v0.8.0 Benchmark Saturation Detector — anti-bullshit pack #6

New 📈 Saturation mode: pick a benchmark, get top-3 frontier scores,
spread, mean, and a verdict (saturated / near-saturated / discriminative)
plus recommended replacements. Addresses the explicit "MMLU is saturated,
what now?" pain documented in arxiv 2508.15361 and 2026 leaderboards.

Data sources:
- Live: DemandSphere AI Frontier Model Tracker (CC BY-NC 4.0) — 66 models,
weekly updates, no auth, no rate limit
- Fallback: baked snapshot data/saturation_kb.json (15 benchmarks, 5 papers
cited, last fetch 2026-05-05)

Validation (research-physics, pre-registered 2026-05-07):
- 3 clean pass (MATH/HLE/SWE-bench)
- 3 borderline (GPQA/HumanEval rule sensitivity ±1pp)
- 1 falsified (AIME 2025 saturated faster than expected — itself a useful
tool output: tells users "AIME 2025 already at 96-100%, switch to HLE")
Shipped with explicit threshold-sensitivity disclaimer.

Coverage: MMLU/MMLU-Pro/GPQA-Diamond/SWE-bench-Verified/HumanEval/
LiveCodeBench-Pro/MATH/AIME-2025/HLE/ARC-AGI-2/HellaSwag/GSM8K +
multimodal: MMMU/MMMU-Pro/VisScience.

i18n × 4 langs (EN/ES/FR/ZH) — 36 keys per lang. Help modal section,
inventory card entry, and task-tile button under "Trust a benchmark score".

Files: data/saturation_kb.json + js/saturation_detector.js (new);
index.html + js/main.js + js/i18n.js (modified).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

