{
  "AJ-01": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "Late-2024 snapshot (o1-preview vs GPT-4o, PandaLM). Direction (reasoning models dominate trained judges) has held through 2026 follow-ons; the specific accuracies are two generations stale. Re-rank at the deployment-tier bakeoff."
  },
  "AJ-02": {
    "grade": "A",
    "basis": "validation-design result",
    "measured_on": "structural",
    "note": "A result about how gold labels are elicited and aggregated, not about model skill. Applies to validating Fable/GPT-5.5-class judges exactly as it did to GPT-4o-class."
  },
  "AJ-03": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "Includes April-2026 frontier models; the kappa-deflation statistics are properties of agreement measurement and transfer; per-judge rankings do not. Deployment tier and human-writeup QA untested."
  },
  "AJ-04": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "March-2026 on GPT-4o / Claude-Sonnet-4.5 / o3. One tier behind deployment; hard-subset ~55% is a prior, not a ceiling, for Fable/GPT-5.5-class. The +7–12pp per-criterion-reasoning effect is the durable part."
  },
  "AJ-05": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "June-2026, 41 judges. Subjective-vs-verifiable axis geometry is the mechanism; Indic community-health scope noted in corpus. Consensus-is-not-validity conclusion is method-level and durable."
  },
  "AJ-06": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "Measured on mini-tier judges (GPT-4o-mini / GPT-4.1-mini). Flip rates for deployment-tier reasoners are unknown and plausibly far lower; k-sampling remains cheap hygiene either way."
  },
  "AJ-07": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "GPT-4o-era kappa 0.28±0.32. Use as the canonical warning against percent-agreement narratives, not as an estimate of current judge–human agreement."
  },
  "AJ-08": {
    "grade": "B",
    "basis": "current-gen mechanism",
    "measured_on": "model-outputs",
    "note": "ICLR 2026. Lineage-based contamination is structural to training pipelines and likely persists at the deployment tier; exact PLS magnitudes are tier-bound. Cross-family routing stays the defense."
  },
  "AJ-09": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "Mid-2025 trained-judge landscape (J1, CompassJudger-2, Skywork-V2). This market moves quarterly; treat as a map of the design space, not a current ranking."
  },
  "AJ-10": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "June-2025; top models <40% on Precise Instruction Following predates GPT-5-class. The warning (instruction-compliance checking is not solved) held then; current magnitude unknown — exactly why it needs its own gold set."
  },
  "AJ-11": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "2024 study (IUI 2025). The 64–68% expert-agreement figures predate reasoning-model maturity. Mechanism (expertise-heavy axes are the weak lane) is plausible but the deployment-tier boundary is E1/E3's to measure — in either direction."
  },
  "AJ-12": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "June-2026 benchmark. Long-form instability at the current tier; deployment-tier behavior untested."
  },
  "AQ-01": {
    "grade": "A",
    "basis": "behavioral floor",
    "measured_on": "adjacent-domain",
    "note": "Crowdworker LLM-use base rate (2023, replicated 2025). A floor that rises as tools improve; the decision input (assume pervasive assistance) only strengthens."
  },
  "AQ-02": {
    "grade": "A",
    "basis": "adversarial floor",
    "measured_on": "adjacent-domain",
    "note": "Survey-agent result. Attack capability is a lower bound that strengthens with every model generation; the economic asymmetry ($0.05 vs $1.50) widens. Content detection stays beaten."
  },
  "AQ-03": {
    "grade": "A",
    "basis": "validation-design result",
    "measured_on": "structural",
    "note": "How to validate judges under human-label uncertainty — statistics of the gold set, not model capability."
  },
  "AQ-04": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "Attribution of annotator disagreement to underspecification/style. Human raters have not changed."
  },
  "AQ-05": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "2022 NLI disagreement taxonomy. Human disagreement structure, not model capability."
  },
  "AQ-06": {
    "grade": "A",
    "basis": "statistical fact",
    "measured_on": "structural",
    "note": "Kappa paradox under extreme base rates. Mathematics."
  },
  "AQ-07": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "Dawid-Skene/MACE-class aggregation. Method, not capability."
  },
  "AQ-08": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "Worker-level contamination auditing without per-item detection. Method survives model improvement on both sides."
