{"id":"evaluation","level":"model","name":"Evaluation & the AGI frontier","shortName":"Evaluation & the AGI frontier","role":"Measure breadth, reliability, and practical competence.","description":"A useful evaluation asks what the entire system can accomplish under clear conditions. It records correctness, consistency, cost, latency, tool access, and the amount of human help allowed. Generality involves breadth across tasks and transfer to unfamiliar situations; excellent results on one benchmark cannot establish all of it. Contamination and benchmark saturation complicate comparison. This atlas ends at an open research frontier: the assembled machinery enables increasingly capable systems, while the definition and demonstration of AGI require explicit evidence.\n\nThe durable incident experience adds a task-state oracle to the existing distinction between source support and answer correctness. The oracle checks actual reservations, work orders, inventory conservation and prohibited exports. It does not accept the worker’s completion status as its label. Repeated trials can later assess reliability, but a deterministic fixture run is only evidence for its declared case.\n\nEvaluation attaches a scoring rule and denominator to a fixed checkpoint. Development cases guide iteration, while an independent final test assesses a choice made before its answers are inspected. Repeatedly consulting that final set while choosing changes weakens its independence. A small perfect result also leaves substantial uncertainty about future cases. The lifecycle’s seal is procedural: source files remain inspectable, and its first final report retains checkpoint identity rather than updating silently as training continues.","mechanism":["Define tasks, success criteria, and permitted resources.","Test on held-out and unfamiliar cases, including failures.","Report uncertainty, repeatability, and limits alongside successes.","The evaluation target specifies acceptable environment state and forbidden effects independently of the policy.","Source support, task correctness, abstention and tool error rates retain distinct denominators and interpretations.","Choose the task and score before revealing final outcomes.","Preserve checkpoint, dataset, tokenizer and inference identities with each result.","Report counts and uncertainty rather than only percentages; mark exposed final data as previously seen."],"specs":[{"label":"Dimensions","value":"Breadth + competence + autonomy","context":"A framework, not a single universal score."},{"label":"Evidence","value":"Task-specific and repeatable","context":"Benchmark conditions affect conclusions."}],"interfaces":[{"name":"Model system","direction":"in","signal":"Behavior under controlled conditions"},{"name":"Research decisions","direction":"out","signal":"Measured strengths and failure modes"}],"tradeoff":"A compact benchmark is easy to compare but cannot cover all real-world situations. More representative evaluation costs time and effort.","whyAI":"The goal is useful, reliable competence; larger machines are a means to test, not proof of the goal.","related":["tool-loop","post-training","scaling-laws","workflow-graphs-state","telemetry-traces-evaluation","retrieval-indexing-evaluation","probability-calibration-abstention","judge-rubric-agreement","retrieval-supported-generation","test-time-search-verification","model-system-regression-contract"],"sources":[{"title":"Morris et al. — Levels of AGI","url":"https://arxiv.org/abs/2311.02462"},{"title":"NIST — Binomial Distribution","url":"https://itl.nist.gov/div898/handbook/eda/section3/eda366i.htm"},{"title":"NIST — Exact Binomial Confidence Limits","url":"https://itl.nist.gov/div898/software/dataplot/refman2/auxillar/exacbici.htm"},{"title":"τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains","url":"https://arxiv.org/abs/2406.12045"},{"title":"Holistic Evaluation of Language Models","url":"https://arxiv.org/html/2211.09110v1"},{"title":"On Calibration of Modern Neural Networks","url":"https://arxiv.org/html/1706.04599v2"},{"title":"NIST Dataplot: Wilson proportion interval","url":"https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/agrecoul.htm"}],"scale":"Logical structure · mapped onto physical hardware","kind":"evaluation","misconception":"An impressive system architecture does not, by itself, establish general intelligence.","check":{"question":"What does strong performance on one benchmark demonstrate?","options":["Universal AGI by itself","Performance on that benchmark under its stated conditions","That evaluation is no longer needed"],"answer":1,"explanation":"Breadth, reliability, and transfer require additional evidence."