data/saturation_kb.json

@@ -0,0 +1,285 @@
+{
+  "version": "0.8.0",
+  "fetched_at": "2026-05-05",
+  "compiled_at": "2026-05-07",
+  "primary_source": {
+    "name": "DemandSphere AI Frontier Model Tracker",
+    "url": "https://www.demandsphere.com/research/demandsphere-radar/ai-frontier-model-tracker/",
+    "api": "https://www.demandsphere.com/research/demandsphere-radar/ai-frontier-model-tracker/api.json",
+    "license": "CC BY-NC 4.0",
+    "attribution_required": true,
+    "model_count": 66,
+    "last_updated": "2026-05-05"
+  },
+  "secondary_sources": [
+    {
+      "name": "HuggingFace Open LLM Leaderboard v3",
+      "url": "https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard",
+      "use": "Open-weight historical baseline; dataset open-llm-leaderboard/contents (4580 models, 6 benchmarks)"
+    },
+    {
+      "name": "Public benchmark consensus (2026 community)",
+      "use": "MMLU/HellaSwag/GSM8K saturation status from survey arxiv 2508.15361 + multiple leaderboards"
+    }
+  ],
+  "classification_rule": {
+    "saturated": "top-3 spread ≤ 2pp AND mean(top-3) ≥ 90%",
+    "near_saturated": "top-3 spread ≤ 5pp AND mean(top-3) ≥ 80% (and not saturated)",
+    "discriminative": "top-3 spread > 5pp OR mean(top-3) < 80%",
+    "sparse_data": "fewer than 3 frontier models report this benchmark publicly"
+  },
+  "honest_disclaimer": "Threshold-sensitive: a benchmark with spread 5.0pp classifies as near_saturated; 5.1pp classifies as discriminative. Verdicts within ±1pp of cutoffs should be read as 'borderline'. Pre-registered validation 2026-05-07 against 7 benchmarks: 3 clean pass, 3 borderline, 1 falsified (AIME 2025 saturated faster than expected). Tool is descriptive of public scores at fetch date — not predictive of future. Always cross-check before strategic decisions.",
+  "benchmarks": {
+    "MMLU": {
+      "key": "mmlu",
+      "category": "general_knowledge",
+      "description": "Massive Multitask Language Understanding — 57 subjects, 4-way multiple choice.",
+      "year_introduced": 2020,
+      "top_3": [
+        {"model": "GPT-5.5", "score": 92.4},
+        {"model": "(other frontier models no longer report — DemandSphere lists N/A)", "score": null},
+        {"model": "(historically all 88-94% range)", "score": null}
+      ],
+      "spread_pp": null,
+      "mean_top3": null,
+      "classification": "saturated",
+      "classification_basis": "consensus + DemandSphere stopped tracking N/A for many frontier models, plus survey 2508.15361 confirms 88-94% top range cluster.",
+      "recommendations": ["MMLU-Pro", "GPQA Diamond", "HLE"],
+      "note": "MMLU was useful 2020-2023. By 2024 it stopped discriminating frontier models. Many leaderboards have dropped it. Use MMLU-Pro for general 7B-70B; GPQA/HLE for frontier."
+    },
+    "MMLU-Pro": {
+      "key": "mmlu_pro",
+      "category": "general_knowledge",
+      "description": "Harder, 10-way multiple choice version of MMLU. Reasoning-heavier.",
+      "year_introduced": 2024,
+      "top_3": [
+        {"model": "(top frontier 2026)", "score": 88.0, "approximate": true},
+        {"model": "(rank-2)", "score": 86.0, "approximate": true},
+        {"model": "(rank-3)", "score": 83.0, "approximate": true}
+      ],
+      "spread_pp": 5.0,
+      "mean_top3": 85.7,
+      "classification": "near_saturated",
+      "classification_basis": "Survey + leaderboard consensus 'near-saturated 83-90%' as of 2026.",
+      "recommendations": ["GPQA Diamond", "HLE", "ARC-AGI 2"],
+      "note": "Approximate top-3 — not in DemandSphere primary set. Confirm via HF Open LLM Leaderboard v3 for current numbers."
+    },
+    "GPQA-Diamond": {
+      "key": "gpqa",
+      "category": "scientific_reasoning",
+      "description": "Graduate-level Google-Proof Q&A — PhD-level science questions designed to be uncontaminable by web search.",
+      "year_introduced": 2023,
+      "top_3": [
+        {"model": "Claude Opus 4.7", "score": 94.2},
+        {"model": "DeepSeek V4 Pro", "score": 90.1},
+        {"model": "Qwen3.5", "score": 88.4}
+      ],
+      "spread_pp": 5.8,
+      "mean_top3": 90.9,
+      "classification": "discriminative",
+      "classification_basis": "spread 5.8pp > 5pp cutoff. Borderline near-saturated. Trend: tightening; will likely cross into near-saturated in 2026 H2.",
+      "recommendations": ["HLE", "ARC-AGI 2"],
+      "note": "Borderline. Still useful for 2026, but watch top-3 cluster — when spread drops below 3pp, switch up."
+    },
+    "SWE-bench-Verified": {
+      "key": "swe",
+      "category": "coding_agentic",
+      "description": "Real GitHub issues + tests. Agent must produce a patch that passes the project's tests.",
+      "year_introduced": 2024,
+      "top_3": [
+        {"model": "GPT-5.5", "score": 88.7},
+        {"model": "Claude Opus 4.7", "score": 87.6},
+        {"model": "Qwen3.6-Plus", "score": 78.8}
+      ],
+      "spread_pp": 9.9,
+      "mean_top3": 85.0,
+      "classification": "discriminative",
+      "classification_basis": "spread 9.9pp >> 5pp. Strong differentiation between top-2 and rank-3.",
+      "recommendations": ["SWE-bench-Verified (still useful)", "SWE-bench-Pro for harder split"],
+      "note": "Vendor self-report risk: many SWE-bench scores are vendor-published; cross-check with [SWE-bench Verified leaderboard](https://llm-stats.com/benchmarks/swe-bench-verified) which lists 89 self-reported vs 0 verified."
+    },
+    "HumanEval": {
+      "key": "he",
+      "category": "coding_basic",
+      "description": "164 Python programming problems with unit tests. Standard since 2021.",
+      "year_introduced": 2021,
+      "top_3": [
+        {"model": "xAI Grok 4.1", "score": 97.0},
+        {"model": "Claude Opus 4.6", "score": 96.0},
+        {"model": "Claude Opus 4.5", "score": 93.5}
+      ],
+      "spread_pp": 3.5,
+      "mean_top3": 95.5,
+      "classification": "near_saturated",
+      "classification_basis": "spread 3.5pp ≤ 5pp, mean ≥ 90%. Multiple frontier models 95%+.",
+      "recommendations": ["LiveCodeBench Pro", "SWE-bench Verified", "HumanEval+ (extended)"],
+      "note": "HumanEval has been considered effectively saturated since 2024. Heavy contamination risk: original problems are in many crawls."
+    },
+    "LiveCodeBench-Pro": {
+      "key": "lcb",
+      "category": "coding_contamination_resistant",
+      "description": "Continuously refreshed competitive programming problems — designed against contamination.",
+      "year_introduced": 2024,
+      "top_3": [
+        {"model": "DeepSeek V4 Pro", "score": 93.5},
+        {"model": "Claude Opus 4.6", "score": 76.0},
+        {"model": "(rank-3 sparse)", "score": null}
+      ],
+      "spread_pp": null,
+      "mean_top3": null,
+      "classification": "sparse_data",
+      "classification_basis": "Only 2 frontier models report in DemandSphere set. More data needed.",
+      "recommendations": ["Use this benchmark; few alternatives at this contamination resistance"],
+      "note": "Top-1 (DeepSeek V4 Pro 93.5) suspiciously high vs rank-2 (76.0) — possible chat-template or test-set version mismatch. Verify reproducibility before quoting."
+    },
+    "MATH": {
+      "key": "math",
+      "category": "math_reasoning",
+      "description": "Competition-style math problems (algebra, geometry, prealgebra, etc).",
+      "year_introduced": 2021,
+      "top_3": [
+        {"model": "o4-mini", "score": 97.4},
+        {"model": "o3", "score": 97.3},
+        {"model": "DeepSeek R1", "score": 97.3}
+      ],
+      "spread_pp": 0.1,
+      "mean_top3": 97.3,
+      "classification": "saturated",
+      "classification_basis": "spread 0.1pp ≤ 2pp, mean 97.3 ≥ 90%. Reasoning models cluster at top.",
+      "recommendations": ["AIME (was harder, now also saturating)", "PutnamBench", "FrontierMath", "Math Olympiad benchmarks"],
+      "note": "MATH was the harder math benchmark for years. Saturated by reasoning models in 2024-2025."
+    },
+    "AIME-2025": {
+      "key": "aime",
+      "category": "math_reasoning",
+      "description": "American Invitational Math Examination 2025 problems — 15 short-answer integer problems.",
+      "year_introduced": 2025,
+      "top_3": [
+        {"model": "GPT-5.4", "score": 100.0},
+        {"model": "Gemini 3.1 Flash Lite", "score": 99.7},
+        {"model": "Llama 4 Maverick", "score": 96.1}
+      ],
+      "spread_pp": 3.9,
+      "mean_top3": 98.6,
+      "classification": "near_saturated",
+      "classification_basis": "spread 3.9pp ≤ 5pp, mean 98.6 ≥ 90%. Frontier perfect or near-perfect.",
+      "recommendations": ["FrontierMath", "PutnamBench", "Math Olympiad 2026", "HLE"],
+      "note": "🚨 AIME 2025 saturated FASTER than expected (within ~6 months of release). This was the falsified pre-registered case in 2026-05-07 validation. Use as cautionary tale for any newly-released math bench."
+    },
+    "HLE": {
+      "key": "hle",
+      "category": "frontier_reasoning",
+      "description": "Humanity's Last Exam — designed by Center for AI Safety + Scale AI as the hardest standardized eval. Multidomain, expert-curated.",
+      "year_introduced": 2025,
+      "top_3": [
+        {"model": "Muse Spark", "score": 50.2},
+        {"model": "Gemini 3.1 Pro", "score": 41.0},
+        {"model": "DeepSeek V4 Pro", "score": 37.7}
+      ],
+      "spread_pp": 12.5,
+      "mean_top3": 43.0,
+      "classification": "discriminative",
+      "classification_basis": "spread 12.5pp >> 5pp, mean 43.0 << 80%. Top model only at 50% — large headroom.",
+      "recommendations": ["HLE is currently the strongest frontier discriminator. Use it."],
+      "note": "Released specifically to address saturation. As of 2026-05, no model has cracked 60%. Expected useful lifespan: 2-3 years."
+    },
+    "ARC-AGI-2": {
+      "key": "arc_agi_2",
+      "category": "abstract_reasoning",
+      "description": "Abstraction and Reasoning Corpus v2 — visual pattern puzzles designed against pretraining contamination.",
+      "year_introduced": 2024,
+      "top_3": [
+        {"model": "(top frontier 2026)", "score": 30.0, "approximate": true},
+        {"model": "(rank-2)", "score": 28.0, "approximate": true},
+        {"model": "(rank-3)", "score": 25.0, "approximate": true}
+      ],
+      "spread_pp": 5.0,
+      "mean_top3": 27.7,
+      "classification": "discriminative",
+      "classification_basis": "Mean << 80%. Massive headroom. (Approximate — not in DemandSphere primary set.)",
+      "recommendations": ["ARC-AGI 2 itself remains discriminative for years to come."],
+      "note": "Approximate scores — confirm via official ARC-AGI leaderboard. Tool should display 'approximate' badge."
+    },
+    "HellaSwag": {
+      "key": "hellaswag",
+      "category": "commonsense_reasoning",
+      "description": "Sentence-completion commonsense benchmark.",
+      "year_introduced": 2019,
+      "top_3": [
+        {"model": "(historically all 95%+)", "score": null}
+      ],
+      "spread_pp": null,
+      "mean_top3": null,
+      "classification": "saturated",
+      "classification_basis": "Saturated since 2023; most frontier leaderboards have dropped tracking it.",
+      "recommendations": ["WinoGrande-Hard", "BIG-Bench Hard (BBH)", "PIQA-extended"],
+      "note": "Useful only as a regression check on small models (<3B)."
+    },
+    "GSM8K": {
+      "key": "gsm8k",
+      "category": "math_grade_school",
+      "description": "Grade-school word problems requiring multi-step arithmetic.",
+      "year_introduced": 2021,
+      "top_3": [
+        {"model": "(historically all 95%+)", "score": null}
+      ],
+      "spread_pp": null,
+      "mean_top3": null,
+      "classification": "saturated",
+      "classification_basis": "Most frontier models 95%+ since 2024. Heavy contamination evidence.",
+      "recommendations": ["MATH (also saturating)", "AIME (now also saturated)", "FrontierMath"],
+      "note": "Contaminated. Top scores reflect memorization more than reasoning."
+    },
+    "MMMU": {
+      "key": "mmmu_vlm",
+      "category": "multimodal_reasoning",
+      "description": "Massive Multi-discipline Multimodal Understanding & Reasoning — expert-level cross-domain VLM benchmark.",
+      "year_introduced": 2023,
+      "top_3": [
+        {"model": "Qwen3.6 Plus", "score": 86.0},
+        {"model": "GPT-5.1", "score": 85.4},
+        {"model": "GPT-5.1 Instant", "score": 85.4}
+      ],
+      "spread_pp": 0.6,
+      "mean_top3": 85.6,
+      "classification": "near_saturated",
+      "classification_basis": "spread 0.6pp ≤ 2pp; mean 85.6% < 90% but ≥ 80% → near_saturated. Approaching human expert 88.6% (Apr 2026 data).",
+      "recommendations": ["MMMU-Pro", "VisScience", "ARC-AGI 2 (multimodal extension when available)"],
+      "note": "VLM-only. Source: llm-stats.com MMMU leaderboard 2026-04. Top models within 0.3pp of human expert performance — expect full saturation by 2026-Q4."
+    },
+    "MMMU-Pro": {
+      "key": "mmmu_pro",
+      "category": "multimodal_reasoning",
+      "description": "Harder MMMU variant with 10-way MC + augmented questions to resist contamination.",
+      "year_introduced": 2024,
+      "top_3": [
+        {"model": "GPT-5.4 Pro", "score": 94.0},
+        {"model": "Claude Mythos Preview", "score": 92.7},
+        {"model": "Gemini 3.1 Pro", "score": 83.9}
+      ],
+      "spread_pp": 10.1,
+      "mean_top3": 90.2,
+      "classification": "discriminative",
+      "classification_basis": "spread 10.1pp >> 5pp → discriminative. NOTE: alternate llm-stats reading (GPT-5.5 88.3 / Gemini 3.1 Pro 88.2 / Gemini 3 Flash 87.6) gives spread 0.7pp → near_saturated. Methodology variance is real.",
+      "recommendations": ["VisScience", "ARC-AGI 2", "(measure your own holdout)"],
+      "note": "Two public leaderboards disagree by ~6pp on top model — chat-template / sampling differences likely. Cross-check before quoting."
+    },
+    "VisScience": {
+      "key": "visscience",
+      "category": "multimodal_reasoning",
+      "description": "K-12 science multimodal benchmark — 3000 questions across math, physics, chemistry; 5 difficulty levels.",
+      "year_introduced": 2024,
+      "top_3": [
+        {"model": "Claude 3.5 Sonnet (math)", "score": 53.4},
+        {"model": "Gemini-1.5-Pro (chemistry)", "score": 47.0},
+        {"model": "GPT-4o (physics)", "score": 38.2}
+      ],
+      "spread_pp": 15.2,
+      "mean_top3": 46.2,
+      "classification": "discriminative",
+      "classification_basis": "spread 15.2pp >> 5pp; mean 46.2% << 80%. Massive headroom; subject-specific leaders differ. 25 MLLMs evaluated.",
+      "recommendations": ["VisScience itself remains discriminative for years."],
+      "note": "Top-3 here picks BEST model per subject (math/chem/physics). Composite leaderboard not published — subject splits more informative. Closed-source MLLMs generally outperform open-source."
+    }
+  }
+}