  },
  "AQ-09": {
    "grade": "A",
    "basis": "field practice",
    "measured_on": "structural",
    "note": "Disagreement-as-signal institutionalized (LeWiDi-2025). Practice fact."
  },
  "AQ-10": {
    "grade": "A",
    "basis": "dataset audit",
    "measured_on": "structural",
    "note": "≥3.3% gold-label error floor across canonical test sets. Property of datasets and labeling processes, not of judges."
  },
  "AQ-11": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "Anchoring on shown suggestions is human behavior; better suggestions anchor harder. Blind-first design stays mandatory."
  },
  "AQ-12": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "Jan-2026: RLVR training degrades disagreement modeling. Mechanism relevant to judge-model selection at the deployment tier; magnitude tier-bound."
  },
  "CM-01": {
    "grade": "C",
    "basis": "superseded-gen capability + human floor",
    "measured_on": "human-work",
    "note": "CriticGPT (GPT-4-era, 2024). Critic-assisted-beats-either-alone is a floor that should strengthen with critic quality, but the 63% preference number is two generations stale and unreplicated in 24+ months."
  },
  "CM-02": {
    "grade": "C",
    "basis": "superseded-gen capability + human floor",
    "measured_on": "human-work",
    "note": "The 24%-vs-6% flawless-slice lift is GPT-4-era. Direction (audit confident passes with critic assistance) is the durable design input; the 4× magnitude is not a current estimate."
  },
  "CM-03": {
    "grade": "B",
    "basis": "current-gen + human science",
    "measured_on": "adjacent-domain",
    "note": "2025 DeepMind amplified-oversight. The format effect (verdict-visible → over-reliance; evidence-only safe) is human behavior and durable; the 91.3/87.7/75.1 magnitudes are tier- and task-bound (fact verification, not annotation QA)."
  },
  "CM-04": {
    "grade": "A",
    "basis": "human science (direction)",
    "measured_on": "adjacent-domain",
    "note": "Meta-analytic human-AI complementarity result. As deployment-tier models widen the human-AI gap, the warning against naive verify-the-AI designs strengthens."
  },
  "CM-05": {
    "grade": "B",
    "basis": "human response to AI feedback",
    "measured_on": "human-work",
    "note": "2025 RCT on >20k real peer reviews — one of the few human-work measurements in the corpus. Revision behavior is human; feedback quality rises with model tier, so 27% is closer to a floor than a ceiling. Domain transfer (peer review → paid annotation) untested (E6)."
  },
  "CM-06": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "AbsenceBench on 2025 flash-tier (Gemini-2.5-flash). Omission-blindness is attention-mechanics and plausibly persists, but deployment-tier recall is unmeasured — E2 measures it before any absence-claim verdict is trusted."
  },
  "CM-07": {
    "grade": "C",
    "basis": "superseded-gen method demo",
    "measured_on": "model-outputs",
    "note": "MetaCritique (GPT-4, 2024). The critique-decomposition METHOD is reusable; its reported reliability numbers are stale."
  },
  "CM-08": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "Right-verdict/poor-critique dissociation measured on 2024–25 models; re-check found headline numbers span sibling benchmarks. Keep the failure mode, drop the numbers."
  },
  "CM-09": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "Same study family as AJ-03; June-2026 window."
  },
  "CM-10": {
    "grade": "A",
    "basis": "production practice",
    "measured_on": "structural",
    "note": "Toloka's deployed pass/fail/unable-to-verify QA with human escalation. Existence proof of the architecture in production — vendor-reported, flagged as such."
  },
  "CM-11": {
    "grade": "A",
    "basis": "human science (direction)",
    "measured_on": "human-work",
    "note": "2022 foundational result: model critiques help humans find flaws they'd miss. Floor-type: strengthens with critic quality."
  },
  "CM-12": {
    "grade": "A",
    "basis": "adversarial floor",
    "measured_on": "human-work",
    "note": "Deployed checklist assistant gamed in one shot (NeurIPS 2024). Attackers improve with model tier; the design lesson is permanent."
  },
  "GF-01": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "The ~77 bacc AggreFact ceiling and specialists-within-2.5-points are GPT-4o/Llama-3.1-era. Deployment-tier entailment ceilings are unknown; the two-tier screening economics need repricing at the bakeoff."
