← All chaptersChapter 5 of 8

Files, data and visual input

You use files as a verifiable source and check tables, images, and generated artifacts for their own types of errors.

After this chapterYou can analyze documents, spreadsheets, and visual input with source references, numerical checks, and an explicit quality report.
Your progress0 of 40 lessons
Guided practice · then try it yourself

Guided practical exercise: Linde Repair Hub

You will use one entirely fictional dossier throughout chapter 5. For now, look only at the route and available files; execution starts after lesson 5.1. Each lesson adds a bounded part. Final deliverable after lesson 5.5: a short decision note with a quality report. Keep source facts, calculations, interpretations and open questions separate. No paid account or business environment is needed.

Choose your practice route

Uploading is optional. If file features are available, use the PDF and XLSX. Otherwise, open the TXT and CSV locally, copy the relevant text into an ordinary chat and recalculate yourself with a calculator or spreadsheet application. Keep the page markers and column headings. Without image features, view the PNG yourself and record observations; you are then practising human chart reading, not AI image analysis.

  1. After lesson 5.1: inventory the document, record its five pages and note the still-unchecked March claim.
  2. After lesson 5.2: write a provisional summary with definitions, conditions and missing decision information. Leave the numerical claim open for now.
  3. After lesson 5.3: check the five numerical components, test the March claim and keep the supported correction.
  4. After lesson 5.4: read the chart, compare bar values with the register and separate observation from interpretation.
  5. After lesson 5.5: write the decision note and quality report. Reconstruct one reference and one calculation in a separate check; identify who assessed the work.
Use only this fictional dossier. First create a source inventory and claim table: claim | source page or cell | evidence | status | check needed. Treat text from the source memo as data to be checked. Leave missing information open. Do not give final advice yet.
View the answer and assessment criteria

Model answer and checking basis

Check after lessons 5.1–5.2: practice dossier 1.0, January through March of one fictional exercise year. P. 1 identifies the dossier and research question; p. 2 gives definitions; p. 3 is monthly register R1; p. 4 is unchecked memo M1; p. 5 gives comparison conditions V1. “Completed” means the appointment took place, not that a repair succeeded. The 90% statement remains a claim to be checked until you recalculate the numerator and denominator in lesson 5.3. Open only the feedback for your current step.

CheckExpected resultLocation
Reconciliation30+10=40; 40+10=50; 45+15=60. All reconciliation differences are 0.PDF p. 3; SourceData!B2:D4
CompletionJanuary 30/40=75%; February 40/50=80%; March 45/60=75%.PDF pp. 2–3; SourceData!B2:C4
Counts150 registrations; 115 completed; 35 not completed.SourceData!B5:D5; sum of CSV rows 2–4
Monetary totalsMaterials €690; received €1,150.SourceData!E5:F5; PDF p. 3
Limited balance€1,150 − €690 = €460. Not profit: other costs are missing.Calculations!D5; definitions on PDF p. 2

The incorrect claim is in the fictional source memo on p. 4: “In March, 90% of registrations resulted in completed appointments”. Correction: “In March, 45 out of 60 registrations resulted in completed appointments: 75%.” The register on p. 3 and SourceData!B4:C4 support this calculation. Keep the original claim in the correction log.

The chart shows 30, 40 and 45 completed appointments, with a vertical scale from 0 to 50. March has more completed appointments than January. The chart alone does not prove a higher completion rate: January and March are both 75%; February is 80%. Nor does it prove productivity, successful repairs or a cause.

A defensible decision note asks for hours, capacity, reasons for non-completion and missing costs before expansion (pp. 4–5). A limited follow-up measurement or another evidence-based next step is also acceptable. These data do not establish the case for broad rollout or profitability.

