Review: "Where Should an Agent Correction Live?" (ACL 2027)

Summary
When an agent fails, the fix can go into memory, a skill, a runtime rule, or abstention. MIGRATE says the choice should be measured on matched task streams and gated on regressions, not assumed, and tests this on ALFWorld and WebShop-100k with a scripted controller.

Strengths
- The question is real, and almost nobody measures it.
- Paired recovered/regressed counts (Table 16) beat headline success rates.
- Locked goal manifests, checksums, a claim-verifier script, a reported negative result.

Weaknesses
- No language model anywhere (Appendix H). Memory scores 0 because the scripted controller never reads the note. That is the harness, not memory. ExpeL and Agent Workflow Memory move ALFWorld, WebShop, and WebArena numbers with notes an LLM reads.
- The six ALFWorld "clusters" are the six task types of the valid-unseen split (24+18+31+23+21+17 = 134). The "skill" is a hand-written solver per type; 134/134 is a solver, not a repair.
- The WebShop cells are policies of very different size (a buy-time guard against the whole V3 controller), not one IR compiled four ways. The matrix measures effort, not placement.
- Protected stream of 14 tasks; the skill cell is rejected on 4 flips.
- The gated policy nets 451/1021 against 498 raw. The gate costs 47 tasks; the paper never says so, nor prices a regression against a recovery.
- Held-out gain is +1.86 pp over 20 splits, yet the full-set gated comparison is +13.71 pp (p = 5.66e-36). Either the latter is in-sample or the protocols differ. Say which.
- Missing prior work: Xie et al. (ACL 2021), negative flips; SPIBB (Laroche et al., 2019); AgentSpec (ICSE 2026); ACE (ICLR 2026); Wen et al. (TACL 2024), abstention.

Detailed review
Novelty. Placement as a gated decision is a clean framing I have not seen for LLM agents. The parts are old: regression-free updates, safe policy improvement, regression-test gating, rule DSLs. Framing novelty moderate; technical novelty low.

The experiments do not test the framing; three of four cells are fixed by construction. What would:

1. Put an LLM agent in the loop so every cell is live: a retrieved note, a callable skill, an AgentSpec-style guard, a confidence-based abstain. AgentErrorBench (Zhu et al., cited) has LLM failures on both benchmarks.
2. Hold engineering effort roughly equal across cells; print one IR with its four compiled artifacts.
3. Use the 311 rerank successes as the protected stream everywhere, state the gating cost, and fit the gate cross-validated.
4. Estimate the lambdas in Eq. 1 or delete it. It is never used.
5. The heuristic router matches 0/6. Learn one held out by correction family, or stop calling an argmax a router.
6. Prose: the paper twice tells reviewers what not to read (Sec. 4.2, App. I.1); "load-bearing" appears seven times; the abstract is a results table in sentence form.

Not ready for ACL. With an LLM controller and equal-effort compilers, the same framing could carry a strong paper.