  },
  "GF-02": {
    "grade": "B",
    "basis": "current-gen architecture result",
    "measured_on": "model-outputs",
    "note": "FaithJudge (late-2025, o3-mini-era). Exemplar-anchoring-beats-zero-shot is an architecture effect expected to persist; margins are tier-bound."
  },
  "GF-03": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "MiniCheck/GPT-4o-mini-era decomposition harm. Adaptive granularity stays the design default; the 80→71 magnitude is not a current estimate."
  },
  "GF-04": {
    "grade": "B",
    "basis": "current-gen mechanism",
    "measured_on": "model-outputs",
    "note": "2025 decomposition/decontextualization tension analysis. Design-level tension, partially structural."
  },
  "GF-05": {
    "grade": "A",
    "basis": "method / typology",
    "measured_on": "structural",
    "note": "Verifiability triage before entailment (VeriScore/FactBench). A pipeline design pattern, not a capability claim."
  },
  "GF-06": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "ContextualJudgeBench ~55% on o1-era judges. Holistic-vs-claim-level gap direction is consistent across studies; magnitude stale."
  },
  "GF-07": {
    "grade": "B",
    "basis": "current-gen capability (flash-tier graders)",
    "measured_on": "model-outputs",
    "note": "FACTS (Dec 2025, gemini-2.5-flash graders). The agreement-default asymmetry (85/46) is the single most decision-relevant grounding mechanism — and it was measured on flash-tier graders. Deployment-tier asymmetry is E2's first number."
  },
  "GF-08": {
    "grade": "B",
    "basis": "current-gen practice",
    "measured_on": "model-outputs",
    "note": "FACTS v1→v2 ensemble evolution; includes GPT-5-class by v2. Cross-provider ensembling as self-preference mitigation is practice-level."
  },
  "GF-09": {
    "grade": "B",
    "basis": "current-gen method critique",
    "measured_on": "model-outputs",
    "note": "2025: span-restricted citation checking is insufficient. Method-level; motivates the dual-pass citation design."
  },
  "GF-10": {
    "grade": "B",
    "basis": "current-gen brittleness",
    "measured_on": "model-outputs",
    "note": "FactEval (2025): verifier verdicts flip under meaning-preserving perturbation. Perturbation harness stays a ship gate; deployment-tier flip rates unknown (could be far lower — measure, don't assume)."
  },
  "GF-11": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "FaithLens (ACL 2026, GPT-5.2-era comparison) — among the most current capability anchors in the corpus. Still model-output verification, not human-writeup QA."
  },
  "GF-12": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "Lynx-vs-GPT-4-era comparison (2024). Fine-tuned-specialist economics need current repricing."
  },
  "HS-01": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "PPI/PPI++ unbiasedness and judge-robust intervals are mathematics; they get MORE useful as judges improve. (Empirical demos in the paper are GPT-4-era; the method is the decision input. Title changed on arXiv v3 — noted in re-check.)"
  },
  "HS-02": {
    "grade": "A",
    "basis": "statistical method (numbers dated)",
    "measured_on": "model-outputs",
    "note": "Trust-or-Escalate's guarantee construction is method-level and durable. Its reported coverage numbers are GPT-4-era; recalibrate on deployment-tier judges — the framework exists precisely to do that."
  },
  "HS-03": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "Closed-form gold-set sizing. Mathematics."
  },
  "HS-04": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "Confidence-driven human-annotation allocation. Method."
  },
  "HS-05": {
    "grade": "A",
    "basis": "tooling fact",
    "measured_on": "structural",
    "note": "GLIDE library unifying PPI-class estimators (May 2026). Tooling availability, current."
  },
  "HS-06": {
    "grade": "A",
    "basis": "statistical theory",
    "measured_on": "structural",
    "note": "EIF/PPI++ efficiency results. Mathematics (interval-width empirics corpus-flagged as not independently recomputed)."
  },
  "HS-07": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "'Noisy but Valid' certification: judge-assisted testing provably more powerful under conditions. Mathematics."
  },
  "HS-08": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "Sensitivity/specificity-corrected pass-rate reporting. Mathematics."