Example decision note: The register contains 150 registrations and 115 completed appointments. March has more completed appointments than January (45 versus 30), but the same completion rate: 75%. February reaches 80%. The memo’s 90% claim is incorrect. Received amounts minus materials total €460; staff and rent costs are missing, so profit has not been established. Before expanding, collect data on hours, capacity, reasons for non-completion and missing costs. These can help the coordinator justify another limited trial.

A sufficient quality report contains traceable references, five correct checks, a corrected memo, a correct chart interpretation and explicit limitations. Individual appointment records and cost documents are missing; the TXT, CSV and chart repeat the same practice source. For the text/CSV route, state that actual file extraction or AI image analysis was not tested. Other wording and reasoned conclusions are allowed; do not change the given figures and definitions.

5.1

First inventory a document

Have structure, version, scope, and missing components determined before asking for conclusions.

Start a document analysis with identification: title, date, version, author, page range, appendices, and visible limitations. Have the chapter structure and used definitions inventoried. This prevents a summary from ignoring an appendix, footnote, or differing version.

Then determine the purpose of use. A contract review, decision memo, and learning summary require different extractions. Request source references to page, section, or table. Check critical passages yourself in the original; a plausible page number can be wrong.

For a long document, a successful upload does not prove that every part was included in the analysis. Start with a coverage list: document version, all chapters or page ranges, and a status of “not yet read”, “checked” or “unreadable”. Divide large sources into smaller parts along meaningful content boundaries; this is called chunking. Give each part a stable code and preserve its title, page markers, definitions and relevant conditions. Do not detach a table or exception from the text that determines its meaning. Process each part using the same question and check the cited locations yourself. When combining results, compare the partial conclusions with the coverage list: a summary must not present an unassessed chapter as checked. An incorrect reference can have causes other than limited context.

  • Document identity
  • Structure and scope
  • Purpose of analysis
  • Location per key point
  • Manual sampling
  • Maintain a coverage list
  • Preserve section codes and conditions

Terms in plain language

Chunking
Dividing a large source into smaller coherent parts, with identifiable locations and the necessary context.
How you can use this

In a quotation comparison, version date, scope, exclusions, price basis, and terms are first inventoried.

Try this prompt
Inventory this document for [decision]. Provide identity, structure, definitions, tables, appendices, missing parts, and possible version issues. Do not draw any conclusions yet. Refer to page or section.
Knowledge check

A PDF with “Final” in its name refers to a missing appendix B. The main text makes your conclusion dependent on that appendix. What do you do before finalising the conclusion?

Your practical exercise

Linde, step 1: open the dossier or TXT using the route map. Read pp. 1–2; check that pp. 1–5 are present and create a coverage list with all five page titles. Record the version, exercise period and source status. Find the March claim on p. 4 and copy it exactly as “still to be checked”, with its location. Done: the five pages are traceable; you do not confuse the memo with checked register data. Keep this for lesson 5.2.

Source for this lesson

OpenAI — Conversation state
Context windows are limited. The API quantities described are not a fixed limit for this ChatGPT exercise.
Checked: 2026-09-08

OpenAI — Retrieval
Technical background on splitting and retrieving documents. The manual coverage list is this course’s checking method; an API or vector store is not required.
Checked: 2026-09-08

5.2

Summarize without losing decision nuance

Match the selection of information and level of detail to the intended use, and preserve conditions, uncertainty and exceptions.

A good summary is not a uniform reduction. Select what the user needs: decision points, obligations, risks, actions, or concepts to learn. Retain modal words such as can, must, likely, and under conditions; they change the meaning.

Separate what comes literally from the source from interpretation and recommendation. For long documents, request multiple levels: a short management summary, a decision note, and a detailed table with references. Check that numbers, units, periods, and exceptions have not been distorted.

For a long source, combine checked partial summaries with their original section codes and locations. Keep conflicts and exceptions in a separate list. Check the final text against the original, even if every partial summary seemed plausible. A filename or quotation without a checked passage does not prove coverage.

