Guides · Process data

From paper batch records to CPV data, without breaking data integrity

Let the CPV plan decide which values to take, use the master batch record as the map, keep each value as written beside its normalised form and its page, have copied values checked by a second person or a validated system, record what cannot be read, and version the dataset.

Many continued process verification (CPV) programmes still start with paper. The values that matter are written by hand in executed batch records, and somebody copies them into a spreadsheet or a database before any trend chart exists. That copy is a new record. It needs the same care as the batch record it came from: it must say who made it, keep a link to the original page, and be checked.

In brief: the CPV plan decides which values to take, the master batch record tells you where each one is written, every value is kept as written beside its normalised form, every copied value is verified, values that cannot be read are recorded as such, and the dataset is versioned. The rest of this guide explains each step and the regulatory text behind it. A CSV template with the columns described below and two fictional example rows is free to download.

Why does this copy matter to a regulator?

Because the dataset is what your process decisions rest on. FDA’s process validation guidance describes stage 3, continued process verification, as gaining “ongoing assurance … during routine production that the process remains in a state of control”. It ties this to 21 CFR 211.180(e) and says the data “should be statistically trended and reviewed by trained personnel”, with the quality unit reviewing the information.

In the EU, Annex 15 (sections 5.28 to 5.32) calls the same activity ongoing process verification. It asks manufacturers to monitor product quality with “the relevant process trends evaluated”, under an approved protocol or equivalent document, with a report and statistical tools where appropriate. ICH Q10 section 3.2.1 asks for a process performance and product quality monitoring system that analyses the parameters and attributes in the control strategy.

The annual reviews draw on the same data. 21 CFR 211.180(e) requires records to be kept so that their data can be used to evaluate each product’s quality standards at least annually, including a review of a representative number of batches. EU GMP Chapter 1, section 1.10, lists “a review of critical in-process controls and finished product results” in the product quality review.

If a value is copied wrongly, the trend is wrong, and so is any conclusion drawn from it. Data integrity guidance treats such a copy as data in its own right. The MHRA guidance (section 6.11.2) says: “Where manual transcriptions occur, these should be verified by a second person or validated system.”

Which values should you take?

The CPV plan decides, not the person doing the copying. FDA recommends that “a statistician or person with adequate training in statistical process control techniques develop the data collection plan”, and Annex 15 expects an approved protocol. Practice varies a great deal between companies, and your quality system decides the format, but the selection usually comes from the control strategy and the process risk assessment.

Values commonly taken from executed batch records:

  • Critical process parameters (CPPs) and other monitored parameters: temperatures, times, speeds, pressures, pH, set points and actual readings.
  • In-process controls (IPCs): weights, hardness, loss on drying, fill volumes, bioburden sample results recorded on the record.
  • Yields and reconciliations at each stage. FDA lists “process yield variations” and “batch records” among the sources that can reveal variation.
  • Times and durations: hold times, start and end of steps, time out of refrigeration. Many are derived from two written clock times, so you need both.
  • Context values that explain the others: equipment ID, raw material lot numbers, the MBR version.

Write the list down per product, with each parameter’s name, unit, step and limits. Add values later through a controlled change to the plan, not by quietly adding columns. A list fixed in advance also protects you against the suspicion that values were picked after looking at the results.

How do you find each value on the record?

Use the approved master batch record (MBR) as the map. For each parameter in the plan, note where it sits on the blank MBR: page, section, table, row and column, and the limits printed beside it. That map does three jobs:

  1. It tells the person copying exactly which entry to take, so two people take the same one.
  2. It gives each limit a source. Note whether a printed range is a normal operating range (NOR), a proven acceptable range (PAR), a specification or an alert limit; the NOR and PAR guide explains the difference.
  3. It tells you when the map has changed. A new MBR version can move a row, add a decimal place or change a limit. Record the MBR version with every batch so a value is always read against the form it was written on.

What goes wrong when values are copied by hand?

The same errors come back on every site. Knowing them tells the person verifying what to look for.

Error What it looks like Why it gets through
Decimal 1,8 typed as 18; 0.25 typed as 2.5 Both look plausible in a column of numbers
Unit 2.4 kg kept when the column is in g The number itself is right
Date 03/04 read as 3 April when written as 4 March Day-first and month-first records mix on global supply chains
Wrong row or column The 30-minute reading taken from the 60-minute row Adjacent rows hold similar numbers
Wrong batch or page A value from the previous batch’s record Records look alike
Correction The struck-out value copied instead of the new one Both are legible on the page
Illegible entry A guessed digit The person copying wants a complete row

A value that is wrong but still inside its limits is the hardest case. It produces no out-of-trend signal, so nothing downstream will catch it. Only a check against the original page will.

How should the copy be verified?

Every transcribed value should be checked against the original. The MHRA guidance asks for verification “by a second person or validated system”; PIC/S PI 041-1 (section 9.7) and Annex 11 (section 6) ask it for critical data; and 21 CFR 211.68(b) sets the degree and frequency of the check by the system’s complexity and reliability. Verifying transcribed data sets out what each text says and compares the methods: a side-by-side check, independent double entry, or software reading with human review.

How you verify is a choice for your quality system. Whatever the method, record who verified each value and when, beside the value. If software reads the values, ask how it decides what goes to a person: a reading that is wrong, sure of itself and within limits may never be shown to anyone, so the tool’s error rate on your forms needs measuring, for example by spot checks.

What does ALCOA+ mean for a transcribed dataset?

The MHRA guidance (section 6.1) applies ALCOA+ to all data, and a dataset built from batch records is data in its own right. In practice: each value carries who copied it, who verified it and when; it is stored as written, character for character; the copy records when it was made as well as when the value was written in the record; the executed record, or its verified true copy, stays the original and the dataset points to it; and every planned value for every batch has a row, including the ones that could not be read. ALCOA+ for batch records takes each principle in turn.

