Preclinical study report writing, explained.

A GLP final report is the one document that speaks for a finished laboratory study

From the last study day to a signed GLP report, in five steps

1. The study ends and specialists finish their workVLT-TOX-028's recovery animals complete their period on September 29, 2026. Pathology and toxicokinetics reports follow.
2. The data package reaches the study directorOn October 5: the signed plan and amendment, data exports, one signed contributor report and the report template.
3. The report is written and checked against the recordsTables, results, summary and conclusion, with every number traced. Missing values and missing data are chased here.
4. Quality assurance audits, and its statement is preparedQA compares the report with the raw data, reports what it found and signs a statement of its inspection dates.
5. The study director signs, and the records go to the archiveA later change can only be made by a signed amendment that says what changed and why.

Step 3 is the subject of this guide, with its hand-offs to steps 4 and 5. The study, people and figures in this example are fictional.

What surprises people new to GLP reports: a one-page rule, a list of dates, exports and separate signatures

Study director, raw data, QA statement and amendment defined

Fourteen terms are defined in the one-page summary.

The 14 items in 21 CFR 58.185, and where a study director gets each one

In shortFDA's GLP rule names 14 things a final report "shall include", and leaves the order and the headings to you. Four come from the approved protocol and six from facility and test article records. Two are written by the study director from the data, and two are other people's signed documents. A report can follow every heading in a template and still miss one.

The regulation opens with a floor: the report "shall include, but not necessarily be limited to" the list.2 It gives no section numbers and no required order. A house template with nine sections and a template with forty can both carry all 14 items, and each item needs checking on its own. That is why an audit against the template and an audit against the rule can find different things.

The last two columns are our reading and a common pattern. The rule names neither sources nor sections.

What the protocol has to carry before the report can

Those four items are only as good as the plan. 21 CFR 58.120 requires an approved written protocol listing, among other things, the purpose, the test system, each dosage level, the type and frequency of tests, the records to be kept and the proposed statistical methods. Any change "shall be documented, signed by the study director, dated, and maintained with the protocol".9 A protocol amendment that nobody carried into the report is one way for the report to disagree with the file.

What the report leaves out by design

Writers sometimes ask how much to reproduce. A report summarizes, and the raw data stays in the archive. OECD asks for "a summary of results" and "all information and data required by the study plan".4 FDA's IND rule separately asks for "a full tabulation of data suitable for detailed review" for each toxicology study meant to support a clinical trial, so a report headed for an IND may carry individual animal tables as appendices.10

FDA Part 58, EPA and OECD GLP: what each says about the final report, November 2026

In shortFDA's Part 58 dates from December 1978 and OECD's Principles from 1997, and the report paragraphs have barely moved. FDA's 2016 rewrite is still only a proposal. The two rulebooks agree on most things and differ on vocabulary, on what the QA statement says and on where the compliance statement lives.

Protocol or study plan, test article or test item: FDA and OECD side by side

FDA's own comparison page, current April 7, 2015, sets these texts side by side with EPA's GLP rules for pesticides and chemicals. EPA's text says that modifying a final report to meet EPA's submission requirements is no correction, addition or amendment.14

Other rulebooks follow the same pattern. In the European Union, Directives 2004/9/EC and 2004/10/EC apply. National authorities inspect the test facilities, and the European Medicines Agency (EMA) coordinates these inspections for studies in its centralized procedure.15 The UK has its own 1999 regulations.16 Swiss authorities report that the EU is evaluating its GLP Directive in a 2025 to 2026 project. We found no published outcome.17 A report written to both FDA and OECD can take the stricter reading of each row above.

Why a number in a GLP report has to lead back to an original record, through every copy

In shortFDA defines raw data by what the report needs: the original records "necessary for the reconstruction and evaluation" of it. Between that record and a sentence, a number may pass through a balance file, a system export, a spreadsheet, a table and a summary. Each copy is a place to go wrong, and each needs an owner. Decide in writing what your facility treats as raw data before you trace anything to it.