  },
  "HS-09": {
    "grade": "B",
    "basis": "current-gen method + measurement",
    "measured_on": "model-outputs",
    "note": "Conformal per-item uncertainty for judge ratings (2025–26). Method durable; measured widths tier-bound."
  },
  "HS-10": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "Anchor-set drift attribution with anytime-valid statistics. Method."
  },
  "HS-11": {
    "grade": "A",
    "basis": "statistical finding",
    "measured_on": "structural",
    "note": "Sequential estimation: uncertainty-based querying bought little vs uniform. Regime-dependent statistics; the design response (A/B routing against uniform) is already encoded."
  },
  "HS-12": {
    "grade": "A",
    "basis": "tooling fact",
    "measured_on": "structural",
    "note": "Cleanlab TLM: trust-scored judgments with calibration tooling. Vendor tooling fact, current."
  },
  "RR-01": {
    "grade": "B",
    "basis": "current-gen practice + capability",
    "measured_on": "model-outputs",
    "note": "HealthBench grader meta-evaluation (2025, GPT-4.1/o3-era graders). The per-criterion meta-evaluation PATTERN is the durable input; MF1 0.709 is tier-bound."
  },
  "RR-02": {
    "grade": "A",
    "basis": "production practice",
    "measured_on": "structural",
    "note": "48,562 physician-written per-item criteria: what expert-grounded rubric authoring costs at production scale. Practice fact."
  },
  "RR-03": {
    "grade": "B",
    "basis": "current-gen result",
    "measured_on": "model-outputs",
    "note": "Rubrics-as-Rewards (2025): reference-grounded synthesis matches human authorship except pitfall criteria. The pitfall exception drives the compilation design; re-test as generators improve."
  },
  "RR-04": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "RubricBench (Mar 2026, DeepSeek-v3.2/Gemini-3-era): 26–27pp self-vs-expert rubric gap that test-time compute doesn't close. Current-window; deployment-tier gap unknown."
  },
  "RR-05": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "EvalGen criteria drift: human graders' criteria change as they grade. Human behavior; durable."
  },
  "RR-06": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "RIFT (Apr 2026, GPT-5.2/Gemini-3-era linting): Subjective detectable (F1 .925), Hackable not (~0). Current-window; the human-red-team conclusion for anti-gaming review stands until a later tier changes it."
  },
  "RR-07": {
    "grade": "B",
    "basis": "current-gen result",
    "measured_on": "model-outputs",
    "note": "Support Vector Rubrics (Jun 2026): contrastive mining closes the rubric gap to ~0.3pp given preference pairs. Current-window; depends on adjudication data the pilot creates."
  },
  "RR-08": {
    "grade": "B",
    "basis": "current-gen practice + capability",
    "measured_on": "model-outputs",
    "note": "PaperBench (2025, o1/o3-era judge F1 0.83 at ~1/100 cost). Decomposition-at-scale pattern durable; judge F1 and cost tier-bound."
  },
  "RR-09": {
    "grade": "A",
    "basis": "production practice",
    "measured_on": "structural",
    "note": "OpenAI grader doctrine: every model grader validated on its own expert-graded eval. Practice prescription, current."
  },
  "RR-10": {
    "grade": "B",
    "basis": "current-gen literature position",
    "measured_on": "model-outputs",
    "note": "2026 RL-rubrics cluster: static rubrics are exploitable specifications under optimization. Conceptual; strengthens with optimizer capability."
  },
  "RR-11": {
    "grade": "A",
    "basis": "tooling fact",
    "measured_on": "structural",
    "note": "Autorubric open-source framework. Tooling availability."
  },
  "RR-12": {
    "grade": "A",
    "basis": "tooling fact",
    "measured_on": "structural",
    "note": "LangSmith Align Evals: judge-calibration workflow productized. Tooling fact."
  },
  "IP-01": {
    "grade": "A",
    "basis": "production practice",
    "measured_on": "structural",
    "note": "An annotation platform and its labeler network QA stack: double-blind labeling, α≥0.80 bar, judge flags-not-sanctions. Vendor-reported practice; existence proof for A4's flag-only posture."
  },
  "IP-02": {
    "grade": "B",
    "basis": "current-gen vendor eval",
    "measured_on": "human-work",
    "note": "The banking meta-eval (May 2026): 3,204 practicing-banker labels — a rare human-work meta-eval. Vendor research; benchmark release pending per corpus."
