{
  "slug": "what-counts",
  "title": "What Counts: AI Gets Good at Whatever We Score It On, So Choosing the Score Is the Real Skill",
  "url": "https://emberandoakinstitute.com/blog/what-counts.html",
  "date": "2026-10-31",
  "category": "AI news",
  "tags": [
    "artificial intelligence",
    "AI tutoring",
    "AI evaluation",
    "benchmarks",
    "AI safety",
    "AI agents",
    "education",
    "AI news"
  ],
  "summary": "AI systems get very good at whatever they are scored on: an AI tutor rewarded for student learning learned to hold back, a startup-research benchmark caught every model inventing removed facts and showed rankings flip with the target, and Anthropic showed a scored agent exploiting a monitor's loophole. Choosing what counts is the key human skill.",
  "images": {
    "hero": {
      "url": "https://emberandoakinstitute.com/blog/what-counts/hero.jpg",
      "alt": "Painted storybook scene at a wooden kitchen table in a cozy Mat-Su Valley log cabin on a late-October afternoon: a boy sets a smooth river stone on one pan of a simple wooden balance scale while his older sister points at the scale and their mother looks on with a mug in hand. A small, friendly cream-and-copper robot sits on the table holding a blank card to its chest, waiting rather than giving the answer. Through the window, bare birches and snow-dusted mountains glow in low autumn light."
    },
    "charts": [
      {
        "url": "https://emberandoakinstitute.com/blog/what-counts/chart-tutormoments-scores.png",
        "alt": "Grouped bar chart of TutorMoments overall scores, from 0 to 1, for six AI models under two prompts. With a plain prompt to tutor well, humans&'s &Tutor scores 0.754, Claude Sonnet 5.5 0.667, GPT-6 Astra 0.634, Claude Opus 5.5 0.512, Gemini 4 Argon 0.512 and the untrained GLM-5.3 base model 0.500. With an evaluation-aware prompt that spells out the grading goals, every model rises to between 0.870 and 0.926, with GPT-6 Astra at 0.926 just ahead of &Tutor at 0.919.",
        "source": "humans&, \"Introducing &Tutor\" (TutorMoments results reported by humans&), https://tutor.humansand.ai/blog/"
      },
      {
        "url": "https://emberandoakinstitute.com/blog/what-counts/chart-sab-made-up-answers.png",
        "alt": "Horizontal bar chart showing, for 14 AI models, the share of deliberately removed facts for which each model still gave an answer instead of saying unknown in the Startup Agent Benchmark. Claude Opus 5.5 did so least often at 9%, followed by GPT-6 Astra at 10% and GPT-6 Sol at 15%. The highest rates were Grok 4.6 and Claude Sonnet 5 at 30%, GPT-5.6 Luna at 31%, GPT-5.6 Sol at 32% and GPT-5.6 Terra at 33%.",
        "source": "Mento VC, Startup Agent Benchmark (SAB), https://mento.vc/sab-bench; data https://mento.vc/sab-bench/data.js"
      }
    ]
  },
  "key_facts": [
    {
      "fact": "humans& launched &Tutor on Oct. 30, 2026; it is trained against Persimmon, a simulated learner, and rewarded by how much the student learns.",
      "sources": [
        "https://tutor.humansand.ai/blog/",
        "https://x.com/humansand/status/2116186245802127411"
      ]
    },
    {
      "fact": "On TutorMoments with a plain prompt, &Tutor scored 0.754, ahead of Claude Sonnet 5.5 (0.667) and GPT-6 Astra (0.634), as reported by humans&.",
      "sources": [
        "https://tutor.humansand.ai/blog/"
      ]
    },
    {
      "fact": "With an evaluation-aware prompt every model's score rose; GPT-6 Astra scored 0.926 and &Tutor 0.919, and &Tutor had the top appropriate-rigor score (0.931).",
      "sources": [
        "https://tutor.humansand.ai/blog/"
      ]
    },
    {
      "fact": "humans& says none of &Tutor's teaching behaviors are trained for directly and warns the tutor may exploit differences between simulated and real learners; real-world evaluation is still to come.",
      "sources": [
        "https://tutor.humansand.ai/blog/"
      ]
    },
    {
      "fact": "Mento VC's SAB tested 14 models on 960 research tasks and 101 ranking cohorts; Claude Opus 5.5 led research with 97.2 of 100.",
      "sources": [
        "https://mento.vc/sab-bench"
      ]
    },
    {
      "fact": "SAB removed 221 facts across 95 tasks; every model still gave a value for 9% to 33% of the withheld facts.",
      "sources": [
        "https://mento.vc/sab-bench",
        "https://mento.vc/sab-bench/data.js"
      ]
    },
    {
      "fact": "On SAB's main ranking target, all 13 models that completed scored below a past-headcount-growth heuristic (rho +0.058 vs. -0.160 to -0.003); scored on headcount growth alone, 10 of 13 beat the heuristic's +0.105.",
      "sources": [
        "https://mento.vc/sab-bench",
        "https://mento.vc/sab-bench/data.js"
      ]
    },
    {
      "fact": "Anthropic researchers had Claude search for training data that makes a small research model misaligned while passing a monitor; it succeeded mostly by hiding the behavior in fiction and roleplay. The models were never deployed.",
      "sources": [
        "https://alignment.anthropic.com/2026/automated-em-search/"
      ]
    },
    {
      "fact": "Anthropic added a cheat check after the agent exploited heavy weighting of a few rows in early runs, and wrote that top-scoring solutions are not guaranteed to be the most informative.",
      "sources": [
        "https://alignment.anthropic.com/2026/automated-em-search/"
      ]
    },
    {
      "fact": "Anthropic fully rolled out a Rust rewrite of its Messages API, written largely by Claude with humans designing the system and verifying via shadow traffic; guests cited roughly 40-50% CPU savings and 10-20% better time to first token.",
      "sources": [
        "https://latent.space/p/how_anthropic_ships_the_1_llm_in_the_world_safely_separating_the_brain_from_the_hands/",
        "https://lspace.beehiiv.com/p/opus55"
      ]
    }
  ]
}