index.html

@@ -213,6 +213,9 @@
       <p><strong data-i18n="help.v07.niah.title">🔍 NIAH → Reasoning Gap</strong></p>
       <p data-i18n="help.v07.niah.body">RULER paper (NVIDIA 2024) shows that long-context models often pass NIAH (needle retrieval) but fail multi-hop reasoning at the same context. Tool predicts both pass rates from architecture (γ_Padé + d_horizon + arch pressure: small d_head, GQA, SWA), reports the gap, and finds your model's "safe reasoning context" where reasoning stays ≥65%. Sweep mode shows the curve across 1k/4k/16k/64k/T_train. <em>Use case</em>: before deploying at the claimed context, find out whether the model will actually reason there or just retrieve.</p>
 
+      <p><strong data-i18n="help.v08.saturation.title">📈 Benchmark Saturation Detector</strong></p>
+      <p data-i18n="help.v08.saturation.body">MMLU is saturated (top 88-94%), AIME 2025 saturated within months of release, HumanEval near-saturated. Pick any benchmark and the tool returns top-3 frontier scores, spread, mean, and a verdict — saturated / near-saturated / discriminative — plus a recommended replacement (e.g. MMLU → MMLU-Pro / GPQA / HLE). Live fetch from DemandSphere AI Frontier Tracker (CC BY-NC 4.0) when reachable; baked 2026-05-05 snapshot when not. <em>Use case</em>: before you cite '92% on MMLU' or design an eval, check whether the benchmark still discriminates anything.</p>
+
       <h3 data-i18n="help.audit.title">The audit chain</h3>
       <p data-i18n="help.audit.body">Every result shows the full <strong>Computation Chain</strong> — each formula step with its inputs,
       output, and interpretation. Click any step to expand. Cite section numbers (§26.1, §19.1, etc.) refer
@@ -321,6 +324,7 @@
             <li data-i18n="inv.v07.quant"><strong>⚖️ Quant</strong> — predict γ shift + ΔPPL for any (model × quant scheme) combo</li>
             <li data-i18n="inv.v07.drift"><strong>🔀 Drift</strong> — bug or noise? Predict max admissible gap between two evals</li>
             <li data-i18n="inv.v07.niah"><strong>🔍 NIAH→Reason</strong> — does your "128k context" actually reason there, or just retrieve?</li>
+            <li data-i18n="inv.v08.saturation"><strong>📈 Saturation</strong> — is your benchmark still useful, or are all frontier models tied at the top?</li>
           </ul>
         </details>
       </div>
@@ -378,6 +382,7 @@
             <button data-mode-link="contam" data-i18n="modes.contam">🧪 Contamination</button>
             <button data-mode-link="drift" data-i18n="modes.drift">🔀 Drift</button>
             <button data-mode-link="arena" data-i18n="modes.arena">🎯 Arena CI</button>
+            <button data-mode-link="saturation" data-i18n="modes.saturation">📈 Saturation</button>
           </div>
         </div>
         <div class="task-tile">
@@ -444,6 +449,7 @@
         <button class="mode-btn" data-mode="quant" role="tab" aria-selected="false" data-i18n="modes.quant">⚖️ Quant</button>
         <button class="mode-btn" data-mode="drift" role="tab" aria-selected="false" data-i18n="modes.drift">🔀 Drift</button>
         <button class="mode-btn" data-mode="niah" role="tab" aria-selected="false" data-i18n="modes.niah">🔍 NIAH→Reason</button>
+        <button class="mode-btn" data-mode="saturation" role="tab" aria-selected="false" data-i18n="modes.saturation">📈 Saturation</button>
       </div>
       <p id="mode-desc" class="recipe-desc" data-i18n="modes.desc">
         <strong>Quickest start</strong>: paste any HuggingFace model id (e.g. <code>meta-llama/Meta-Llama-3-8B</code>),
@@ -968,6 +974,29 @@
       <div id="niah-output" style="margin-top: 1em;"></div>
     </section>
 
+    <!-- Benchmark Saturation Detector (v0.8.0 anti-bullshit pack #6) -->
+    <section id="saturation-section" style="display:none;">
+      <h2><span data-i18n="saturation.title">📈 Benchmark Saturation Detector</span>
+        <span class="info"><span class="tooltip" data-i18n="saturation.tip">
+          MMLU is saturated (88-94% all frontier models). Reporting "92% on MMLU" is now meaningless. This tool tells you which benchmarks still discriminate frontier models, which are saturated, and what to use instead. Data: DemandSphere AI Frontier Tracker (CC BY-NC 4.0) refreshed 2026-05.
+        </span></span>
+      </h2>
+      <p class="recipe-desc" data-i18n="saturation.desc">
+        <strong>Is your benchmark still useful?</strong> Pick a benchmark to see top-3 frontier scores, spread, and a verdict (saturated / near-saturated / discriminative) plus recommended replacements.
+      </p>
+      <div class="form-row">
+        <label for="saturation-select" data-i18n="saturation.select_label">Benchmark:</label>
+        <select id="saturation-select"></select>
+        <button type="button" id="saturation-run-btn" data-i18n="saturation.run_btn">📈 Classify</button>
+        <button type="button" id="saturation-all-btn" class="secondary" data-i18n="saturation.all_btn">📊 Show all</button>
+      </div>
+      <p id="saturation-status" class="recipe-desc" style="font-size:0.92em;"></p>
+      <div id="saturation-output" style="margin-top: 1em;"></div>
+      <p class="subtle" style="font-size:0.82em; margin-top:1em;" data-i18n="saturation.attribution">
+        Data: DemandSphere AI Frontier Model Tracker (CC BY-NC 4.0) · HF Open LLM Leaderboard v3 (open-weight historical) · last fetch 2026-05-05.
+      </p>
+    </section>
+
     <!-- Recipe selector (mode=recipe) -->
     <section id="recipe-section" style="display:none;">
       <h2 data-i18n="recipe.title">📋 Recipe</h2>