  },
  "IP-03": {
    "grade": "A",
    "basis": "market practice",
    "measured_on": "structural",
    "note": "2026 eval-tooling convergence on human-calibrated judges. Market fact."
  },
  "IP-04": {
    "grade": "A",
    "basis": "documented postmortem",
    "measured_on": "adjacent-domain",
    "note": "Scale/Bulba spam postmortem (journalism-of-documents, disputed by Scale, pre-LLM-judge era — corpus flags all three). Economic lesson durable; treat operational details as reported, not established."
  },
  "IP-05": {
    "grade": "A",
    "basis": "market fact",
    "measured_on": "structural",
    "note": "An AI data vendor acquired a data-quality startup for no-second-human flagging (Jan 2026). Market signal, current."
  },
  "IP-06": {
    "grade": "A",
    "basis": "market practice",
    "measured_on": "structural",
    "note": "Surge QA-as-dashboards practice. Vendor-reported."
  },
  "IP-07": {
    "grade": "A",
    "basis": "market practice",
    "measured_on": "structural",
    "note": "micro1 real-time QA verification layers. Vendor-reported."
  },
  "IP-08": {
    "grade": "B",
    "basis": "market pricing",
    "measured_on": "structural",
    "note": "Braintrust cost arithmetic (judge $5–15/10k vs expert $800–1,800/500). Pricing is current but moves with model economics — reprice at bakeoff; the routing logic it supports is durable."
  },
  "IP-09": {
    "grade": "C",
    "basis": "superseded-gen capability",
    "measured_on": "model-outputs",
    "note": "Lynx vs GPT-4o-era comparisons. Specialist-vs-frontier margins need current repricing."
  },
  "IP-10": {
    "grade": "A",
    "basis": "tooling practice",
    "measured_on": "structural",
    "note": "DeepEval G-Eval→DAG progression. Tooling doctrine."
  },
  "IP-11": {
    "grade": "A",
    "basis": "structural/incentive analysis",
    "measured_on": "human-work",
    "note": "Scale-vendor reviewer layer unreliability for incentive reasons. Structural; the AutoQA must not inherit it."
  },
  "IP-12": {
    "grade": "A",
    "basis": "practice observation",
    "measured_on": "structural",
    "note": "Vendors attribute low precision to instruction/rubric defects. Consistent with AQ-04; practice-level."
  },
  "CT-01": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "Contrarian read of the June-2026 meta-evaluation (same source as AJ-03; not independent)."
  },
  "CT-02": {
    "grade": "A",
    "basis": "human science (direction)",
    "measured_on": "adjacent-domain",
    "note": "g=−0.23 across 370 effect sizes, mostly older-gen AI and adjacent tasks (incl. hiring deference). The mechanism — humans defer when AI outperforms them — is human behavior and strengthens as the capability gap widens. Magnitudes are not current estimates."
  },
  "CT-03": {
    "grade": "A",
    "basis": "theorem",
    "measured_on": "structural",
    "note": "The 2× ceiling is mathematics. Note the flip side: if deployment-tier judges materially EXCEED attempter accuracy on a lane, the constraint relaxes there — a bakeoff-measurable condition, not an assumption."
  },
  "CT-04": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "Grader criteria instability. Human behavior."
  },
  "CT-05": {
    "grade": "B",
    "basis": "current-gen mechanism",
    "measured_on": "model-outputs",
    "note": "Mistake-similarity bias + industry-wide error-correlation trend (2025). Mechanism expected to persist; magnitudes tier-bound. Cross-family routing + permanent audit remain the response."
  },
  "CT-06": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "2025: judges don't follow their own rubrics (schema explains under half of verdict variance). Motivates the schematic-adherence ship gate; deployment-tier adherence is measurable there."
  },
  "CT-07": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "2025 intra-rater reliability findings. k-sampling + entropy routing remain hygiene."
  },
  "CT-08": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "2025: judges weakest on questions they cannot themselves answer. Same mechanism family as re-solving (F26-01); E3 probes it on our workload."