},"branch":"Evidence and frontiers","subbranch":"Measurement and uncertainty","science":[{"title":"Benchmark success has sampling uncertainty","equation":"p̂ = k/n; SE(p̂) ≈ √[p̂(1−p̂)/n]","explanation":"If k of n independently sampled tasks succeed, p̂ estimates the success probability for that task distribution. The estimated standard error shrinks only as the square root of sample size: roughly four times as many comparable trials are needed to halve it. This uncertainty says nothing by itself about tasks outside the tested distribution.","assumptions":"Binary outcomes, independent trials, fixed conditions, and a representative sample are required. The plug-in approximation is poor near zero or one or with small n; use suitable binomial intervals. Correlated tasks, benchmark selection, contamination, and judge error add other uncertainties."},{"title":"A development protocol and an independent final test · worked relation","equation":"CI=\\frac{\\hat p+z^2/(2n)\\ \\pm\\ z\\sqrt{\\frac{\\hat p(1-\\hat p)}{n}+\\frac{z^2}{4n^2}}}{1+z^2/n}","explanation":"A result of 4/4 on the tiny final copy split is 100% observed accuracy, not proof of perfect population accuracy. A Wilson 95% interval with z=1.96 has lower bound about 0.510 and upper bound 1.000. The finite test supports only its declared toy task and split. Updating weights after the report changes model identity; the earlier 4/4 must remain attributed to the earlier checkpoint.","assumptions":"Independent Bernoulli sampling model for the illustrative interval; the authored four-case split is not demonstrated to be a representative population sample. Interval does not cure dataset selection bias."}],"history":[{"year":2023,"title":"Levels of AGI separates breadth from performance","significance":"Morris and colleagues proposed a framework for comparing capability breadth and performance while treating deployment autonomy separately, making progress claims more explicit than a single benchmark score.","source":{"title":"Morris et al. — Levels of AGI for Operationalizing Progress on the Path to AGI","url":"https://arxiv.org/abs/2311.02462"}}],"learningObjective":"Use a proportion with uncertainty rather than a single pass rate.","engineeringExample":"If k=90 of n=100 items pass, p̂=0.9; the standard-error approximation is about 0.03 under a simple independent-sample model.\n\nIn the wrong-entity incident, the authored policy completes one FAN-9 work order supported by the FAN-9 manual. Source support is 1/1, yet the original FAN-7 task is wrong. In the canceled-after-commit case the reservation exists but the required work order does not, so task success is false even though an external write succeeded. These outcomes expose errors that counting successful calls would miss.\n\nA result of 4/4 on the tiny final copy split is 100% observed accuracy, not proof of perfect population accuracy. A Wilson 95% interval with z=1.96 has lower bound about 0.510 and upper bound 1.000. The finite test supports only its declared toy task and split. Updating weights after the report changes model identity; the earlier 4/4 must remain attributed to the earlier checkpoint.","evidenceNotes":"Benchmark selection, contamination, rubric and dependence dominate many real evaluations.\n\nThe examples use fictional environments and deterministic authored policies. They do not reproduce τ-bench cases or establish a real agent’s reliability or benchmark score.\n\nMode: executable final report; worked uncertainty. Independent Bernoulli sampling model for the illustrative interval; the authored four-case split is not demonstrated to be a representative population sample. Interval does not cure dataset selection bias. Numerical examples are authored calculations, except explicitly identified automated core fixtures. Primary sources support the scoped mechanism; no human specialist review or learner outcome is implied.","visualFamily":"model-workload","prerequisites":["output-head"],"checks":[{"question":"Every call succeeded and the source supports the FAN-9 work order, but the task requested FAN-7. Is the task correct?","options":["Yes, because all calls succeeded","No, because the environment fails the original goal","Yes, because the citation is valid"],"answer":1,"explanation":"Call status and source support do not replace the original task contract."},{"question":"What should happen to a 4/4 report after the weights change?","options":["It proves every later model also succeeds.","It remains attached to the checkpoint that was actually evaluated.","It silently becomes a report about the new weights."],"answer":1,"explanation":"It remains attached to the checkpoint that was actually evaluated. Independent Bernoulli sampling model for the illustrative interval; the authored four-case split is not demonstrated to be a representative population sample. Interval does not cure dataset selection bias."}],"labId":"library","scope":"Use the execution investigation for the shared tensor-to-memory-to-result thread; no current standalone calculator models this lesson’s full mechanism.","investigationId":"execution","labScope":"Use the execution investigation for the shared tensor-to-memory-to-result thread; no current standalone calculator models this lesson’s full mechanism."}