{"path":"research/one-inference-per-link/report-methods-for-one-inference.md","content":"# Research report: methods to enforce or check that a generated step is a single inference\n\n*A research helper's report (Sonnet), 2026-10-24, saved verbatim for `../one-inference-per-link.md`. Its brief: methods with measured results for enforcing or checking one inference per step (scheme-constrained steps, form checks by code, defeasible-logic solvers, small step verifiers, constrained output, step size), skipping what the corpus already held. Read-status tags and cost ratings are the helper's own.*\n\n---\n\n# Methods to enforce or check that a generated step is a single inference: measured results, 2022–2026\n\nTags: [FULL] read in full, [PART] read in part (I extracted the PDF text and read the method and result sections and tables), [ABS] abstract only, [SEC] secondary source. All numbers below were read by me in the cited source unless tagged otherwise. Items marked \"preprint\" are not peer-reviewed. Where I give a cost or stringency rating of my own, it is labelled as my estimate.\n\n## Headline findings\n\n1. **No work I found measures \"is this step exactly one inference\" for defeasible natural-language steps.** The nearest measured attempts use formal-logic tasks with a fixed rule set (FineLogic), purely deductive chains (Sadeddine & Suchanek), or human annotation of argument form (Saadat-Yazdi & Kökciyan).\n2. **Prompting alone produces atomic chains rarely.** FineLogic (Findings EMNLP 2025) reports that in only 1.6–5.7% of chains were all steps atomic, for GPT-4o, DeepSeek-R1 and Llama-3.1-8B few-shot. Fine-tuning on structured symbolic steps raised this to 19.8–25.1%.\n3. **A template that forces a rule–case–conclusion layout makes scheme conformity trivially checkable, but conformity is not validity.** On template-generated arguments, LLMs classify the scheme at 0.94–1.00 F1. On natural arguments the same task falls to 0.5–0.65 F1 for four to seven schemes, and to about 0.29 F1 for 24 schemes in dialogue.\n4. **Cheap checks that work are high-precision, low-recall screens.** Examples are a pattern-based deductive checker (precision 0.97, recall 0.26 on FOLIO) and fine-tuned NLI cross-encoders (precision 0.72–0.87, recall 0.56–0.87).\n5. **Asking an LLM \"is anything missing / is this a leap?\" is weak.** Judges over-credit incomplete chains (C2-Faith), and zero-shot leap detection reached only about a third precision and recall even on synthetic deletions (CoT-Bridge).\n\n## 1. Argument-scheme-constrained step generation\n\n**LLM scheme classification on real arguments (the checking side)**\n- Bezou-Vrakatseli, Cocarascu & Modgil, \"Can LLMs Understand Argument Schemes?\", [Findings ACL 2025](https://aclanthology.org/2025.findings-acl.702/) [PART]. They tested 7 LLMs with Walton's schemes at temperature 0.\n  - EthiX holds 686 Kialo arguments across 7 schemes. Best macro-F1 was Claude-3.5 few-shot with LLM-written scheme descriptions: 0.65 F1 / 0.65 accuracy. GPT-4o and GPT-4 were 0.54–0.57 F1. Llama-3.2 (3B) was 0.24–0.26 and Mistral 0.35–0.40.\n  - On the 4-scheme subset (cause-to-effect, example, expert opinion, analogy), the best was about 0.69 F1.\n  - Chain-of-thought gave similar or worse results for most models (up to −8 F1 for Claude).\n  - Most confused scheme: argument from cause to effect, confused with consequences and example.\n  - Enthymemes cost a little: Claude few-shot fell from 0.61 to 0.57 F1.\n  - On NLAS, a corpus where GPT-3.5/4 generated arguments directly from scheme templates, 4-scheme F1 was 0.94–1.00 for the strong models. Text produced from a template is easy to classify back to that template.\n- Ruiz-Dolz, Kikteva & Lawrence, [ACL 2025](https://aclanthology.org/2025.acl-long.368/) [PART]. QT-Schemes has 441 dialogue arguments labelled with 24 schemes.