Verifiable Rewards) training-loop insight to the LLM section of the research landscape notes.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
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.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/research-rlvr-note

---

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

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/notes/research-landscape-2026.md`:
- Around line 124-125: The document contains inconsistent spelling variations of
the same word: both "labelled" (British spelling) and "labeled" (American
spelling) appear in the same section. Choose one spelling variant (either
British "labelled" or American "labeled") and search through the entire section
containing the "proposal → verdict pair is a labelled training example" sentence
to find and replace all instances with the chosen variant for consistency
throughout.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 106cc323-7695-4e00-8a5f-5225b81edf83

📥 Commits

Reviewing files that changed from the base of the PR and between bd6e3e7 and 5ca702e.

📒 Files selected for processing (1)

• docs/notes/research-landscape-2026.md