  },
  "CT-09": {
    "grade": "A",
    "basis": "adversarial floor",
    "measured_on": "model-outputs",
    "note": "Universal adversarial phrases transfer to unseen judges (2024). Attack floors strengthen with attacker capability; comparative/anchored scoring and perturbation probes stay day-one requirements."
  },
  "CT-10": {
    "grade": "A",
    "basis": "economic history",
    "measured_on": "adjacent-domain",
    "note": "Spam economics beat industrial QA once already. Economic asymmetry is durable."
  },
  "CT-11": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "human-work",
    "note": "NUTMEG: separating systematic subpopulation disagreement from noise. Method for human-label modeling."
  },
  "CT-12": {
    "grade": "A",
    "basis": "human science (direction)",
    "measured_on": "human-work",
    "note": "AI assistance homogenizes population output while raising individual quality (2025). Monoculture pressure rises with assistance quality; the diversity drift metric stays."
  },
  "F26-01": {
    "grade": "B",
    "basis": "current-gen mechanism (single study, under review)",
    "measured_on": "model-outputs",
    "note": "Re-solving collapse (68→9%) is the corpus's own flagged single-source provisional claim; C2 holds it at 'replicate before hard-gating'. Current-window measurement."
  },
  "F26-02": {
    "grade": "B",
    "basis": "current-gen vendor result",
    "measured_on": "human-work",
    "note": "The vendor's agent-judge (May 2026): architecture>model on expert-graded human work. Vendor research; among the most decision-relevant current results — and exactly the kind the bakeoff must reproduce in-house."
  },
  "F26-03": {
    "grade": "B",
    "basis": "current-gen mechanism",
    "measured_on": "model-outputs",
    "note": "Evaluation faking under consequence framing (2026). Stakes-sterile prompting is cheap insurance regardless of tier; behavioral bias measurement stays mandatory."
  },
  "F26-04": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "HyPAC dual-threshold routing with error budgets. Method."
  },
  "F26-05": {
    "grade": "B",
    "basis": "current-gen method + measurement",
    "measured_on": "model-outputs",
    "note": "Per-criterion conformal reliability ceilings (2026). Method durable; measured ceilings tier-bound — recompute per project."
  },
  "F26-06": {
    "grade": "B",
    "basis": "current-gen capability",
    "measured_on": "model-outputs",
    "note": "RubricEval 2026 restatement (same underlying source family as AJ-04)."
  },
  "F26-07": {
    "grade": "B",
    "basis": "current-gen result",
    "measured_on": "model-outputs",
    "note": "LLM-as-a-Verifier (Jul 2026): verification as a scaling axis; decomposition monotonically improves it. The most current capability signal in the corpus; still model-output verification."
  },
  "F26-08": {
    "grade": "B",
    "basis": "current-gen practice",
    "measured_on": "human-work",
    "note": "CoVal (Jan 2026): rubric-vs-collective-eval concordance on crowd judgments. Re-check flagged in-sample vs out-of-sample number nuance — see ledger entry."
  },
  "F26-09": {
    "grade": "A",
    "basis": "statistical method",
    "measured_on": "structural",
    "note": "Corrected pass-rate certification under imperfect judges. Mathematics (re-check corrected a per-class-vs-total sizing detail — see ledger entry)."
  },
  "F26-10": {
    "grade": "A",
    "basis": "dataset availability",
    "measured_on": "structural",
    "note": "FindTheFlaws: expert-annotated flawed-solution datasets exist for seeded-recall measurement (E2's raw material)."
  },
  "F26-11": {
    "grade": "A",
    "basis": "production practice",
    "measured_on": "structural",
    "note": "2026 norm: closed calibration loop on sampled human corrections. Practice pattern."
  },
  "F26-12": {
    "grade": "A",
    "basis": "tooling fact",
    "measured_on": "structural",
    "note": "Autorubric framework packaging 2026 judge-reliability techniques."
  },
  "CG-01": {
    "grade": "A",
    "basis": "psychometric method",
    "measured_on": "human-work",
    "note": "MFRM adjustment changes rankings built on human ratings; estimability needs 3–5 ratings/output linkage. Method + human-rater behavior; applies unchanged."
  },
  "CG-02": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "Weigle 1998: training buys self-consistency, not severity equality. Canonical, replicated; humans unchanged."
