{"slug": "an-evaluation-score-is-only-as-useful-as-its-target", "title": "An Evaluation Score Is Only as Useful as Its Target", "summary": "A developer building a document-evidence assistant found that a 90% agreement score (36 of 40 observations) masked a critical failure: the model missed the single reference labeled as documented evidence while reproducing all 33 \"not located\" labels. The prototype, evaluated only on its own training observations against AI-generated references, had no explicit-absence examples and correlated observations drawn from just ten documents, prompting the developer to treat target definition and document-level splitting as part of system design.", "body_md": "My small document-evidence model agreed with its reference labels on 36 of 40 observations. The headline was 90%.\n\nThen I inspected the confusion matrix. It missed the only reference labeled as documented evidence.\n\nThat error mattered more to the intended workflow than the headline suggested. I was building an assistant to help someone locate and check evidence in a document. Reproducing the dominant “not found” state while missing the sole positive example did not establish that the assistant was useful.\n\nThis was a local prototype, evaluated on its own training observations against AI-origin references. It was not an independent test, a measure of candidate ability, or evidence of improved hiring outcomes. Those qualifications change what the score means.\n\nThe exercise made me treat the target definition as part of the system design.\n\nThe prototype assigned a documentary state to each of four dimensions in a document. Its four possible states were documented evidence, not located, explicit absence, and ambiguous.\n\nEach state needs a precise interpretation. “Not located” describes what the available document and process revealed. It does not establish that a person lacks a skill. A document can omit something the person knows. Extraction can fail. A relevant passage can use unexpected language.\n\n“Documented” also requires care. A statement can mention a tool without describing completed work. An intention to learn is different from evidence of practice. Conflicting passages may require a reviewer rather than a forced positive or negative label.\n\nThe references for this run came from AI, without a new independent human annotation. Agreement with them measured reproduction of those references. Calling that number “accuracy” without naming the target would invite a much broader interpretation than the experiment supported.\n\nThe forty observations came from ten documents, with four dimensions per document. Their reference distribution was highly uneven:\n\n| Reference state | Observations | \n|---|---|\n| Documented evidence | 1 | \n| Not located | 33 | \n| Explicit absence | 0 | \n| Ambiguous | 6 | \n\nThe model reproduced all 33 “not located” references and three of the six ambiguous references. It predicted “not located” for the remaining three ambiguous observations and for the one documented-evidence observation.\n\nThose outcomes explain the 36/40 agreement. They also show why the score could conceal a failure in the capability I wanted to investigate.\n\nThere were no explicit-absence references. A reporting convention can include that class in a four-class macro-F1 calculation, but its numerical contribution does not become measured performance on that state. The class had no evaluation support.\n\nI want a report to make unsupported capabilities visible. A reader should be able to distinguish “the method failed on these examples” from “this dataset did not test the capability.”\n\nThe four dimensions within one document share text, formatting, extraction quality, and source context. That creates correlation between observations.\n\nIt would be misleading to describe this as forty independent cases. It would also be risky to split the rows randomly and allow dimensions from one document to appear in both training and evaluation. A model could encounter much of the same underlying input on both sides.\n\nFor a future document-level evaluation, I would group observations by source document before assigning splits. Closely related versions and duplicated documents need the same treatment. The unit of separation should reflect how information can cross the boundary, rather than whichever unit makes the table easiest to partition.\n\nThis particular run had an even simpler limitation: all forty observations were training observations. It demonstrated technical learning and exposed errors. It did not measure performance on unseen documents.\n\nAn earlier, separate benchmark compared a TF-IDF logistic-regression baseline with a small PyTorch model on a frozen document split. Its labels were derived from a lexical rubric, not from independently observed outcomes.\n\nAcross five additional seeds on that same split, the baseline's test macro-F1 stayed at approximately 0.902. The neural model ranged from approximately 0.576 to 0.902.