FDA defines raw data as worksheets, records, notes "or exact copies thereof" that result from original observations. It can include "recorded data from automated instruments" and computer media.8 OECD says "original test facility records and documentation, or verified copies".4 Both let a copy stand in for an original, and both require that it really is one.

That carries a trap for anyone tracing a report to a spreadsheet. In its 1981 questions and answers, FDA was asked whether a computer printout derived from data transferred to computer media from laboratory data sheets is raw data. The answer was "No."5

Of computerized data-acquisition systems that meet its criteria, the same document says "either the magnetic media or hardcopy printouts are considered to be raw data". It also says that if raw data are transcribed and sent to a sponsor for formatting, QA should assure the computer-formatted data accurately reflect them.5

So tracing a report number to an export cell proves the report matches the export. Showing the export matches the original record takes a further step, and your procedures have to say who takes it.

The ALCOA words behind a traceable GLP number

Data integrity guidance across regulated work rests on the letters ALCOA: attributable, legible, contemporaneous, original and accurate. The UK regulator, the MHRA, adds that the "+" sometimes used refers to complete, consistent, enduring and available, and says there is "no difference in expectations".19 OECD's advisory document on GLP data integrity applies the same thinking to studies. For a report writer, three clauses matter most:

How long the records behind a final report are kept

Part 58 ties retention to the filing: at least two years after a marketing application is approved, at least five years after results are submitted in support of an application, or at least two years after the study ends if it is never submitted, whichever is shortest.20 Records go to an indexed archive with a named person responsible. The report must say where they are.2,20 Other authorities set their own periods, so a global program plans for the longest it faces.

Contributor reports, QA statement, compliance statement and study director signature: what each signature stands for

In shortA GLP report carries several signatures, and people mix them up. Specialists sign their own reports. Quality assurance signs a statement about its inspections. The study director signs the final report and accepts responsibility for the data. Each signature means something different, and a change after signing has to be made as an amendment.

Why FDA keeps specialists' reports separate from the integrated report

When FDA asked for comments on modernizing Part 58, many respondents suggested one integrated final report in place of the individual scientists' reports. In its 2016 proposal FDA recalled that individual reports exist "to ensure the findings of the individual scientists are accurately reflected", said merged reports "would obscure the individual scientist's accountability" and wrote: "We continue to affirm these statements." In the same passage it said: "we support processes used for the efficient review of the draft study report."6 Faster drafting fits that. Removing the contributor's signed report does not.

What a QA audit of the final report looks for, in OECD's words

OECD's consensus document on quality assurance lists five questions for the report audit:

QA must be told of all changes made during the audit. Before signing, it should confirm that "no changes to the report have been made which would require a further audit".4

QA is never the author. FDA requires the unit to be "entirely separate from and independent of the personnel engaged in the direction and conduct of that study".3 So a report checker that QA runs on its own is a QA tool. A checker the study team runs before the audit is the study team's tool, and QA still does its own review.

Draft, audited draft and final: what each is good for at FDA

FDA's 1995 guidance on Phase 1 INDs accepts an integrated summary built on "unaudited draft toxicologic reports", with final, "fully quality-assured" reports available on request within 120 days of the start of the human study.21 GLP still applies in full, and a draft can be read by regulators, so its labeling and version control matter from the first release.

Keeping a toxicology report consistent: a percent change and a mean, worked through

In short"Internally consistent" sounds simple until you count the places a value appears. A calculated figure also depends on which inputs and which rounding you used, and two honest people can get different answers. The rule asks you to describe your calculations. Writing the rounding rule down, and calculating one way, prevents most arguments.

A weight gain that appears in three sections of the report

In the fictional study, mean male body weight gain from Day 1 to Day 29 is 142.7 g in the top dose group and 156.8 g in controls. Those two values sit in one spreadsheet row. The difference is:

That figure appears in the results, in the summary and in the conclusion. If a later correction changes 142.7, all three have to change, as well as the table, any table legend and any later summary a sponsor has built on it.