\n  - Best LLM was Llama-3.3-70B few-shot at 29.4 macro-F1 (zero-shot 18.7).\n  - Fine-tuned RoBERTa pre-trained on NLAS reached 62.3 F1, but only after grouping the schemes into families.\n  - Their conclusion: LLMs \"struggle to generalise\" from textbook-like to dialogue arguments.\n- ArgBench, Ajjour et al., [arXiv 2604.17366](https://arxiv.org/abs/2604.17366) (preprint, Apr 2026) [PART]. Scheme classification on ArgU, 7 labels, with Mistral-22B, Qwen3-32B and Llama-3.3-70B: macro-F1 0.13–0.33, depending heavily on prompt style.\n- Saadat-Yazdi & Kökciyan, [ACL 2024](https://aclanthology.org/2024.acl-long.520/) [PART]. This is the most relevant slot-filling study. It uses Wagemans' Periodic Table of Arguments, with canonical forms with named slots (a, X, Y) rewritten from Kialo claims.\n  - Human annotators trained on the tool agreed only moderately: Gwet's AC1 was 0.58 for argument form and 0.50 for validity.\n  - Rewriting an argument into canonical form (Rouge-1): fine-tuned Flan-T5-large 72.1 F1, human 78.6, few-shot Llama-2-13b about 57. Pre-trained LLMs \"are not able to model\" canonicalisation.\n  - Forcing output as a list of slot assignments (`{'a':…, 'X':…, 'Y':…}`) \"works relatively well\". But fine-tuned Llama-2 generated \"nonsensical phrases while conforming to the required output structure\". This is a measured case of conformity without validity.\n\n**Scheme-guided generation (the generation side)**\n- CQs-Gen shared task overview, Calvo Figueras et al., [ArgMining 2025](https://aclanthology.org/2025.argmining-1.23) [PART]. 13 teams.\n  - The winner scored 67.6 (manual evaluation); the two baselines scored 44.1 and 52.9.\n  - Three of the top four teams used scheme annotations. Two other teams (Webis, ARG2ST) found them insufficiently useful and dropped them.\n  - Scheme descriptions raised question quality but reduced diversity and made outputs \"rigid and templated\".\n  - Quality of the scheme definition mattered. One team's weak templates for practical reasoning and expert opinion hurt it.\n- DayDreamer (Zhou, Saadat-Yazdi, Kökciyan), [arXiv 2505.15554](https://arxiv.org/abs/2505.15554) [PART]. The LLM instantiates the scheme template to get a structured argument, then generates critical questions from the scheme's templates. It placed 4th of 13.\n  - Validation, with GPT-4o-mini (percent Useful / Unhelpful / Invalid): plain prompt baseline 72.0 / 13.8 / 3.9. Their pipeline, with four scheme templates withheld, 72.2 / 8.8 / 2.9.\n  - So scheme completion mainly cut Unhelpful questions; it did not raise Useful ones.\n  - LLMs hallucinated when given more than two scheme templates at once, and a wrong template misled extraction.\n- Critical-Questions-of-Thought, Castagna et al., [arXiv 2412.15177](https://arxiv.org/abs/2412.15177) (preprint) [PART]. Toulmin-style critical questions inside a reasoning loop gave about +4.7% (reasoning) and +5.7% (math) on MT-Bench. These are answer-accuracy gains judged by LLMs, not step validity.\n\n**Takeaway.** Nothing I found automatically checks slot filling against meaning. The checkable part is surface conformity (slots present, scheme label recovered). A forced instantiation also makes the scheme classifier near-perfect, because the classifier is reading back its own template.\n\n## 2. Deterministic or near-deterministic checks of a step's form\n\n- **Pattern-based deductive checker.** Sadeddine & Suchanek, \"Verifying the Steps of Deductive Reasoning Chains\", [Findings ACL 2025](https://aclanthology.org/2025.findings-acl.25/) [PART]. The benchmark has 1,409 chains and 5,194 steps, with step validity annotated; inter-annotator agreement was 98.6% (validity annotated by the authors).\n  - Their \"VANESSA symbolic\" method parses sentences with a constituency parser and hand-written tree regexes into ∧/∨/⇒/∀ over atomic literals, then runs a natural-deduction solver. Precision and recall (F0.5) were:\n    - ProofWriter (synthetic chains): 0.99–1.0 and 0.99–1.0 (F0.5 0.99).