// evaluation.reliability

    Evaluation is not reliable

    At scale, evaluation becomes the weakest and least reliable layer in any learning system.

    // system.assumptions

    The model that breaks at scale

    01

    Scores represent performance

    02

    Evaluation is a neutral step

    03

    Results are sufficient for decisions

    This assumption holds at small scale. Evaluation appears stable. Scores appear comparable.

    // failure.modes

    Three failures that compound at scale

    01

    Inconsistency

    Evaluation outcomes vary across evaluators, contexts, and time. Reliability cannot be guaranteed.

    02

    Loss of information

    Evaluation generates detailed signals during assessment, but systems retain only aggregated outcomes.

    03

    Inability to reuse

    Scores cannot be used to improve learning, drive personalisation, or support consistent decisions.

    // signal.capture

    Every response contains learning signals

    SCORES RETAIN

    72

    SCORE_ONLY

    SIGNALS.GENERATED

    concept_gaptrue
    error_type"partial_recall"
    rubric_alignment0.71
    step_reasoningincomplete

    When evaluation is reduced to a score, these signals are lost. What remains is a number.

    // system.impact

    When evaluation cannot be trusted, the system fails

    SYSTEM FAILURES

    No memory of learning
    Personalisation becomes inconsistent
    Decisions based on unstable inputs
    Evaluation cannot be audited

    WHAT THIS AFFECTS

    Learning outcomes
    Program reliability
    Governance and accountability
    Trust in the system itself

    This is not a measurement issue. It is a system reliability problem.

    // evaluation.control_point

    Evaluation is where signals are either captured — or lost

    Responses are evaluated
    Judgments are made
    Learning signals are generated

    If signals are not captured here, they are lost permanently. No downstream system can recover them.

    SIGNAL.CAPTURE.POINT

    STUDENT.RESPONSEreceived
    EVALUATION.RUNSactive
    SIGNALS.GENERATEDcaptured ✓
    MEMORY.WRITTENpersisted ✓
    // without capture at this point, signals are discarded

    // structural.requirement

    Reliability requires a structural shift

    This is not an optimization — it is a structural requirement for modern learning systems.

    SIGNALS MUST BE

    Captured during evaluation
    Structured into machine-readable formats
    Stored as persistent learning memory

    // cgf.evaluation_infrastructure

    CrazyGoldFish makes evaluation reliable

    // 93% ICC

    Consistent scoring

    93% reliability in total scoring alignment (ICC) — independent of evaluator drift.

    // signal.capture

    Signals retained

    Every response produces a LearningSignal — concept_gap, confidence_score, error_type — captured before they can be discarded.

    // memory.layer

    Persistent learning memory

    Signals are stored as persistent learning memory and reused for personalisation, remediation, and structured learning intelligence.

    // COMMON.QUESTIONS

    Questions about evaluation reliability

    Evaluation reliability fails because human interpretation varies across evaluators, rubrics, and contexts — and systems discard the signals that would allow this variability to be governed. CrazyGoldFish captures learning signals at the point of evaluation and structures them as persistent learning memory, achieving 93% reliability in total scoring alignment (ICC).

    Why is evaluation unreliable?

    Evaluation depends on human interpretation, which varies across evaluators, rubrics, and contexts. Small variations compound at scale into systemic divergence. No two evaluators reliably produce the same score for the same response without a governing signal layer.

    What are the three structural failures in evaluation?

    Inconsistency: outcomes vary across evaluators and time. Loss of information: detailed signals are discarded and only scores retained. Inability to reuse: scores cannot drive personalisation, remediation, or consistent decisions across systems.

    What happens when evaluation is unreliable at scale?

    Systems make decisions on unstable inputs. Personalisation becomes inconsistent. Programs cannot be audited or governed. Trust in the system itself erodes. This is not a measurement problem — it is a system reliability problem.

    How does CrazyGoldFish make evaluation reliable?

    CrazyGoldFish captures learning signals during evaluation and structures them as persistent learning memory. Reliability: 93% alignment with human scoring (ICC), 89.7% adjudicated correctness vs teacher 82.8% (+6.9%). Every response produces a LearningSignal — governed, auditable, and reusable.

    What is evaluation infrastructure?

    Evaluation infrastructure is the system layer that captures learning signals during assessment and stores them as persistent learning memory. Unlike grading tools that produce scores and stop, evaluation infrastructure converts every assessment into structured, reusable intelligence — making evaluation the control point where learning is interpreted and remembered.

    // book.demo

    Evaluation built to be trusted at scale

    See how CrazyGoldFish captures learning signals and makes evaluation reliable across your platform.