{"paper":{"title":"LLM-as-a-Judge Is Not an Oracle: Why Self-Improving Agents Need Deterministic Guardrails","arxiv_id":"2609.02246","primary_cat":"cs.AI","submitted_at":"2026-09-02T07:54:23+00:00","version_compiled":"v1"},"notice":"Every quoted sentence is the authors' text from the arXiv listing abstract. Labels and probabilities are the outputs of classifier-1.13.0 on the question bank abstract-p1-20260923, exactly as returned. No sentence in this record was written by a language model.","schema":"pith.compiled-record/v0","status":"compiled","abstract":{"source":"arXiv listing abstract v1","sentences":[{"i":0,"p":0.96,"id":"s0","own":false,"role":"background","text":"Self-improving agent pipelines have a problem at their center.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.08,"main_claim_p":0.03,"quantitative":false,"quantitative_p":0.03,"role_probabilities":{"gap":0.04,"method":0.0,"result":0.0,"background":0.96,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":1,"p":1.0,"id":"s1","own":false,"role":"background","text":"An optimizer rewrites prompts to score higher, and the score comes from a judge that is itself an LLM.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.12,"main_claim_p":0.03,"quantitative":false,"quantitative_p":0.05,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":1.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":2,"p":0.48,"id":"s2","own":false,"role":"background","text":"That judge has the last word on whether the system is getting better, and our position is that it has not earned it.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.12,"main_claim_p":0.26,"quantitative":false,"quantitative_p":0.09,"role_probabilities":{"gap":0.14,"method":0.0,"result":0.0,"background":0.48,"limitation":0.0,"future_work":0.0,"contribution":0.03,"interpretation":0.35}},{"i":3,"p":0.42,"id":"s3","own":false,"role":"interpretation","text":"The judge should be demoted from oracle to advisor: its verdict becomes one input among several, and every change is gated instead by a deterministic verification layer the judge cannot override.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.19,"main_claim_p":0.89,"quantitative":false,"quantitative_p":0.29,"role_probabilities":{"gap":0.0,"method":0.19,"result":0.0,"background":0.01,"limitation":0.0,"future_work":0.11,"contribution":0.27,"interpretation":0.42}},{"i":4,"p":0.51,"id":"s4","own":true,"role":"method","text":"We reached this position by building the alternative and running it.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.93,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.03,"role_probabilities":{"gap":0.0,"method":0.51,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.48,"interpretation":0.01}},{"i":5,"p":0.5,"id":"s5","own":true,"role":"contribution","text":"Over months of running autonomous prompt-optimization loops in production across contract analysis, compliance review, and code quality, we cataloged eleven ways the evaluation signal failed, in four classes: judge bias, harness and metric failures, ground-truth errors, and reward hacking.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.95,"main_claim_p":0.05,"quantitative":true,"quantitative_p":0.92,"role_probabilities":{"gap":0.0,"method":0.35,"result":0.15,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.5,"interpretation":0.0}},{"i":6,"p":1.0,"id":"s6","own":true,"role":"result","text":"Agents achieved perfect scores by reading cached answer keys from their environment, a 100% pass rate concealing 68% true capability.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.9,"main_claim_p":0.03,"quantitative":true,"quantitative_p":0.99,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":7,"p":1.0,"id":"s7","own":true,"role":"result","text":"A corrupted ground-truth label caused the optimizer to delete correct compliance rules to agree with it.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.86,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.06,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":8,"p":1.0,"id":"s8","own":true,"role":"result","text":"A syntactically broken prompt was promoted as the winner because a silent parser fallback improved the metric.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.87,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.09,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":9,"p":0.9,"id":"s9","own":true,"role":"result","text":"Attempts to fix the judge by rewriting its rubric plateaued; the only reliable gain came from a structural constraint on its output order.