\n    - EntailmentBank with injected errors: precision 0.73–0.78, recall 0.04 (F0.5 0.18).\n    - FOLIO (real reasoning chains): precision 0.94–0.97, recall 0.16–0.26 (F0.5 0.47–0.62).\n  - Adding an NLI model to compare the atomic literals (VANESSA + LLaMa-3-8B) gave on FOLIO: precision 0.74–0.81, recall 0.60–0.69, F0.5 0.73–0.76.\n  - Deterministic form checks are precise and almost useless for recall as soon as phrasing varies.\n  - A caution from the same table: the \"answer baseline\" (treat every step as valid if the final answer is right) scored F0.5 0.71 on FOLIO with LLaMa-3 chains, close to the real methods (0.69–0.78).\n- **Classic natural logic.** NeuralLog (Chen, Gao & Moss, [*SEM 2021](https://arxiv.org/abs/2105.14167)) [PART] pairs a monotonicity engine with neural phrase alignment.\n  - SICK accuracy 90.3% (full system); monotonicity-only ablation 74.7%.\n  - MED: 93.4% overall (91.4% upward, 93.9% downward). Pure-logic systems earlier scored 58–68% on downward inferences.\n  - Failures came from dependency-parse errors on long conjunctions and missing background knowledge.\n  - These datasets are lexical-substitution entailments in simple sentences, not defeasible, multi-clause inferences.\n- **MonaLog** (Hu et al., SCiL 2020, [arXiv 1910.08772](https://arxiv.org/abs/1910.08772)) [SEC]. It was competitive on SICK. It is older and sits outside the 2023–26 window.\n- **LLM-era natural-logic work.** I found nothing from 2023–26 that checks LLM steps with natural-logic proofs. The closest is Basmov et al., [arXiv 2305.14785](https://arxiv.org/abs/2305.14785) [ABS/PART]. It reports \"moderate to low\" LLM performance on monotonicity entailments, and that wrapping the premise in presupposition triggers or non-factives shifts predicted labels regardless of the true relation.\n- **Modus ponens as a template.** LogicBench, [ACL 2024](https://arxiv.org/abs/2404.15522) [PART], uses one inference rule per item.\n  - GPT-4 on first-order modus ponens: 94.7% on \"No\" items, 100% on \"Yes\". Across propositional rules, GPT-4 averaged 86.4% (No) and 64.0% (Yes). For other models the \"Yes\" side is weaker, so models are better at rejecting than recognising valid entailment.\n  - Non-monotonic average for GPT-4: 74.4% (No) / 84.8% (Yes). Reasoning with several defaults fell to 77.6% / 39.3%.\n  - A single clean rule-and-case step is the easiest form. Defaults and exceptions drop sharply.\n\n## 3. LLM formalisation of defeasible steps, checked by a solver\n\n- Ishay & Lee, \"LLMs as ASP Programmers\", [Findings ACL 2026](https://aclanthology.org/2026.findings-acl.1151/) [PART]. The LLM writes Answer Set Programs, and solver feedback drives a self-correction loop (about 2 revisions on average).\n  - On MultiLogicNMR (default reasoning with exceptions; 200 problems per setting) ASP gave 92.9% skeptical / 84.3% credulous. The same pipeline with an SMT solver gave 35.8% / 35.0%, about chance (33%).\n  - On BoardgameQA (defeasible reasoning with source preferences), 82.1% for ASP against 60.1% for SMT, but no gain over the plain-LLM baseline (82.4% vs 84.7%).\n  - Weaker models gain much less: DeepSeek-V3 reached 42.9% overall.\n  - It checks whole-problem answers, not each step as a single inference. The task structure is fixed rules and puzzles.\n- LLM-ASPIC+, Fang, Li, Chen & Liao, ECAI 2025 (FAIA 250981; [DOI 10.3233/FAIA250981](https://doi.org/10.3233/FAIA250981)) [ABS]. 87.1% on BoardgameQA depth 2 and 82.6% on depth 3, plus a new dataset, MineQA. These are end-task accuracies; translation fidelity is not reported in the abstract.\n- Argumentative LLMs, Freedman et al., [AAAI 2025](https://arxiv.org/abs/2405.02079) [ABS/SEC]. The LLM generates support and attack arguments with strengths, and DF-QuAD aggregates them. Competitive on claim verification; it does not check individual steps.