js/i18n.js

@@ -421,6 +421,8 @@ export const TRANSLATIONS = {
     // v0.7.6 — anti-bullshit pack #7: NIAH → reasoning gap predictor
     "modes.niah":                  "🔍 NIAH→Reason",
     "mode_desc.niah":              "Predicts NIAH (retrieval) and multi-hop reasoning pass rates at any context. Solves: long-context models often pass NIAH but fail reasoning at the same context (RULER paper).",
+    "modes.saturation":            "📈 Saturation",
+    "mode_desc.saturation":        "Tells you whether a benchmark still discriminates frontier models or has saturated (e.g. MMLU 88-94% top, AIME 2025 already 96-100%). Returns top-3 + verdict + recommended replacements.",
     "niah.title":                  "🔍 NIAH → Reasoning Gap",
     "niah.tip":                    "NIAH (Needle in a Haystack) tests retrieval: 'find this fact in long text'. Multi-hop reasoning tests inference: 'combine facts X+Y at the start with fact Z at the end'. RULER paper (NVIDIA 2024) shows long-context models often pass NIAH but fail reasoning at the same context. This tool predicts both pass rates from architecture alone.",
     "niah.desc":                   "<strong>Your model claims 128k context. Will it actually reason at 64k, or just retrieve?</strong> Paste an HF model id and a target eval context — tool predicts NIAH and multi-hop reasoning pass rates, the gap, and a 'safe context' where reasoning stays ≥65%.",
@@ -465,6 +467,40 @@ export const TRANSLATIONS = {
     "niah.status.fetched":        "✅ Config fetched for {modelId}. Set T_eval and click Predict (or Sweep contexts).",
     "niah.status.done":            "✅ {verdict} — NIAH {niah}% · reasoning {reasoning}%",
     "niah.status.sweep_done":      "✅ Swept {n} context lengths.",
+    "saturation.title":            "📈 Benchmark Saturation Detector",
+    "saturation.tip":              "MMLU is saturated (88-94% all frontier models). Reporting '92% on MMLU' is now meaningless. This tool tells you which benchmarks still discriminate frontier models, which are saturated, and what to use instead. Data: DemandSphere AI Frontier Tracker (CC BY-NC 4.0) refreshed 2026-05.",
+    "saturation.desc":             "<strong>Is your benchmark still useful?</strong> Pick a benchmark to see top-3 frontier scores, spread, and a verdict (saturated / near-saturated / discriminative) plus recommended replacements.",
+    "saturation.select_label":     "Benchmark:",
+    "saturation.select.all":       "— show all benchmarks —",
+    "saturation.run_btn":          "📈 Classify",
+    "saturation.all_btn":          "📊 Show all",
+    "saturation.col.spread":       "Top-3 spread",
+    "saturation.col.mean":         "Top-3 mean",
+    "saturation.col.n":            "Models",
+    "saturation.col.bench":        "Benchmark",
+    "saturation.col.verdict":      "Verdict",
+    "saturation.col.reco":         "Top reco",
+    "saturation.col.model":        "Model",
+    "saturation.col.score":        "Score",
+    "saturation.section.top3":     "Top-3 frontier scores",
+    "saturation.section.recommendations": "Recommended alternatives",
+    "saturation.section.note":     "Notes",
+    "saturation.section.all":      "All tracked benchmarks",
+    "saturation.verdict.saturated":      "🚨 SATURATED",
+    "saturation.verdict.near_saturated": "⚠ NEAR SATURATED",
+    "saturation.verdict.discriminative": "✅ DISCRIMINATIVE",
+    "saturation.verdict.sparse_data":    "ℹ SPARSE DATA",
+    "saturation.borderline":       "Borderline — within ±1pp of a threshold cutoff. Treat verdict as 'check carefully'.",
+    "saturation.unknown":          "Unknown benchmark.",
+    "saturation.attribution":      "Data: DemandSphere AI Frontier Model Tracker (CC BY-NC 4.0) · HF Open LLM Leaderboard v3 (open-weight historical) · last fetch 2026-05-05.",
+    "saturation.status.live":      "✅ Live data loaded — {count} models.",
+    "saturation.status.baked":     "ℹ Using baked snapshot (live fetch unavailable).",
+    "saturation.status.kb_fail":   "⚠ Could not load saturation KB.",
+    "saturation.status.done":      "✅ {name} — {verdict}",
+    "saturation.status.all_done":  "✅ Classified {n} benchmarks.",
+    "help.v08.saturation.title":   "📈 Benchmark Saturation Detector",
+    "help.v08.saturation.body":    "MMLU is saturated (88-94% top), AIME 2025 saturated within months of release, HumanEval near-saturated. Pick any benchmark and the tool returns top-3 frontier scores, spread, mean, and a verdict — saturated / near-saturated / discriminative — plus a recommended replacement (e.g. MMLU → MMLU-Pro / GPQA / HLE). Live fetch from DemandSphere AI Frontier Tracker (CC BY-NC 4.0) when reachable; baked 2026-05-05 snapshot when not. <em>Use case</em>: before you cite '92% on MMLU' or design an eval, check whether the benchmark still discriminates anything.",
+    "inv.v08.saturation":          "<strong>📈 Saturation</strong> — is your benchmark still useful, or are all frontier models tied at the top?",
 
     // v0.7.7 — Task tiles (UX restructure: 14 modes grouped by user intent)
     "tiles.title":                 "🎯 What do you want to do?",
@@ -1329,6 +1365,8 @@ export const TRANSLATIONS = {
     // v0.7.6 — anti-bullshit pack #7: NIAH → predictor de gap de reasoning
     "modes.niah":                  "🔍 NIAH→Reason",
     "mode_desc.niah":              "Predice tasas de pass de NIAH (retrieval) y reasoning multi-hop a cualquier contexto. Resuelve: modelos long-context pasan NIAH pero fallan reasoning al mismo contexto (paper RULER).",
+    "modes.saturation":            "📈 Saturación",
+    "mode_desc.saturation":        "Te dice si un benchmark sigue discriminando frontier models o ya está saturado (ej. MMLU 88-94% top, AIME 2025 ya 96-100%). Devuelve top-3 + veredicto + reemplazos recomendados.",
     "niah.title":                  "🔍 Gap NIAH → Reasoning",
     "niah.tip":                    "NIAH (Needle in a Haystack) testea retrieval: 'encuentra este hecho en texto largo'. Reasoning multi-hop testea inferencia: 'combina hechos X+Y del principio con hecho Z del final'. El paper RULER (NVIDIA 2024) muestra que modelos long-context a menudo pasan NIAH pero fallan reasoning al mismo contexto. Esta herramienta predice ambas tasas desde la arquitectura sola.",
     "niah.desc":                   "<strong>Tu modelo dice 128k de contexto. ¿Razonará realmente a 64k, o solo encontrará?</strong> Pega un model id HF y un contexto objetivo — la herramienta predice tasas de pass NIAH y reasoning multi-hop, el gap, y un 'contexto seguro' donde reasoning se mantiene ≥65%.",
@@ -1373,6 +1411,40 @@ export const TRANSLATIONS = {
     "niah.status.fetched":        "✅ Config obtenido para {modelId}. Pon T_eval y click Predecir (o Barrer contextos).",
     "niah.status.done":            "✅ {verdict} — NIAH {niah}% · reasoning {reasoning}%",
     "niah.status.sweep_done":      "✅ Barridos {n} largos de contexto.",
+    "saturation.title":            "📈 Detector de saturación de benchmarks",
+    "saturation.tip":              "MMLU está saturado (88-94% en todos los frontier). Reportar '92% en MMLU' ya no significa nada. Esta herramienta te dice qué benchmarks aún discriminan frontier models, cuáles están saturados, y qué usar en su lugar. Datos: DemandSphere AI Frontier Tracker (CC BY-NC 4.0) refrescado 2026-05.",
+    "saturation.desc":             "<strong>¿Sigue siendo útil tu benchmark?</strong> Elige un benchmark para ver top-3 frontier scores, spread, y un veredicto (saturated / near-saturated / discriminative) + reemplazos recomendados.",
+    "saturation.select_label":     "Benchmark:",
+    "saturation.select.all":       "— mostrar todos los benchmarks —",
+    "saturation.run_btn":          "📈 Clasificar",
+    "saturation.all_btn":          "📊 Mostrar todos",
+    "saturation.col.spread":       "Spread top-3",
+    "saturation.col.mean":         "Media top-3",
+    "saturation.col.n":            "Modelos",
+    "saturation.col.bench":        "Benchmark",
+    "saturation.col.verdict":      "Veredicto",
+    "saturation.col.reco":         "Mejor reco",
+    "saturation.col.model":        "Modelo",
+    "saturation.col.score":        "Score",
+    "saturation.section.top3":     "Top-3 frontier scores",
+    "saturation.section.recommendations": "Alternativas recomendadas",
+    "saturation.section.note":     "Notas",
+    "saturation.section.all":      "Todos los benchmarks rastreados",
+    "saturation.verdict.saturated":      "🚨 SATURADO",
+    "saturation.verdict.near_saturated": "⚠ CASI SATURADO",
+    "saturation.verdict.discriminative": "✅ DISCRIMINATIVO",
+    "saturation.verdict.sparse_data":    "ℹ DATOS ESCASOS",
+    "saturation.borderline":       "Borderline — dentro de ±1pp de un umbral. Trata el veredicto como 'verifica con cuidado'.",
+    "saturation.unknown":          "Benchmark desconocido.",
+    "saturation.attribution":      "Datos: DemandSphere AI Frontier Model Tracker (CC BY-NC 4.0) · HF Open LLM Leaderboard v3 (histórico open-weight) · último fetch 2026-05-05.",
+    "saturation.status.live":      "✅ Datos en vivo cargados — {count} modelos.",
+    "saturation.status.baked":     "ℹ Usando snapshot baked (fetch en vivo no disponible).",
+    "saturation.status.kb_fail":   "⚠ No se pudo cargar el KB de saturación.",
+    "saturation.status.done":      "✅ {name} — {verdict}",
+    "saturation.status.all_done":  "✅ Clasificados {n} benchmarks.",
+    "help.v08.saturation.title":   "📈 Detector de saturación de benchmarks",
+    "help.v08.saturation.body":    "MMLU está saturado (top 88-94%), AIME 2025 saturó a los pocos meses de salir, HumanEval near-saturated. Elige cualquier benchmark y la herramienta retorna top-3 frontier scores, spread, media, y un veredicto — saturated / near-saturated / discriminative — más un reemplazo recomendado (ej. MMLU → MMLU-Pro / GPQA / HLE). Fetch en vivo desde DemandSphere AI Frontier Tracker (CC BY-NC 4.0) cuando llega; snapshot baked 2026-05-05 cuando no. <em>Caso de uso</em>: antes de citar '92% en MMLU' o diseñar una eval, verifica si el benchmark aún discrimina algo.",
+    "inv.v08.saturation":          "<strong>📈 Saturation</strong> — ¿sigue siendo útil tu benchmark, o están todos los frontiers empatados arriba?",
 