Two honest answers to the same difference in food consumption

Now take the four-week mean of food consumption. The weekly values for the top dose group are 24.6, 25.1, 25.7 and 25.9, and for controls 25.8, 26.4, 26.9 and 27.1 grams per animal per day.

Fictional example, our arithmetic. The screens in Part B show 25.3, 26.6 and −4.9%, calculated from the two displayed cells.

Neither figure is a mistake, and neither changes what the study found. The trouble comes when the table uses one method and the text the other, or when a QA auditor recalculates with a calculator and gets a different answer from the report. 58.185(a)(11) requires "a description of the transformations, calculations, or operations performed on the data".2 Say in the methods whether percent changes come from displayed or stored values, and apply that everywhere.

Words that have to stay the same from table to conclusion

Consistency reaches past digits. Pathology terms should match across the contributor report, the table and the text. INHAND, a joint naming project of toxicologic pathology societies, offers one shared vocabulary, and its organizers have worked with FDA and the standards body CDISC to bring its terms into SEND, the data standard described in section 11.23 Group names, dose labels, day numbering and units all drift. A report that writes "Day 29" in one table and "Week 4" in another for the same time point is a question QA will ask.

When a report is translated, every check runs again

FDA's drug center has presented a case in which English translations of final study reports did not accurately reflect the raw data: misspellings, omitted paragraphs or amended reports, mislabeled parameters, wrong tables and formatting errors. Some translated versions carried the study director's name printed by the translator. The compliance statement in one translation named different regulations from the original.24 A translation is one more copy, and it needs its own control.

What a GLP report says about a missing value, an excluded value and a plan deviation

In shortA number nobody recorded has to stay absent from the report, and the report says so. An excluded value stays on file with a documented reason. Anything that may have affected the data's quality has to be described. Writing these down costs a sentence. Leaving them out is what FDA cites.

Four things you can do with an empty cell in a data export, and one you never do

OECD's data integrity document is plain about exclusions: data may only be invalidated "where it can be demonstrated through sound scientific or technical justification or logical sense that the data are not representative of the recorded event". For common cases, "the rules to exclude or invalidate data should be defined in advance in the study plan or in SOPs".11

Deviations and unforeseen events: FDA's item (9) in practice

Item (9) of the report list asks for "all circumstances that may have affected the quality or integrity of the data". The word "may" makes this wider than confirmed problems, and it is one of the items FDA cites by number.2,25 The report is the last link: if nobody noted a problem when it happened, nobody can describe it later. 58.33(c) puts the duty on the study director to see that such circumstances "are noted when they occur".7

NOAEL and adversity in a study report: words the study director has to own

In shortMost of a report describes. The conclusion decides. Calling a finding adverse, and naming a no-observed-adverse-effect level, are judgments made and signed by people. This guide treats them only as report content: what the text should show, who writes it, and what stays out of a draft until they have.

The no-observed-adverse-effect level, or NOAEL, is, in plain words, the highest dose tested at which the study's scientists find no adverse effect, compared with the control group. The Society of Toxicologic Pathology's Adversity Working Group calls for it to be set "at the level of the overall study report".26 How a sponsor then uses that figure belongs to other documents and other people, and this guide does not cover it.

The same working group published best-practice recommendations for reporting adversity. In its words: "Adverse is a term indicating 'harm' to the test animal", "Test article-related effects should be described on their own merits", and decisions "to consider them as adverse or nonadverse should be justified".26 For a report writer, that gives a structure: first describe each change, with its group, sex, direction and size. Then give the reasoning. Then, separately and in the study director's name, the call.

How a draft should carry a conclusion that is not yet decided

A contributor's findings can arrive after the first draft, as the recovery pathology does in this guide's example. A conclusion written before those findings are signed is a draft of a draft, and the report should say so.

FDA GLP inspection findings: where the final report itself gets cited

In shortFDA finalized 44 GLP inspections in fiscal 2024, and in a separate review of 37 inspections, 17 ended with a Form 483. Several recurring items concern the report: deviations left out, circumstances not described, corrections made outside an amendment. Three warning letters show what that looks like in a real file.