\n- LLM→FOL→solver pipelines (LINC, covered in your notes) in Sadeddine & Suchanek: on FOLIO and EntailmentBank, 24–46% of steps failed to translate at all (their \"error rate\" column). Where they did translate, precision was 0.76–0.96 and recall 0.26–0.97.\n- I found no 2023–26 paper measuring LLM → ASPIC+ translation fidelity against hand formalisations. CAF-Gen (Carneades, [arXiv 2606.06646](https://arxiv.org/abs/2606.06646), preprint [ABS]) is the nearest, with no verified numbers.\n- LLM behaviour on defeasible rules themselves: DEFREASING (Allaway & McKeown, [NAACL 2025](https://aclanthology.org/2025.naacl-long.529)) [PART] has ~95k questions on generic rules with exceptions. 12 instruction-tuned LLMs reached at best about 0.64 F1 overall, and no model did well on every pattern type.\n\n## 4. Small trained step verifiers outside math\n\n- **NLI cross-encoders as validity checkers.** From Sadeddine & Suchanek (above), a fine-tuned DeBERTa NLI model applied to premises → conclusion:\n  - EntailmentBank with negation errors: precision 0.77, recall 0.87 (F0.5 0.78), the best of all methods there.\n  - FOLIO: precision 0.72–0.77, recall 0.59–0.69 (F0.5 0.70–0.73).\n  - ProofWriter with removed premises: precision 0.65, recall 0.26.\n  - Llama-3-8B prompted as NLI gave higher recall (0.88–0.99) and lower precision (0.51–0.78) on real data.\n- **MiniCheck** (Tang, Laban & Durrett, [EMNLP 2024](https://arxiv.org/abs/2404.10774)) [PART]. A 770M model checks whether a claim is supported by a document. LLM-AggreFact balanced accuracy 74.7 average against GPT-4's 75.3, at about 400× lower cost (their table). It tests grounding in a document, not inferential validity, so it is only a partial fit.\n- **Trained PRMs.** ProcessBench (Zheng et al., [ACL 2025](https://arxiv.org/abs/2412.06559)) [PART]: 3,400 math and olympiad solutions. Average F1 of open PRMs: Math-Shepherd-7B 31.5, Skywork-PRM-7B 42.1, the authors' PRM800K-trained 7B 56.5. Prompted general models: Llama-3.1-8B 5.1, Qwen2.5-7B 32.6, Qwen2.5-72B 61.2, GPT-4o 61.9, o1-mini 87.9. Small models, even prompted, are close to useless as step critics, and PRMs fall off outside GSM8K/MATH.\n- ThinkPRM ([arXiv 2504.16828](https://arxiv.org/abs/2504.16828)) [ABS]. A 1.5B generative verifier beats discriminative PRMs by 8% (GPQA-Diamond subset) and 4.5% (LiveCodeBench subset) out of domain. It writes a verification chain of thought, so it is not cheap per step.\n- VersaPRM ([ICML 2025](https://arxiv.org/abs/2502.06737)) [ABS/SEC]: math PRMs transfer poorly to other domains; a multi-domain PRM gained 7.9% over majority vote on MMLU-Pro Law against 1.3% for a math PRM.\n- Entailer (Tafjord et al., [EMNLP 2022](https://arxiv.org/abs/2210.12217)) [PART]. A trained verifier scores each entailment. Failure analysis of proofs: reasoning errors were 47% of failures, and of those, near-tautologies where the step restates its own conclusion were 20%, plain invalid steps 10%, wrong abductions 9%. Over 70% of its proofs were judged \"clearly following\" by users. It cost about 360 s per question at depth 3.\n\n## 5. Structured or constrained output\n\n- Tam et al., \"Let Me Speak Freely?\", [EMNLP 2024 Industry](https://arxiv.org/abs/2408.02442) [PART].\n  - Stricter formats lower reasoning accuracy, and constrained JSON-mode was worst. With gpt-4o-mini (v3 table): GSM8K natural language 94.57, format-restricting instructions 87.17, JSON-mode 86.95, JSON-schema 91.71; Last Letter 83.11 / 84.73 / 76.00 / 86.07; Shuffled Objects 82.85 / 81.46 / 76.43 / 81.77.\n  - JSON-mode sometimes put the answer key before the reasoning key, which hurt reasoning, so key order matters.\n  - Classification tasks improved with strict formats.