Data are captured, documented, stored, transformed and used for a purpose. They are then retained or deleted according to that purpose and usage conditions. Metadata describe origin, authorship, date, fields, units, licence or permission and transformations. An identifier distinguishes an observation but does not itself explain how it was obtained or what it represents. A context-free table may contain readable numbers yet remain impossible to interpret correctly.

The audit uses supplied, reusable teaching HTML stored locally, without collecting people’s data or extracting external pages. An observation code is enough to follow a row; adding classmates’ names introduces unnecessary information. The data dictionary defines every field, type, scale and permitted missing value. Traceability connects original, export, cleaned version, chart and model. Keeping everything indefinitely is not an automatic choice either: state which files enable reproduction and when unnecessary copies will be removed. This policy belongs to the design before training.

HTML organises content through tags. In a table, table delimits the whole structure, tr represents a row, th a header cell and td a data cell. Appearance alone does not identify this structure: styles can change its appearance without changing data, and different structures can look similar. The extractor needs to recognise the target table and check headers, column count and content correspondence. Finding a table completes only the first step because it may still be a different table from the intended one.

HTMLParser reports events as it encounters opening tags, text and closing tags. To turn them into a data table, the program retains state: whether it is inside the selected table, which row it is collecting and which cell is active. This context determines where each fragment belongs. A cell may contain several pieces of text, for example when another tag occurs inside it; these are joined before its contents are stored. Even when the result contains digits, it remains text. Locating the cell does not yet establish that all its contents have been preserved.

How each row is completed also affects extraction. If it is stored only when the next row begins, the last row can be lost because no further row opening follows. Checking closing events, counts and identifiers reveals that loss and supports comparison with the source. HTMLParser does not automatically validate the document’s complete correctness. This work uses a local file with a known structure: explaining that condition clarifies what the extractor checks and why it is not presented as a universal solution for every page.

Scraping extracts information from a representation such as HTML; an API provides a defined interface for exchanging data. The former depends on tags and structure; the latter uses a contract of fields, operations and responses. Neither automatically guarantees quality, permission or freedom from changes. A public page does not automatically permit every collection and reuse. This course processes only the supplied local file without automated requests to external sites.

CSV preserves rows and fields using separators, quoting and encoding that must be interpreted compatibly. Python’s csv library manages details that manually joining commas can break. The format neither defines meaning nor adds units: that information remains in the dictionary. The extractor checks the intended table, headers and columns before exporting. Then compare source and result, including problematic and final records. Opening CSV without errors only proves that a tool could interpret it, not that every field corresponds to the correct observation.

Cleaning prepares data for a purpose through justified, recorded decisions. Preserve the original and work on a copy. An exact duplicate of the same observation may be removed after checking origin; independent observations with equal values are not necessarily duplicates. A repeated identifier with different content requires investigation, not arbitrary deletion. A change log reconstructs inclusions, exclusions and conversions.

A missing value is not zero: replacing it invents an observation and changes averages or learning. If excluded from a calculation, the denominator must reflect available values. Check numeric format, range and unit separately; successful conversion does not guarantee domain membership. A Google Sheets chart needs title, axes, scale or unit and coherent data selection. A bar or point summarises a record but proves neither its origin nor causality. Compare the chart with the cleaned table and explain excluded rows so visualisation does not hide previous decisions.

A classifier assigns a class from input data. This unit’s model uses a reading and a boundary: below the threshold it predicts aviso, otherwise normal. Learning selects that threshold using examples with known labels rather than writing each observation’s answer. Labels come from the dataset’s teaching criterion, not an automatically discovered truth about the classroom.

The program compares candidates between training values and counts matches with labels. It retains the selected candidate and applies the rule to new values without readjusting during prediction. If candidates tie, document the program’s policy; data leaving an example-free gap cannot establish a unique boundary. Equality depends on the operator and must be tested. This model makes each decision inspectable but explains neither physical causes nor all AI. The approved object alternative retains provenance, labels and independent testing; it does not automatically replace the local extraction required by the project.

Training selects the threshold from known examples; evaluation checks how it works on observations that did not participate in that choice. Training and test data are therefore separated before fitting. Independence requires checking identifiers and provenance: different IDs can refer to copies of the same example or nearly identical captures. When that information appears in both sets, results can look better without establishing equivalent ability on new data. This is information leakage, which changes what the test allows us to conclude.

Independence also depends on how results are used. If we repeatedly inspect the test to choose changes, those examples already influence the design. The record must distinguish the set used for improvement from a new evaluation. In the latter, the matrix crosses reference class with prediction: a false alert predicts aviso when the reference is normal; an omission predicts normal when aviso was required. Separating these errors reveals which decisions the model gets wrong and allows their consequences to be considered in relation to use.

The percentage of correct predictions is obtained by dividing correct predictions by the total number of evaluated cases and expressing the ratio as a percentage; an explicit denominator is necessary. An overall figure can hide errors in an infrequent class. Per-class results, sample size and representativeness complete the interpretation: getting every case in a small sample right does not guarantee performance in other contexts. Conclusions must match the data and procedure actually used. This school exercise supports neither medical decisions nor important decisions about people.

AlphaGo combines networks guiding decisions and evaluating positions with search in a game with defined rules. The original research used expert games and learning through self-play. This procedure differs from our threshold practice: it supports comparing task, data, objective and evaluation without confusing mechanisms. Success at Go establishes neither general capability for every problem nor human intention or understanding.

The audiovisual block takes 3.5 of the eleven sessions: 0.25 preparation, 2.25 authorised viewing or a coherent teacher selection and 1 comparison and transfer. Before viewing, formulate a prediction and a question; during it, describe three observations without transcribing dialogue; afterwards, separate observation, interpretation and a checkable claim. A technical source can confirm a claim or leave it unresolved when evidence is insufficient. The written alternative addresses the same problem through sequence descriptions and guided reading, without requiring viewing outside class. Turn the conclusion into an A01 requirement and a testable application in A02.

An audit connects stages to check how a result was obtained. A row must be traceable from HTML to CSV, cleaning, chart, assigned set, prediction and final retention decision. Each connection needs an appropriate test. A plausible chart may not reveal a lost last row; an error-free model run does not guarantee independent separation. Distinguish extraction integrity, dataset quality and classifier performance.

The laboratory uses extraer_tabla.py and modelo_umbral.py in Thonny with local files and preserved versions. A01 includes executed extraction, comparison of a row and initial evaluation with separate training and test data; it is not just a future plan. A02 extends comparison to records 1, 4 and last, documented cleaning, chart, matrix and verified improvement with new tests. Analyse a possible selection or coverage bias and explain which conclusions it limits. Each partner follows a datum, predicts a boundary and explains a change. Editable files and the log support reviewing causes without inventing data or claiming general capabilities.