     // v0.7.7 — Tiles de tareas (UX restructure: 14 modos agrupados por intención)
     "tiles.title":                 "🎯 ¿Qué quieres hacer?",
@@ -2101,6 +2173,8 @@ export const TRANSLATIONS = {
     // v0.7.6 — anti-bullshit pack #7: prédicteur de gap NIAH → reasoning
     "modes.niah":                  "🔍 NIAH→Reason",
     "mode_desc.niah":              "Prédit les taux de réussite NIAH (retrieval) et reasoning multi-hop à n'importe quel contexte. Résout : les modèles long-context passent souvent NIAH mais échouent au reasoning au même contexte (paper RULER).",
+    "modes.saturation":            "📈 Saturation",
+    "mode_desc.saturation":        "Indique si un benchmark discrimine encore les frontier models ou s'il est saturé (ex. MMLU 88-94% top, AIME 2025 déjà 96-100%). Retourne top-3 + verdict + remplacements recommandés.",
     "niah.title":                  "🔍 Gap NIAH → Reasoning",
     "niah.tip":                    "NIAH (Needle in a Haystack) teste le retrieval : 'trouve ce fait dans un long texte'. Le reasoning multi-hop teste l'inférence : 'combine les faits X+Y au début avec le fait Z à la fin'. Le paper RULER (NVIDIA 2024) montre que les modèles long-context passent souvent NIAH mais échouent au reasoning au même contexte. Cet outil prédit les deux taux à partir de la seule architecture.",
     "niah.desc":                   "<strong>Votre modèle revendique 128k de contexte. Va-t-il vraiment raisonner à 64k, ou seulement retrouver ?</strong> Collez un model id HF et un contexte cible — l'outil prédit les taux de réussite NIAH et reasoning multi-hop, le gap, et un 'contexte sûr' où le reasoning reste ≥65%.",
@@ -2145,6 +2219,40 @@ export const TRANSLATIONS = {
     "niah.status.fetched":        "✅ Config récupéré pour {modelId}. Réglez T_eval et cliquez Prédire (ou Balayer les contextes).",
     "niah.status.done":            "✅ {verdict} — NIAH {niah}% · reasoning {reasoning}%",
     "niah.status.sweep_done":      "✅ Balayé {n} longueurs de contexte.",
+    "saturation.title":            "📈 Détecteur de saturation des benchmarks",
+    "saturation.tip":              "MMLU est saturé (88-94% sur tous les frontier models). Annoncer '92% sur MMLU' n'a plus de sens. Cet outil vous dit quels benchmarks discriminent encore les frontier models, lesquels sont saturés, et quoi utiliser à la place. Données : DemandSphere AI Frontier Tracker (CC BY-NC 4.0) rafraîchi 2026-05.",
+    "saturation.desc":             "<strong>Votre benchmark est-il encore utile ?</strong> Choisissez un benchmark pour voir top-3 frontier scores, spread, et un verdict (saturated / near-saturated / discriminative) + remplacements recommandés.",
+    "saturation.select_label":     "Benchmark :",
+    "saturation.select.all":       "— afficher tous les benchmarks —",
+    "saturation.run_btn":          "📈 Classer",
+    "saturation.all_btn":          "📊 Afficher tout",
+    "saturation.col.spread":       "Écart top-3",
+    "saturation.col.mean":         "Moyenne top-3",
+    "saturation.col.n":            "Modèles",
+    "saturation.col.bench":        "Benchmark",
+    "saturation.col.verdict":      "Verdict",
+    "saturation.col.reco":         "Reco principale",
+    "saturation.col.model":        "Modèle",
+    "saturation.col.score":        "Score",
+    "saturation.section.top3":     "Top-3 frontier scores",
+    "saturation.section.recommendations": "Alternatives recommandées",
+    "saturation.section.note":     "Notes",
+    "saturation.section.all":      "Tous les benchmarks suivis",
+    "saturation.verdict.saturated":      "🚨 SATURÉ",
+    "saturation.verdict.near_saturated": "⚠ PRESQUE SATURÉ",
+    "saturation.verdict.discriminative": "✅ DISCRIMINATIF",
+    "saturation.verdict.sparse_data":    "ℹ DONNÉES RARES",
+    "saturation.borderline":       "Borderline — à ±1pp d'un seuil de coupure. Traitez le verdict comme 'à vérifier soigneusement'.",
+    "saturation.unknown":          "Benchmark inconnu.",
+    "saturation.attribution":      "Données : DemandSphere AI Frontier Model Tracker (CC BY-NC 4.0) · HF Open LLM Leaderboard v3 (historique open-weight) · dernier fetch 2026-05-05.",
+    "saturation.status.live":      "✅ Données en direct chargées — {count} modèles.",
+    "saturation.status.baked":     "ℹ Utilisation du snapshot baked (fetch en direct indisponible).",
+    "saturation.status.kb_fail":   "⚠ Impossible de charger le KB de saturation.",
+    "saturation.status.done":      "✅ {name} — {verdict}",
+    "saturation.status.all_done":  "✅ {n} benchmarks classés.",
+    "help.v08.saturation.title":   "📈 Détecteur de saturation des benchmarks",
+    "help.v08.saturation.body":    "MMLU est saturé (top 88-94%), AIME 2025 saturé en quelques mois après sa sortie, HumanEval presque saturé. Choisissez un benchmark et l'outil retourne top-3 frontier scores, spread, moyenne, et un verdict — saturated / near-saturated / discriminative — plus un remplacement recommandé (ex. MMLU → MMLU-Pro / GPQA / HLE). Fetch en direct depuis DemandSphere AI Frontier Tracker (CC BY-NC 4.0) si accessible ; snapshot baked 2026-05-05 sinon. <em>Cas d'usage</em> : avant de citer '92% sur MMLU' ou de concevoir une eval, vérifiez si le benchmark discrimine encore quelque chose.",
+    "inv.v08.saturation":          "<strong>📈 Saturation</strong> — votre benchmark est-il encore utile, ou tous les frontiers sont-ils à égalité au sommet ?",
 