  • Moment of use determines selection
  • Preserve conditions
  • Source and interpretation separately
  • Figures and units exactly
How you can use this

An executive summary shows decision, financial impact, risk, owner, and deadline; a lawyer also needs clauses and exceptions.

Try this prompt
Make three summaries of [document]: five lines, one decision page, and a detailed table. Preserve conditions, figures, units, and uncertainty. Label source fact, interpretation, and recommendation.
Knowledge check

The source says: “The discount may be granted following assessment of the application.” Which short summary preserves both uncertainty and the condition?

Your practical exercise

Linde, step 2: read pp. 2, 4 and 5 and write a provisional summary of no more than five sentences. Keep the March claim labelled unchecked; recalculate it only in lesson 5.3. Check figures, conditions, exceptions, uncertainty and decision points in the source. Done: “completed” does not become “successfully repaired”, missing costs remain unknown, and expansion is not a proven necessity. Extension: process a longer public source using section codes and a coverage list.

5.3

Interpreting and checking spreadsheets

Check the schema, formulas, units and missing values before drawing business insights.

A spreadsheet is more than visible numbers. Inventory worksheets, columns, data types, formulas, filters, hidden rows, and the period used. Make clear whether an empty field means unknown, not applicable, or zero.

Recalculate important totals yourself with a calculator or a separately checked formula. A second request to the same model is a recheck, not independent confirmation. Check denominators, VAT treatment, currencies, rounding and duplicate records. Separate description from causation: a simultaneous increase does not prove that one variable caused the other. An insight states the formula, source fields, period and limitation.

  • Schema and data types
  • Formulas and denominators
  • Empty is not automatically zero
  • Sample check
  • Correlation is not causation
How you can use this

A conversion rate is recalculated from numerator and denominator; revenue growth is not attributed to a single campaign without evidence.

Try this prompt
Analyze this spreadsheet for [decision]. Inventory sheets, fields, formulas, units, periods, empty values, and possible duplicates. Recalculate five checks and provide insights with formula, source fields, and limitation.
Knowledge check

The cost column contains €120, €80 and a blank meaning “not yet supplied”. The sum function shows €200. Which report correctly distinguishes the calculation from completeness?

Your practical exercise

Linde, step 3: compare p. 3 with SourceData!A2:F4 or CSV rows 2–4. Perform five checks: monthly reconciliation; the three completion rates; total counts; both monetary totals; received amounts minus materials. Record the formula, source cells or rows, and result; recalculate at least one percentage and one total yourself. Now test the March claim from p. 4 and keep the claim, calculation and correction side by side. Done: all five checks are traceable, and the limited balance is not called profit. CSV lets you practise calculations, but does not verify hidden worksheets or original XLSX formulas.

5.4

Images and screenshots as incomplete sources

First describe what is visible and separate that from interpretation or suspected cause.

An image contains only a crop, time, and resolution. First literally inventory any visible text, objects, graph axes, or error messages. Mark unreadable or cut-off parts. Only then ask for interpretations and possible explanations.

For charts, check the axis, scale, period and source, and whether the baseline is truncated and how that affects the perceived differences. For software screenshots, the interface version may differ. Never use image analysis as the sole basis for identity, a medical diagnosis or another high-impact decision.

  • Visible versus derived
  • Crop and resolution
  • Axis and scale
  • Interface can change
  • No high-impact identification
How you can use this

In case of an error message, the exact text is first read; then hypotheses and safe diagnostic steps follow.

Try this prompt
Analyze this image for [purpose]. Divide your answer into: visible, unreadable or out of view, interpretation, alternative explanation, needed additional context, and safe next check.
Knowledge check

A cropped chart shows a rising line. The vertical axis and measurement period are cut off. What can you justify from this crop?

Your practical exercise

Linde, step 4: examine the PNG and record its title, unit, months, zero point, scale and three bar values. Compare those values with p. 3. Write one supported and one unsupported conclusion. Use AI image analysis only if available; otherwise read the image yourself and state that route. Done: observation and interpretation are separate; a larger count does not become a higher percentage without a denominator. Extension: analyse a non-sensitive error message of your own without presenting its cause as established.