  },
  "CG-03": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "FOR training d≈0.50 ceiling on the best rater-training intervention. Durable."
  },
  "CG-04": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "G-theory: interactions dominate rater main effects. Durable; the strongest independent argument for per-claim verification."
  },
  "CG-05": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "Within-session rater severity drift. Durable; rolling-window reviewer effects stay."
  },
  "CG-06": {
    "grade": "B",
    "basis": "current-gen method application",
    "measured_on": "model-outputs",
    "note": "MFRM applied to 2025 LLM judges. The symmetric-measurement idea is method-level; the specific judge rankings are tier-bound."
  },
  "CG-07": {
    "grade": "A",
    "basis": "law",
    "measured_on": "structural",
    "note": "Platform Work Directive Art 10(5). Binding for EU workers from transposition (2026-12-02); capability-independent by definition."
  },
  "CG-08": {
    "grade": "A",
    "basis": "law",
    "measured_on": "structural",
    "note": "PWD Art 11 explanation/review rights."
  },
  "CG-09": {
    "grade": "A",
    "basis": "law",
    "measured_on": "structural",
    "note": "PWD scope covers online-performed platform work."
  },
  "CG-10": {
    "grade": "A",
    "basis": "law",
    "measured_on": "structural",
    "note": "PWD Art 7 telemetry prohibitions + DPIA."
  },
  "CG-11": {
    "grade": "A",
    "basis": "case law",
    "measured_on": "structural",
    "note": "SCHUFA doctrine; active enforcement. (Re-check: doctrine confirmed; the phrase 'determining role' traces to the CJEU judgment rather than the fetched commentary — see ledger entry.)"
  },
  "CG-12": {
    "grade": "A",
    "basis": "standard",
    "measured_on": "structural",
    "note": "ISO/IEC 5259 published series. Standards fact."
  },
  "CG-13": {
    "grade": "A",
    "basis": "production practice",
    "measured_on": "adjacent-domain",
    "note": "Facebook/Cognizant audit architecture. Decade-scale deployed practice for humans-reviewing-human-judgment."
  },
  "CG-14": {
    "grade": "A",
    "basis": "industry codification",
    "measured_on": "adjacent-domain",
    "note": "TSPA QA curriculum: seeding indistinguishability, appeals bias, wrong-selection errors."
  },
  "CG-15": {
    "grade": "B",
    "basis": "current-gen practice",
    "measured_on": "human-work",
    "note": "Pinterest decision-quality framework (Feb 2026, GPT-5-class agents in scope): reliability-vs-correctness separation, dual-labeling on policy change. Most current deployed analog."
  },
  "CG-16": {
    "grade": "A",
    "basis": "human incentive result",
    "measured_on": "human-work",
    "note": "Consensus-rewarded auditing induces conformity (Community Notes). Incentive design; durable."
  },
  "CG-17": {
    "grade": "A",
    "basis": "method",
    "measured_on": "human-work",
    "note": "Per-rubric-question agreement decomposition (Google). Method."
  },
  "CG-18": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "adjacent-domain",
    "note": "Throughput pressure degrades moderator judgment (2025 observational). Human factors; durable."
  },
  "CG-19": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "Kluger & DeNisi FIT: a third of feedback interventions backfire; task-referenced helps. Foundational; durable."
  },
  "CG-20": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "adjacent-domain",
    "note": "Cochrane 2025 audit-and-feedback update: modest median effects, concentrated in low performers. Durable."
  },
  "CG-21": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "human-work",
    "note": "Dow CSCW 2012: task-specific feedback improves paid microtask quality; self-ratings inflate. Closest crowdwork RCT; durable."
  },
  "CG-22": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "adjacent-domain",
    "note": "Information content dominates feedback efficacy (d≈0.99 high-information vs 0.46 bare verdicts). Durable."
  },
  "CG-23": {
    "grade": "A",
    "basis": "human science",
    "measured_on": "adjacent-domain",
    "note": "25-year workplace-feedback retrospective: field immature; effects heterogeneous. Calibrates expectations."
  },
  "CG-24": {
    "grade": "A",
    "basis": "human incentive result",
    "measured_on": "adjacent-domain",
    "note": "Rank feedback under incentives induces gaming/sabotage. Benchmark against exemplar work, never ranked people."
  }
}