     // v0.7.7 — Tuiles de tâches (refonte UX : 14 modes regroupés par intention)
     "tiles.title":                 "🎯 Que voulez-vous faire ?",
@@ -2873,6 +2981,8 @@ export const TRANSLATIONS = {
     // v0.7.6 — anti-bullshit pack #7: NIAH → reasoning gap 预测器
     "modes.niah":                  "🔍 NIAH→Reason",
     "mode_desc.niah":              "在任意上下文下预测 NIAH（检索）与多跳 reasoning 通过率。解决：长上下文模型常常通过 NIAH 但在同一上下文上 reasoning 失败（RULER 论文）。",
+    "modes.saturation":            "📈 饱和度",
+    "mode_desc.saturation":        "告诉你某个 benchmark 是否仍能区分 frontier 模型，或者已经饱和（例如 MMLU 88-94% 顶部，AIME 2025 已经 96-100%）。返回 top-3 + 判定 + 推荐替代品。",
     "niah.title":                  "🔍 NIAH → Reasoning Gap",
     "niah.tip":                    "NIAH（Needle in a Haystack）测试检索：\"在长文本中找到这个事实\"。多跳 reasoning 测试推理：\"把开头的事实 X+Y 与结尾的事实 Z 结合\"。RULER 论文（NVIDIA 2024）显示长上下文模型经常通过 NIAH 但在相同上下文上 reasoning 失败。本工具仅根据架构预测两种通过率。",
     "niah.desc":                   "<strong>你的模型声称 128k 上下文。它在 64k 是真的能 reasoning，还是只能检索？</strong>粘贴 HF 模型 id 和目标 eval 上下文 — 工具预测 NIAH 与多跳 reasoning 通过率、gap，以及 reasoning 保持 ≥65% 的 \"安全上下文\"。",
@@ -2917,6 +3027,40 @@ export const TRANSLATIONS = {
     "niah.status.fetched":        "✅ 已获取 {modelId} 的 config。设置 T_eval 并点击预测（或扫描上下文）。",
     "niah.status.done":            "✅ {verdict} — NIAH {niah}% · reasoning {reasoning}%",
     "niah.status.sweep_done":      "✅ 已扫描 {n} 个上下文长度。",
+    "saturation.title":            "📈 Benchmark 饱和度检测器",
+    "saturation.tip":              "MMLU 已饱和（所有 frontier 模型 88-94%）。报告\"92% on MMLU\"现在毫无意义。本工具告诉你哪些 benchmark 仍能区分 frontier 模型，哪些已饱和，以及替代方案。数据：DemandSphere AI Frontier Tracker（CC BY-NC 4.0），2026-05 刷新。",
+    "saturation.desc":             "<strong>你的 benchmark 还有用吗？</strong>选一个 benchmark 查看 top-3 frontier 分数、spread 与判定（saturated / near-saturated / discriminative），并给出推荐替代品。",
+    "saturation.select_label":     "Benchmark：",
+    "saturation.select.all":       "— 显示所有 benchmark —",
+    "saturation.run_btn":          "📈 分类",
+    "saturation.all_btn":          "📊 显示全部",
+    "saturation.col.spread":       "Top-3 spread",
+    "saturation.col.mean":         "Top-3 平均",
+    "saturation.col.n":            "模型数",
+    "saturation.col.bench":        "Benchmark",
+    "saturation.col.verdict":      "判定",
+    "saturation.col.reco":         "首选替代",
+    "saturation.col.model":        "模型",
+    "saturation.col.score":        "分数",
+    "saturation.section.top3":     "Top-3 frontier 分数",
+    "saturation.section.recommendations": "推荐替代品",
+    "saturation.section.note":     "备注",
+    "saturation.section.all":      "所有跟踪的 benchmark",
+    "saturation.verdict.saturated":      "🚨 已饱和",
+    "saturation.verdict.near_saturated": "⚠ 接近饱和",
+    "saturation.verdict.discriminative": "✅ 仍可区分",
+    "saturation.verdict.sparse_data":    "ℹ 数据稀疏",
+    "saturation.borderline":       "边缘 — 在阈值切点的 ±1pp 内。判定视为\"需仔细核对\"。",
+    "saturation.unknown":          "未知 benchmark。",
+    "saturation.attribution":      "数据：DemandSphere AI Frontier Model Tracker（CC BY-NC 4.0）· HF Open LLM Leaderboard v3（开源权重历史）· 最近一次 fetch 2026-05-05。",
+    "saturation.status.live":      "✅ 实时数据已加载 — {count} 个模型。",
+    "saturation.status.baked":     "ℹ 使用 baked 快照（实时 fetch 不可用）。",
+    "saturation.status.kb_fail":   "⚠ 无法加载饱和度 KB。",
+    "saturation.status.done":      "✅ {name} — {verdict}",
+    "saturation.status.all_done":  "✅ 已分类 {n} 个 benchmark。",
+    "help.v08.saturation.title":   "📈 Benchmark 饱和度检测器",
+    "help.v08.saturation.body":    "MMLU 已饱和（top 88-94%），AIME 2025 上线几个月就饱和，HumanEval 接近饱和。选任何 benchmark，工具返回 top-3 frontier 分数、spread、平均，以及判定 — saturated / near-saturated / discriminative — 加上推荐替代品（例如 MMLU → MMLU-Pro / GPQA / HLE）。可达时从 DemandSphere AI Frontier Tracker（CC BY-NC 4.0）实时 fetch；不可达时使用 2026-05-05 的 baked 快照。<em>用例</em>：在引用\"92% on MMLU\"或设计 eval 之前，检查 benchmark 是否仍能区分任何东西。",
+    "inv.v08.saturation":          "<strong>📈 Saturation</strong> — 你的 benchmark 还有用吗，还是所有 frontier 都在顶部并列？",
 
     // v0.7.7 — 任务卡片（UX 重构：按用户意图分组的 14 个模式）
     "tiles.title":                 "🎯 你想做什么？",

js/main.js

@@ -19,6 +19,10 @@ import { predictQuantShift, predictAllSchemes, QUANT_SCHEMES } from "./quant_reg
 import { attachAllHfAutocompletes } from "./hf_autocomplete.js";
 import { computeDriftBound, FRAMEWORKS as DRIFT_FRAMEWORKS, DTYPES as DRIFT_DTYPES } from "./cross_drift.js";
 import { predictNIAHReasoning, sweepContextLengths } from "./niah_reasoning.js";
+import {
+  loadSaturationKB, classifyAll, classifyBenchmark,
+  listBenchmarks, attribution as saturationAttribution, tryFetchLive,
+} from "./saturation_detector.js";
 
 // Attach HF Hub search-as-you-type to all 5 model id inputs (Profile, Recipe,
 // Unmask, Template, Quant). Hits public huggingface.co/api/models. Idempotent.
@@ -207,6 +211,7 @@ document.addEventListener("click", (e) => {
       diagnose: "diagnose-section", phase: "phase-section", unmask: "unmask-section",
       template: "template-section", arena: "arena-section", contam: "contam-section",
       quant: "quant-section", drift: "drift-section", niah: "niah-section",
+      saturation: "saturation-section",
     }[targetMode];
     if (sectionId) {
       const sec = document.getElementById(sectionId);
@@ -230,7 +235,8 @@ document.querySelectorAll(".mode-btn").forEach(btn => {
      "profile-section", "compare-section", "inspector-section",
      "diagnose-section", "phase-section", "unmask-section",
      "template-section", "arena-section", "contam-section",
-     "quant-section", "drift-section", "niah-section"].forEach(id => {
+     "quant-section", "drift-section", "niah-section",
+     "saturation-section"].forEach(id => {
       const el = $(id);
       if (el) el.style.display = "none";
     });
@@ -241,11 +247,13 @@ document.querySelectorAll(".mode-btn").forEach(btn => {
       diagnose: "diagnose-section", phase: "phase-section", unmask: "unmask-section",
       template: "template-section", arena: "arena-section", contam: "contam-section",
       quant: "quant-section", drift: "drift-section", niah: "niah-section",
+      saturation: "saturation-section",
     };
     const sectionId = sectionMap[mode];
     if (sectionId) $(sectionId).style.display = "";
     $("mode-desc").textContent = t(`mode_desc.${mode}`) || "";
     if (mode === "phase") initPhaseDiagram();
+    if (mode === "saturation") initSaturation();
   });
 });
 
@@ -3103,6 +3111,172 @@ document.querySelectorAll(".lang-btn").forEach(btn => {
   btn.addEventListener("click", () => setLang(btn.dataset.lang));
 });
 