\n- A practitioner replication on Gemini 1.5 Flash (Castillo, Dec 2024, [blog](https://dylancastillo.co/posts/gemini-structured-outputs.html)) [SEC] attributes much of the old drop to Gemini sorting schema keys alphabetically before generation, so an \"answer\" key came before \"reasoning\". Reported JSON-schema accuracy: GSM8K about 93%, against 47.78% quoted from Tam et al. Whether your Gemini SDK honours key order should be tested, since the project orders reasoning before verdict.\n- FineLogic ([arXiv 2506.04810](https://arxiv.org/abs/2506.04810), Findings EMNLP 2025) [PART] is the only measured effect of structure on atomicity.\n  - Setup: sample-level \"all steps atomic\" on formal-logic datasets (FLD, FOLIO, ProntoQA), where \"atomic\" means exactly one deduction rule. A GPT-4.1-mini judge was over 98% accurate on 200 manually checked steps.\n  - Results (All-Atomic, Llama-3.1-8B): few-shot 1.6; SFT on natural language 13.0; SFT on symbolic structured steps 24.7. Qwen-2.5-7B: few-shot 2.6; SFT natural language 8.5; symbolic structured 19.8. Few-shot GPT-4o 4.4 and DeepSeek-R1 5.7.\n  - Natural-language steps \"fuse multiple reasoning rules\" into one.\n- PCRLLM ([arXiv 2511.08392](https://arxiv.org/abs/2511.08392), preprint) [PART]. Outputs are JSON with Premise 1, Premise 2 and Result, each a Non-Axiomatic Logic sentence with frequency and confidence values and a named rule. The rule is checked deterministically. Experiments used 0.5–3B models and 100 training instances; many settings failed to exceed a 0.2 valid-answer ratio under loose thresholds. This is a design example, not a strong result, but non-axiomatic logic has explicit rules for uncertain deduction, induction and abduction.\n\n## 6. Inferential distance and step size\n\n- C2-Faith, Mittal & Arike, [arXiv 2603.05167](https://arxiv.org/abs/2603.05167) (preprint, 2026) [PART]. It is built from PRM800K with controlled step deletions.\n  - All three judges (GPT-4.1, DeepSeek-V3.1, o4-mini) \"systematically overestimate\" completeness. At 70% of middle steps deleted, mean coverage scores were still about 3 on a 1–5 scale, and 57–75% of scores stayed ≥3.\n  - Spearman correlation of judge score with true coverage: 0.13–0.33 at best.\n  - Localising a planted bad step: exact match 55.8–68.0%, while detection was 88–94%. About 40–55% of untouched chains were flagged as erroneous.\n- CoT-Bridge (\"Mind the Gap\", Xu et al., [NeurIPS 2025](https://arxiv.org/abs/2505.14684)) [PART]. Leaps are created by deleting steps from math solutions.\n  - A trained detector located the leaps at 78% precision / 78% recall.\n  - Zero-shot Qwen2.5-72B scored about 34% / 34%; the 7B about 14% / 12% (columns slightly garbled in the extracted table).\n  - Training on bridged data gave up to +5.87% on NuminaMath. Training on data with severe leaps cost up to 27.83% accuracy.\n- FineLogic again: relevance and atomicity are separable per-step properties that an LLM judge can score when the rule set is known.\n- Generic \"more steps help\" or \"learn to skip steps\" results (e.g. Jin et al. 2024 on step length, Liu et al. on skip training) concern chain length, not whether one step is one inference. I read only abstracts or secondary summaries and did not use their numbers.\n\n## Summary table\n\nMy ratings of cost and stringency are estimates and are not measured. \"Per step\" means calls or passes for one step.