\n\nThat was a useful diagnostic. The baseline was stable under that setup, while the neural model showed substantial sensitivity to initialization and optimization. Reporting only the best neural run would have hidden that sensitivity.\n\nHowever, the baseline also shared vocabulary with the labeling rubric. A high score could reflect successful reproduction of the proxy. It did not establish the quality of an operational decision.\n\nRepeated seeds examine one source of variability. They do not produce new independent documents, repair a weak target, or make an already inspected test set fresh again. I keep those historical results separate from the forty-observation prototype because they answer different questions.\n\nThe next data problem was semantic compatibility. A larger synthetic corpus used eight graded criteria and levels from zero to four. The documentary prototype used four dimensions and four evidence states.\n\nThese are different targets. Concatenating their rows would not make them equivalent.\n\nConsider this entirely invented passage:\n\n```\nI have explored a video-editing tool through tutorials and would\nlike to use it in a future team project.\n```\n\nA generated grading rubric might award a low level for exposure to the tool. A documentary rule might classify the passage as insufficient evidence of completed editing work. Neither label automatically specifies an operational decision.\n\nThe mapping needs an explicit definition, reviewed examples, and a record of the original target. Ambiguous or incompatible cases should remain visible. A score threshold is not a substitute for deciding what the label means.\n\nA later local experiment preserved the synthetic corpus's own eight-criterion target rather than silently converting it into the earlier four-state task. That preserves an important boundary: a result on the synthetic task remains a result on that task.\n\nSynthetic data makes it easier to construct coverage and contrastive examples. It also makes shortcuts easy to repeat.\n\nTwo documents can have different identifiers, names, section orders, and small wording changes while retaining the same underlying phrase template. An exact-duplicate check can pass even when the train and test examples remain linguistically very similar.\n\nIn the corpus I inspected, variants reused phrase bases with added framing. That is a reason to investigate template overlap, not evidence that every result is invalid or that all leakage has been removed.\n\nFor a stronger evaluation, I would track template families, inspect similarity across splits, and reserve genuinely different evidence formulations. I would also check whether length, section structure, or stylistic cues reveal the generated label. An unseen record is only useful to the extent that it tests the intended difficulty.\n\nThe product question is whether the suggestion helps a reviewer find and assess evidence.\n\nThe next useful comparison would be document plus search versus document plus search plus a model suggestion. It would need independently reviewed references, positive and ambiguous examples, missing-evidence cases, and explicit handling of extraction failures.\n\nI would measure missed evidence, misleading suggestions, corrections, abstentions, and review effort. The model suggestion and retrieved passage should remain separately inspectable: the search result is not automatically an explanation of the neural output.\n\nReference review should also avoid showing the model's suggestion first. Otherwise, the evaluation can inherit the model's framing. Disagreements need resolution or a recorded pending state, rather than an invented consensus.\n\nThat study has not been completed. It is the next evidence needed to judge usefulness, rather than another interpretation of the existing training score.\n\nBefore comparing methods, I write down:\n\nThe 90% headline was useful because inspecting it changed my next step. The system needed better reference coverage and an independent workflow comparison before a claim of utility.\n\nWhat does the target in your current evaluation actually represent?", "url": "https://wpnews.pro/news/an-evaluation-score-is-only-as-useful-as-its-target", "canonical_source": "https://dev.to/rkrisa/an-evaluation-score-is-only-as-useful-as-its-target-4kia", "published_at": "2026-10-16 12:09:52+00:00", "updated_at": "2026-10-16 12:14:56.103965+00:00", "lang": "en", "topics": ["machine-learning", "natural-language-processing", "ai-tools"], "entities": [], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/an-evaluation-score-is-only-as-useful-as-its-target", "markdown": "https://wpnews.pro/news/an-evaluation-score-is-only-as-useful-as-its-target.md", "text": "https://wpnews.pro/news/an-evaluation-score-is-only-as-useful-as-its-target.txt", "jsonld": "https://wpnews.pro/news/an-evaluation-score-is-only-as-useful-as-its-target.jsonld"}}