How should the dataset be structured?

One row per value, not one row per batch. A wide sheet with a column per parameter is easy to chart but loses the information that makes each value traceable. A long table keeps it, and a pivot gives you the wide view when you need it.

Column What goes in it
dataset_version The version of the dataset this row belongs to
product, batch Product code and batch number as on the record
mbr_version The MBR version the record was printed from
source_document, source_sha256 The scan’s file name and its hash, so the exact file can be found and shown unchanged
page, place_on_page Page number and the row, column or field label where the value is written
step, parameter, parameter_class Process step, parameter name from the CPV plan, and class (CPP, IPC, yield, time, context)
value_as_written Exactly what is on the page, including a decimal comma or a written unit
normalised_value, unit The number in the plan’s unit and decimal format, empty if the value cannot be read
lower_limit, upper_limit, limit_kind The printed limits and their kind (NOR, PAR, specification, alert)
recorded_at The date and time written beside the value in the record
status For example: verified, unreadable, corrected on page, query open
transcribed_by, transcribed_on Who copied the value, and when
verified_by, verified_on Who verified it against the page, and when
comment The reason for any status other than verified, or a query reference

Keeping value_as_written beside normalised_value matters. Normalising (decimal commas to points, kg to g, a two-digit year to four) is data processing, and the MHRA guidance (6.9) expects processing to be traceable to the raw data. If the two disagree, a reviewer can see why.

A worked example

A fictional example, with a fictional product and batch, showing a subset of the columns:

batch page step parameter value_as_written normalised_value unit limits (kind) status verified_by
EX-0001 7 Drying Inlet air temperature 62 62 °C 55 to 65 (NOR) verified A. Example, 2026-03-20
EX-0001 8 Drying Loss on drying at end of drying 1,8 % 1.8 % w/w ≤ 2.5 (IPC limit) verified A. Example, 2026-03-20
EX-0001 12 Compression Tablet hardness, 30 min (illegible) N 80 to 120 (NOR) unreadable A. Example, 2026-03-20; query Q-EX-01

The second row shows why the value as written is kept: the decimal comma is on the page, the point is in the dataset, and both are visible. The third row shows an unreadable value recorded as such, with no number, and a query raised.

What do you do with a value you cannot read?

Do not guess, and do not drop the row. Record it as unreadable, leave the normalised value empty, and raise a query with production or QA, who can check the paper original or ask the person who wrote it. If the original is illegible too, that is a finding about the batch record itself, handled through your deviation process, not something the dataset can fix.

Exclusion follows the same logic. The MHRA guidance (6.10) says data may be excluded only with a documented, valid scientific justification, and that “all data (even if excluded) should be retained with the original data set”. In a CPV dataset, a value left out of a chart stays in the table with its reason.

How do you version the dataset?

Treat the dataset as a controlled record.

  • Freeze a version for every CPV report, APR or PQR, and investigation that uses it, and name the version in the report.
  • Never overwrite a verified value. Correct it by a new version or an audited change that keeps the old value, the new value, who changed it, when and why.
  • Keep the dataset where changes are recorded: a validated system with an audit trail, or a controlled file with a change log if your quality system accepts that for the risk.
  • Re-verify when you change a normalisation rule, because every value it touches changes with it.

A report built from version 3 must still be reproducible after version 4 exists.

What might an inspector ask?

Expect questions that follow the guidance above. Prepare an answer and the evidence for each:

  • Where is the plan that says which values are collected, and who approved it?
  • Pick one point on this chart: show me the page it came from.
  • Who copied this value, who verified it, and when?
  • How was the copy verified, and where is the procedure?
  • If software read the records, how was it qualified for this use, and how do you know how often it is wrong on these forms?
  • What happened to values that could not be read, and where are the excluded points?
  • Which version of the dataset does this report use, and what changed since?
  • How do you know the scan you checked against is a true copy of the paper record?

If you can answer the second question for any point in seconds, most of the others follow. For more on checking the copy itself, see verifying transcribed data; for the annual review, see APR and PQR data from batch records.

A checklist for Monday

  • The CPV plan lists each parameter with its unit, step, limits and limit kind.
  • Each parameter is located on the current MBR, and MBR versions are tracked.
  • Scans of executed records are kept as true copies, with their hashes.
  • The dataset is one row per value, with value as written and normalised value.
  • Every value has a transcriber, a verifier and dates.
  • Unreadable values have a row, a status and a query.
  • Each report names the dataset version it used.

Sources

  1. FDA, Process Validation: General Principles and Practices (January 2011), accessed 18 November 2026.
  2. EudraLex Volume 4, Annex 15: Qualification and Validation (in operation since 1 October 2015), accessed 18 November 2026.
  3. ICH Q10 Pharmaceutical Quality System (Step 4, 4 June 2008), accessed 18 November 2026.
  4. EudraLex Volume 4, Part I, Chapter 1: Pharmaceutical Quality System (in operation since 31 January 2013), accessed 18 November 2026.
  5. 21 CFR 211.180 General requirements (records and reports), accessed 18 November 2026.
  6. 21 CFR 211.68 Automatic, mechanical, and electronic equipment, accessed 18 November 2026.
  7. MHRA, GXP Data Integrity Guidance and Definitions, Revision 1 (March 2018), accessed 18 November 2026.
  8. PIC/S PI 041-1, Good Practices for Data Management and Integrity in Regulated GMP/GDP Environments (1 July 2021), accessed 18 November 2026.

General information, not regulatory advice. Check the current text of each regulation; your quality system decides how it applies to you.

Related guides