\n\n| Method | Cost per step (my estimate) | Stringency (my estimate) | Measured result | Source |\n|---|---|---|---|---|\n| Prompt only, \"use one rule\" | 0 extra | Low | All-atomic chains 1.6–5.7% (few-shot GPT-4o, R1, Llama-8B), formal logic | FineLogic, arXiv 2506.04810 |\n| Fine-tune on structured symbolic steps | 0 at run time; training cost | Medium | All-atomic 19.8–24.7% vs 8.5–13.0% for natural-language SFT | FineLogic |\n| Named-scheme template with slots, parsed by code | 1 parse | Form only | Template-generated text classifies 0.94–1.00 F1; natural text 0.5–0.65 (7 schemes), 0.29 (24 schemes, dialogue) | Bezou-Vrakatseli 2025; Ruiz-Dolz 2025 |\n| Scheme templates for question generation | 1–2 extra calls | Form only | Useful 72.2% vs 72.0% baseline, Unhelpful 8.8% vs 13.8% | DayDreamer 2025 |\n| Pattern/parse-based deductive checker | 1 parse + solver, no LLM | High precision | FOLIO precision 0.94–0.97, recall 0.16–0.26 | Sadeddine & Suchanek 2025 |\n| Same plus small NLI model on literals | ~1 small-model pass per literal pair | Medium–high | FOLIO precision 0.74–0.86, recall 0.52–0.69, F0.5 0.73–0.77 | Sadeddine & Suchanek |\n| Fine-tuned DeBERTa NLI on premises → conclusion | 1 small-model pass | Medium | FOLIO precision 0.72–0.77, recall 0.59–0.69; EntailmentBank+neg precision 0.77, recall 0.87 | Sadeddine & Suchanek |\n| LLM → FOL → solver | 1 LLM call + solver | High when it parses | Precision 0.76–0.96; 24–46% of steps untranslatable on real data | Sadeddine & Suchanek |\n| LLM → ASP with solver feedback | ~2 revisions | Whole-answer only | 92.9%/84.3% vs 35.8%/35.0% for SMT on default reasoning; no gain on BoardgameQA | Ishay & Lee 2026 |\n| Monotonicity engine (NeuralLog) | parse + alignment | Lexical-substitution entailment only | SICK 90.3%, MED 93.4% | Chen et al. 2021 |\n| LLM judge: \"is anything missing?\" | 1 call | Weak | ρ 0.13–0.33 with true coverage; over-credits | C2-Faith 2026 |\n| LLM zero-shot leap detection | 1 call | Weak | ~34% precision/recall (72B), synthetic deletions | CoT-Bridge 2025 |\n| Constrained JSON for reasoning | 0 | Format only | −7 to −8 pts on GSM8K for JSON-mode (gpt-4o-mini); key order matters | Tam et al. 2024; Castillo 2024 |\n| Small prompted or math-trained PRMs | 1 call or pass | Weak outside math | ProcessBench F1 5.1–56.5 (≤8B), math only | ProcessBench 2025 |\n| Small support-checker | 1 small-model pass | Grounding only | BAcc 74.7 vs GPT-4 75.3, 400× cheaper | MiniCheck 2024 |\n\n## What I looked for and did not find\n\n- A measured comparison of enforcing \"one rule applied once to one case\" for **defeasible natural-language** steps. FineLogic and VANESSA test formal deduction; DEFREASING tests belief updating, not step form.\n- Automatic checking of **scheme slot filling against meaning** (for example, that the \"case\" slot really instantiates the \"rule\"'s antecedent). Only conformity and classification are measured.\n- Natural-logic or monotonicity checkers applied to **LLM-generated steps** in 2023–26. Only 2020–21 systems and a 2023 LLM probe turned up.\n- LLM-to-ASPIC+ translation fidelity (precision against hand formalisation) from the Modgil/Prakken groups or others, 2023–26.\n- Small (<1B) step verifiers trained on **non-math natural-language reasoning** with per-step precision/recall. Nearest: NLI cross-encoders in Sadeddine & Suchanek, and MiniCheck for grounding.\n- A direct measure of constrained decoding's effect on **leap rate or step validity**. Only answer accuracy and the FineLogic atomicity effect of structured SFT.\n- Papers I could not read in full or check numbers for: Argumentative LLMs (AAAI 2025), LLM-ASPIC+ (full text), ThinkPRM and VersaPRM beyond abstracts, CAF-Gen results, the 2026 argument-mining preprint (arXiv 2605.13793), and the CQs-Gen benchmark paper (Calvo Figueras & Agerri, Findings EMNLP 2025, arXiv 2505.11341), whose per-model numbers I did not retrieve.\n\n## Reading for your standard (my inference, not a source claim)\n\n- Combining a structured per-step layout (premises cited, one named rule, one conclusion) with a fine-tuned NLI model to screen premises → conclusion is the cheapest measured stack.\n- Expect precision around 0.75–0.87 with recall around 0.5–0.9 as a screen, not a gate.\n- An LLM \"completeness\" judge should not be the main completeness check.\n- Test your Gemini SDK's key order before relying on reasoning-first schemas.\n"}