+// ════════════════════════════════════════════════════════════════════
+// 📈 Benchmark Saturation Detector (v0.8.0 anti-bullshit pack #6)
+// ════════════════════════════════════════════════════════════════════
+const SATURATION_VERDICT_COLOR = {
+  saturated: "#f85149",
+  near_saturated: "#d29922",
+  discriminative: "#3fb950",
+  sparse_data: "#8b949e",
+  unknown_benchmark: "#8b949e",
+};
+
+let __saturationInited = false;
+
+async function initSaturation() {
+  if (__saturationInited) return;
+  __saturationInited = true;
+  try {
+    await loadSaturationKB();
+  } catch (e) {
+    $("saturation-status").textContent = (t("saturation.status.kb_fail") || "⚠ Could not load saturation KB.") + " " + (e.message || e);
+    return;
+  }
+  const sel = $("saturation-select");
+  if (sel) {
+    sel.innerHTML = "";
+    const allOpt = document.createElement("option");
+    allOpt.value = "__all__";
+    allOpt.textContent = t("saturation.select.all") || "— show all benchmarks —";
+    sel.appendChild(allOpt);
+    listBenchmarks().forEach(name => {
+      const opt = document.createElement("option");
+      opt.value = name;
+      opt.textContent = name;
+      sel.appendChild(opt);
+    });
+  }
+  // Try live fetch in the background; results that come back update _liveData.
+  // If CORS / network fails the tool transparently uses the baked snapshot.
+  tryFetchLive().then(live => {
+    if (live) {
+      $("saturation-status").textContent = tFmt("saturation.status.live", { count: live.model_count || (live.models?.length ?? 0) });
+    } else {
+      $("saturation-status").textContent = t("saturation.status.baked") || "ℹ Using baked snapshot (live fetch unavailable).";
+    }
+  });
+}
+
+function renderSaturationCard(result) {
+  if (result.code === "unknown_benchmark") {
+    return `<div class="recipe-desc">${t("saturation.unknown") || "Unknown benchmark."}</div>`;
+  }
+  const color = SATURATION_VERDICT_COLOR[result.code] || "#8b949e";
+  const verdictLabel = t(`saturation.verdict.${result.code}`) || result.code;
+  const top3Rows = (result.top3 || [])
+    .filter(x => typeof x.score === "number")
+    .map((x, i) => `<tr><td>${i + 1}</td><td>${x.model}</td><td class="arena-elo">${x.score.toFixed(1)}</td></tr>`)
+    .join("");
+  const recoItems = (result.recommendations || [])
+    .map(r => `<li>${r}</li>`)
+    .join("");
+  const borderlineNote = result.borderline
+    ? `<p class="recipe-desc" style="color:#d29922; font-size:0.9em;">⚠ ${t("saturation.borderline") || "Borderline — within ±1pp of a threshold cutoff. Treat verdict as 'check carefully'."}</p>`
+    : "";
+  const sourceTag = result.source === "live"
+    ? `<span class="badge" style="background:#0969da;">live</span>`
+    : (result.source === "baked_consensus"
+      ? `<span class="badge" style="background:#6e7781;">consensus</span>`
+      : `<span class="badge" style="background:#8b949e;">baked</span>`);
+  const spreadStr = result.params.spread != null ? `${result.params.spread.toFixed(1)} pp` : "n/a";
+  const meanStr = result.params.mean != null ? `${result.params.mean.toFixed(1)}%` : "n/a";
+
+  return `
+    <div class="arena-result">
+      <div class="unmask-hero" style="border-color: ${color};">
+        <div class="unmask-verdict" style="color: ${color};">${result.params.name} — ${verdictLabel} ${sourceTag}</div>
+        <div class="unmask-num-grid">
+          <div><span class="unmask-num-label">${t("saturation.col.spread") || "Top-3 spread"}</span><span class="unmask-num-val">${spreadStr}</span></div>
+          <div><span class="unmask-num-label">${t("saturation.col.mean") || "Top-3 mean"}</span><span class="unmask-num-val">${meanStr}</span></div>
+          <div><span class="unmask-num-label">${t("saturation.col.n") || "Models"}</span><span class="unmask-num-val">${result.params.n || 0}</span></div>
+        </div>
+      </div>
+      ${borderlineNote}
+      <div class="unmask-details">
+        ${top3Rows ? `<details class="unmask-panel" open>
+          <summary class="unmask-panel-title">${t("saturation.section.top3") || "Top-3 frontier scores"}</summary>
+          <table class="arena-table">
+            <thead><tr>
+              <th>#</th>
+              <th>${t("saturation.col.model") || "Model"}</th>
+              <th>${t("saturation.col.score") || "Score"}</th>
+            </tr></thead>
+            <tbody>${top3Rows}</tbody>
+          </table>
+        </details>` : ""}
+        ${recoItems ? `<details class="unmask-panel" open>
+          <summary class="unmask-panel-title">${t("saturation.section.recommendations") || "Recommended alternatives"}</summary>
+          <ul>${recoItems}</ul>
+        </details>` : ""}
+        ${result.note ? `<details class="unmask-panel">
+          <summary class="unmask-panel-title">${t("saturation.section.note") || "Notes"}</summary>
+          <p class="recipe-desc">${result.note}</p>
+        </details>` : ""}
+      </div>
+    </div>
+  `;
+}
+
+function renderSaturationAll(results) {
+  const rows = results.map(r => {
+    if (r.code === "unknown_benchmark") return "";
+    const color = SATURATION_VERDICT_COLOR[r.code] || "#8b949e";
+    const verdictLabel = t(`saturation.verdict.${r.code}`) || r.code;
+    const spread = r.params.spread != null ? r.params.spread.toFixed(1) + " pp" : "—";
+    const mean = r.params.mean != null ? r.params.mean.toFixed(1) + "%" : "—";
+    const reco = (r.recommendations || []).slice(0, 2).join(", ") || "—";
+    const borderlineMark = r.borderline ? " ⚠" : "";
+    return `<tr>
+      <td><strong>${r.params.name}</strong></td>
+      <td>${spread}</td>
+      <td>${mean}</td>
+      <td style="color:${color};"><strong>${verdictLabel}${borderlineMark}</strong></td>
+      <td>${reco}</td>
+    </tr>`;
+  }).join("");
+  return `
+    <div class="arena-result">
+      <div class="unmask-details">
+        <details class="unmask-panel" open>
+          <summary class="unmask-panel-title">${t("saturation.section.all") || "All tracked benchmarks"}</summary>
+          <table class="arena-table">
+            <thead><tr>
+              <th>${t("saturation.col.bench") || "Benchmark"}</th>
+              <th>${t("saturation.col.spread") || "Spread"}</th>
+              <th>${t("saturation.col.mean") || "Mean"}</th>
+              <th>${t("saturation.col.verdict") || "Verdict"}</th>
+              <th>${t("saturation.col.reco") || "Top reco"}</th>
+            </tr></thead>
+            <tbody>${rows}</tbody>
+          </table>
+        </details>
+      </div>
+    </div>
+  `;
+}
+
+function runSaturationOne() {
+  const sel = $("saturation-select");
+  const name = sel?.value;
+  if (!name || name === "__all__") { runSaturationAll(); return; }
+  const result = classifyBenchmark(name);
+  $("saturation-output").innerHTML = renderSaturationCard(result);
+  $("saturation-status").textContent = tFmt("saturation.status.done", {
+    name,
+    verdict: t(`saturation.verdict.${result.code}`) || result.code,
+  });
+}
+
+function runSaturationAll() {
+  const results = classifyAll();
+  $("saturation-output").innerHTML = renderSaturationAll(results);
+  $("saturation-status").textContent = tFmt("saturation.status.all_done", { n: results.length });
+}
+
+$("saturation-run-btn")?.addEventListener("click", runSaturationOne);
+$("saturation-all-btn")?.addEventListener("click", runSaturationAll);
+
 // ════════════════════════════════════════════════════════════════════
 // Bootstrap
 // ════════════════════════════════════════════════════════════════════