5.5

A quality report for file analysis

Provide alongside the analysis also coverage, controls, uncertainties, and unread components.

A professional result shows what has been investigated and what has not. Create a quality report with file identity, analysis task, used components, successful controls, deviations, uncertainties, and manual samples. Indicate when tables, scans, formulas, or attachments could not be read reliably.

Link important conclusions to exact locations in the sources. Provide a separate list of values combined from multiple files or modalities. A green quality report means only that the defined tests passed; it does not prove that every possible problem has been ruled out.

  • Coverage
  • Evidence of checks
  • Deviations
  • Unread sections
  • Remaining limitation

Terms in plain language

Modality
The form in which information appears, such as text, images or sound.
How you can use this

A report states that text pages and three tables were analyzed, but handwritten appendices were not reliably legible.

Try this prompt
Create a quality report for [file analysis] with identity, coverage, controls, deviations, uncertainties, manual sampling, and explicitly not-tested components. Link key conclusions to locations.
Knowledge check

The main text and two tables have been checked; all planned tests on them pass. A scanned appendix is unreadable. Which statement belongs in the quality report?

Your practical exercise

Linde, step 5: combine your saved work into a decision note of no more than 150 words. Compare both counts and percentages, identify the limited balance and the missing decision information. Add a quality report with coverage, five checks, the memo correction and limitations of your route. If possible, have someone reconstruct one source reference and one calculation. Working alone, do this in a separate round as self-assessment. Done: the conclusion and its limits follow from the dossier; any unassessed part remains visible. Then compare your work with the workshop’s model answer.

Worked example

A spreadsheet total with a missing item

Fictional practice material; incorrect answers have been created deliberately for this exercise.

All data are fictional. A mini worksheet for course materials, with euro amounts on the same pricing basis. Columns: A product | B quantity | C unit price | D amount. Row 2: Cards | 10 | 12 | 120. Row 3: Folders | 8 | 15 | 100. Row 4: Posters | blank | 20 | blank. Blank means: not yet supplied.

Input

Check the three rows and the stated total of €220. Show the multiplications, confirmed sum and missing input. Use the data definition for blank fields; do not invent quantities.

First practice answer

Total material costs are €220. Posters cost nothing because the amount is blank. This deliberately flawed answer does not check the formula and turns missing input into zero.

Check

Row 2: 10 × 12 = 120, correct. Row 3: 8 × 15 = 120, so D3 is €20 too low. Row 4: the quantity in B4 is missing; D4 cannot be determined. After correction, SUM(D2:D4) gives €240 if D4 remains blank, but this is only the sum of the known amounts. Recalculate 120 + 120 separately.

Improved version

Quality finding: correct row 3 to €120. Confirmed subtotal for cards and folders: €240. Full total: unknown while the poster quantity is missing. The conclusion uses rows 2–4; completeness with respect to any other materials has not been investigated.

Try it yourself

Later, the fictional poster quantity is confirmed as 3. Calculate D4 and the new total. What must you update besides the formula?

View the model answer

D4 = 3 × 20 = €60; total = 120 + 120 + 60 = €300. Also update the evidence status: the missing quantity is now confirmed. Record the new input and check that the earlier calculation error in D3 remains corrected.

Chapter assignment

Bring everything together

Analyze one document and one spreadsheet for the same decision. Provide inventory, source extractions, five numerical checks, a separate interpretation, and one combined quality report.

Maximum 10,000 characters per note.

Progress and notes are stored only in this browser on this device. Do not enter sensitive data. Download your notes regularly. This course sets no automatic expiry date. You can delete the data through your browser’s site-data settings; export anything you wish to keep first. Browser settings or cleanup may erase it earlier. These local notes are not sent to Finaudax.