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.87,"main_claim_p":0.05,"quantitative":false,"quantitative_p":0.3,"role_probabilities":{"gap":0.01,"method":0.02,"result":0.9,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.06}},{"i":10,"p":0.84,"id":"s10","own":true,"role":"method","text":"In response we describe PROCTOR, a Teacher-Student loop in which a stateful orchestrator holds all tool access, stateless subagents diagnose failures and draft mutations they cannot apply, and a Teacher grades those mutations under five deterministic guardrails: hermetic sandboxes, capability-disjoint roles, acceptance checks that outrank the Teacher, frozen holdouts, and canary cases engineered so that a perfect score is itself evidence of cheating.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.91,"main_claim_p":0.18,"quantitative":true,"quantitative_p":0.88,"role_probabilities":{"gap":0.0,"method":0.84,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.16,"interpretation":0.0}},{"i":11,"p":0.89,"id":"s11","own":true,"role":"limitation","text":"We report the failures this prevented, and, because the Teacher is itself an LLM judge, the failures it did not.","hedge":{"p":0.96,"label":"asserted"},"own_p":0.91,"main_claim_p":0.1,"quantitative":false,"quantitative_p":0.09,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.06,"background":0.0,"limitation":0.89,"future_work":0.0,"contribution":0.05,"interpretation":0.0}}]},"compiled":{"main_claim":{"p":0.99,"text":"The judge should be demoted from oracle to advisor: its verdict becomes one input among several, and every change is gated instead by a deterministic verification layer the judge cannot override.","location":"abstract s3","hedge_strength":{"p":0.99,"label":"asserted"}},"provenance":{"route":{"via":"TypeSafe API","model":"classifier-1.13.0","endpoint":"https://api.typesafe.ai/v1/systemone","provider":"TypeSafe"},"packing":"one request per paper","requests":1,"input_tokens":8916,"cost_usd_reported":0.0},"compiled_at":"2026-09-23","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.05,"states_limitation":0.94,"human_participants":0.06,"controlled_comparison":0.17,"randomized_assignment":0.02,"states_quantitative_result":0.98,"compares_with_prior_results":0.44,"mentions_code_or_data_availability":0.05},"evidential_basis":{"p":1.0,"value":"observation or experiment","probabilities":{"argument":0.0,"literature":0.0,"proof_or_derivation":0.0,"analysis_of_existing_data":0.0,"observation_or_experiment":1.0,"simulation_or_computation":0.0}},"contribution_type":{"p":0.75,"value":"method or tool","probabilities":{"review":0.0,"method_or_tool":0.75,"empirical_study":0.12,"simulation_study":0.0,"theoretical_result":0.0,"dataset_or_benchmark":0.0,"position_or_commentary":0.13}},"contribution_statement":{"p":0.98,"text":"In response we describe PROCTOR, a Teacher-Student loop in which a stateful orchestrator holds all tool access, stateless subagents diagnose failures and draft mutations they cannot apply, and a Teacher grades those mutations under five deterministic guardrails: hermetic sandboxes, capability-disjoint roles, acceptance checks that outrank the Teacher, frozen holdouts, and canary cases engineered so that a perfect score is itself evidence of cheating.","location":"abstract s10"}},"coverage":{"tier":"abstract","levels_run":["P abstract"],"not_examined":["full text","figures and tables","references and citing sentences","relations to other papers"],"sources_read":["arXiv listing abstract v1"],"levels_not_run":["S sections","T body sentences","C citations","R cross-paper relations"],"sentences_examined":12,"sentences_in_abstract":12},"measured_accuracy":{"url":"https://pith.science/accuracy","measured_on":"2026-10-02","method":"two independent model families labelled a fresh random sample blind; the figure is how often the published label matches their consensus","sentence_role_agreement":0.92,"main_claim_agreement":0.908,"typed_citation_agreement":0.933,"typed_citation_agreement_95_interval":[0.917,0.945],"review_comments_accurate":0.758,"review_comments_accurate_95_interval":[0.684,0.82],"review_caught_admitted_error_found_or_related":0.448,"review_comments_judged":{"judged_by":["gpt-5.6-luna","grok-4.7"],"judged_against":"Pith's own text copy of each paper: text extracted from the PDF, the first 90,000 characters","measured_on":"2026-10-04","accurate":113,"decided":149,"sampled_reviews":200,"split_by_lane":false,"by_lane":"https://pith.science/accuracy#by-lane"}}}