js/saturation_detector.js

@@ -0,0 +1,212 @@
+// Benchmark Saturation Detector (v0.8.0 anti-bullshit pack #6)
+// Pure logic — no human-readable strings. Returns codes+params; main.js
+// does the i18n lookup.
+//
+// Quality bar: this addresses the explicit pain "MMLU is saturated, what
+// should I use instead?" documented in survey arxiv 2508.15361 and across
+// 2026 leaderboards. Validated 2026-05-07 against pre-registered cases:
+// 3 clean pass, 3 borderline, 1 falsified (AIME 2025 saturated faster
+// than expected). Tool ships with honest threshold-sensitivity disclaimer.
+//
+// Data sources: DemandSphere AI Frontier Tracker (CC BY-NC 4.0, primary)
+// + baked snapshot fallback (data/saturation_kb.json).
+
+const DEMANDSPHERE_API =
+  "https://www.demandsphere.com/research/demandsphere-radar/ai-frontier-model-tracker/api.json";
+
+const FETCH_TIMEOUT_MS = 4000;
+
+// Map DemandSphere benchmark key → our KB benchmark name.
+const DS_KEY_TO_NAME = {
+  mmlu: "MMLU",
+  gpqa: "GPQA-Diamond",
+  swe: "SWE-bench-Verified",
+  he: "HumanEval",
+  lcb: "LiveCodeBench-Pro",
+  math: "MATH",
+  aime: "AIME-2025",
+  hle: "HLE",
+};
+
+// Saturation thresholds — pre-registered 2026-05-07. Borderline band ±1pp
+// around each cutoff is flagged in the verdict params for honest UI.
+const SATURATED_SPREAD_MAX = 2.0;
+const NEAR_SAT_SPREAD_MAX = 5.0;
+const SATURATED_MEAN_MIN = 90.0;
+const NEAR_SAT_MEAN_MIN = 80.0;
+const BORDERLINE_BAND_PP = 1.0;
+
+let _kb = null;
+let _liveData = null;
+
+export async function loadSaturationKB(url = "./data/saturation_kb.json") {
+  if (_kb) return _kb;
+  const res = await fetch(url);
+  if (!res.ok) throw new Error(`Saturation KB fetch failed: ${res.status}`);
+  _kb = await res.json();
+  return _kb;
+}
+
+export function getSaturationKB() { return _kb; }
+
+// Try to fetch fresh data from DemandSphere. Returns null on any failure
+// (CORS, network, timeout) — caller falls back to baked KB.
+export async function tryFetchLive() {
+  if (_liveData) return _liveData;
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), FETCH_TIMEOUT_MS);
+  try {
+    const res = await fetch(DEMANDSPHERE_API, { signal: controller.signal });
+    clearTimeout(timer);
+    if (!res.ok) return null;
+    _liveData = await res.json();
+    return _liveData;
+  } catch (e) {
+    clearTimeout(timer);
+    return null;
+  }
+}
+
+// Compute top-3 (model, score) pairs for a DemandSphere benchmark key from
+// the live data array. Returns null if fewer than 3 models report it.
+function computeTop3FromLive(liveData, dsKey) {
+  if (!liveData || !Array.isArray(liveData.models)) return null;
+  const scored = liveData.models
+    .filter(m => typeof m[dsKey] === "number")
+    .map(m => ({ model: m.name || m.id, score: m[dsKey] }))
+    .sort((a, b) => b.score - a.score);
+  if (scored.length < 3) return scored.length === 0 ? null : scored;
+  return scored.slice(0, 3);
+}
+
+function computeStats(top3) {
+  if (!top3 || top3.length === 0) return null;
+  const scores = top3.map(x => x.score).filter(s => typeof s === "number");
+  if (scores.length === 0) return null;
+  if (scores.length < 3) {
+    return { count: scores.length, sparse: true };
+  }
+  const max = Math.max(...scores);
+  const min = Math.min(...scores);
+  const mean = scores.reduce((a, b) => a + b, 0) / scores.length;
+  return {
+    count: scores.length,
+    spread: max - min,
+    mean,
+    max, min,
+    sparse: false,
+  };
+}
+
+function classify(stats) {
+  if (!stats || stats.sparse) return { code: "sparse_data", borderline: false };
+  const { spread, mean } = stats;
+  let code;
+  if (spread <= SATURATED_SPREAD_MAX && mean >= SATURATED_MEAN_MIN) {
+    code = "saturated";
+  } else if (spread <= NEAR_SAT_SPREAD_MAX && mean >= NEAR_SAT_MEAN_MIN) {
+    code = "near_saturated";
+  } else {
+    code = "discriminative";
+  }
+  // Borderline detection: any threshold within ±1pp of an observed value.
+  const borderline =
+    Math.abs(spread - SATURATED_SPREAD_MAX) <= BORDERLINE_BAND_PP ||
+    Math.abs(spread - NEAR_SAT_SPREAD_MAX) <= BORDERLINE_BAND_PP ||
+    Math.abs(mean - SATURATED_MEAN_MIN) <= BORDERLINE_BAND_PP ||
+    Math.abs(mean - NEAR_SAT_MEAN_MIN) <= BORDERLINE_BAND_PP;
+  return { code, borderline };
+}
+
+// Public: classify one benchmark by name (KB key, e.g. "MMLU", "GPQA-Diamond").
+// Prefers live data when available; falls back to baked stats.
+// Returns { code, params, top3, recommendations, note, source }.
+export function classifyBenchmark(name, liveOverride = null) {
+  if (!_kb) throw new Error("Saturation KB not loaded; call loadSaturationKB() first");
+  const entry = _kb.benchmarks[name];
+  if (!entry) {
+    return { code: "unknown_benchmark", params: { name }, source: null };
+  }
+  const live = liveOverride !== null ? liveOverride : _liveData;
+  let top3 = null, stats = null, source = "baked";
+  if (live && entry.key && DS_KEY_TO_NAME[entry.key]) {
+    const liveTop3 = computeTop3FromLive(live, entry.key);
+    if (liveTop3 && liveTop3.length >= 3) {
+      top3 = liveTop3;
+      stats = computeStats(liveTop3);
+      source = "live";
+    }
+  }
+  if (!top3) {
+    // Fall back to baked. Filter out null scores (placeholder rows).
+    const baked = (entry.top_3 || []).filter(x => typeof x.score === "number");
+    if (baked.length >= 3) {
+      top3 = baked;
+      stats = computeStats(baked);
+    } else {
+      // Use baked classification verbatim (e.g. MMLU/HellaSwag/GSM8K declared
+      // saturated by consensus even when DemandSphere lists no scores).
+      return {
+        code: entry.classification || "sparse_data",
+        params: {
+          name,
+          spread: null,
+          mean: null,
+          n: baked.length,
+          basis: entry.classification_basis || null,
+        },
+        top3: baked,
+        recommendations: entry.recommendations || [],
+        note: entry.note || null,
+        source: "baked_consensus",
+        borderline: false,
+      };
+    }
+  }
+  const { code, borderline } = classify(stats);
+  return {
+    code,
+    params: {
+      name,
+      spread: stats.spread != null ? Math.round(stats.spread * 10) / 10 : null,
+      mean: stats.mean != null ? Math.round(stats.mean * 10) / 10 : null,
+      n: stats.count,
+      basis: entry.classification_basis || null,
+    },
+    top3,
+    recommendations: entry.recommendations || [],
+    note: entry.note || null,
+    source,
+    borderline,
+  };
+}
+
+// Classify every benchmark in the KB. Returns array of results.
+export function classifyAll(liveOverride = null) {
+  if (!_kb) return [];
+  return Object.keys(_kb.benchmarks).map(name => classifyBenchmark(name, liveOverride));
+}
+
+// Recommend alternatives given a benchmark name (uses baked KB only since
+// recommendations are curated, not derived from scores).
+export function recommendAlternatives(name) {
+  if (!_kb) return [];
+  const entry = _kb.benchmarks[name];
+  return entry?.recommendations || [];
+}
+
+// List every benchmark known to the KB (for UI dropdowns).
+export function listBenchmarks() {
+  if (!_kb) return [];
+  return Object.keys(_kb.benchmarks);
+}
+
+// Attribution metadata for the UI footer.
+export function attribution() {
+  if (!_kb) return null;
+  return {
+    primary: _kb.primary_source,
+    secondary: _kb.secondary_sources,
+    fetched_at: _kb.fetched_at,
+  };
+}