What FDA's FY2024 483 review lists under the study director

FDA sorted its fiscal 2024 GLP citations into four themes: testing facility management, the study director, standard operating procedures and training. It published the themes without counts on its charts. Under the study director theme, three of the nine listed items are about the report:22

Final reports in three FDA warning letters, 2023 to 2026

The same findings at EPA and FDA's veterinary center

EPA's January 2026 update lists its 2025 inspection findings under the reporting subpart: names missing from final reports, "miscalculation and misreporting", and "final report does not reflect raw data". It explains why that matters: "EPA reviewers only see derived data in the final report. Reviewers make their regulatory decision based on the submitted final report."1 FDA's veterinary center put the same item first in its 2023 list of top findings in GLP and clinical studies, and listed two more on documents: amended reports that do not clearly say what changed and why, and a QA inspection of the final report missing from the QA statement.30

The worst outcome is disqualification. FDA may disqualify a testing facility when noncompliance "adversely affected the validity of the nonclinical laboratory studies" and lesser actions will not do.31 Short of that, a finding costs time: a written response, a revised report and a study director pulled from other work.

Transcription error rates: what is measured, and what it means for a toxicology report

In shortNo one has published an error rate for GLP report writing. The nearest careful measure comes from clinical data entry: about one field in 300 wrong before checking, and a check that catches between 40% and 88%, depending on the method and on who runs it. Treat that as a reason to measure your own rate.

In a Japanese kidney transplant registry, two operators keyed 560 forms, and checks found errors in 358 of 104,720 fields, a rate of 0.34%. In reading aloud, one operator reads the keyed data and another compares it with the form. The researchers compared two versions of that with two versions of double entry, each run by the same operator or a different one.32

What those rates would mean on a 400-number report, worked through

Take an illustrative report with 400 hand-copied numbers and borrow the registry's rate. That rate may not transfer, so replace it with your own.

At 4,000 numbers, the arithmetic gives about 14 slips before checking, and 1.6 or 8.2 left. These figures are illustrative. Copying a known figure into a table may be easier or harder than keying a form, and a tired reader at the end of a long report may do worse. The reasonable conclusions are narrower: slips are expected in a long report, the kind of check matters as much as whether one happened, and the weakest check measured left about 60% of the slips in place.

The better check costs more. The same study reports 74.8 hours for double entry against 57.9 hours for reading aloud, on the same forms.32 Across a long report, the second reader's hours add up.

SEND datasets and the study report: FDA expects them to agree, from the draft onward

In shortFor many toxicology studies, FDA also wants the data in a standard electronic form called SEND. It goes in with the PDF report, including a draft. So a modern study has two presentations of the same results, made by different hands, and a reviewer can compare them. Differences have to be explained in a short guide written for the reviewer.

The Standard for Exchange of Nonclinical Data, or SEND, is a format for submitting study data to FDA. FDA's Study Data Technical Conformance Guide, version 6.3 of October 2026, is nonbinding but specific:33

So every change to a report needs one more question: does this change the datasets, or the reviewer's guide? A correction made late in the report and forgotten in the SEND package gives a reviewer two answers. Appendix I of the guide asks for a data submission "consistent with the study report", and the people who create the datasets are often not the people who write the text.33

Research has reached the same seam from the other side. In 2026, authors at one large drug company described a pipeline that reads existing toxicology reports and turns them into structured data aligned with SEND terms. Testing on 200 of their own reports, they found sensitivity "typically above 95%" and precision "frequently exceeding 97%" for histopathology, organ weight and NOAEL-related endpoints, with weaker results on route and substance identifiers, which they tie to uneven reporting habits. All authors are current or past employees of that company.34 The study reads reports. Searching PubMed and the web on November 3, 2026, we found no published test of the opposite job: drafting a GLP report from study data.

AI in toxicology report writing: what FDA, EMA and the EU say as of November 2026

In shortNo regulator has written a rule for AI drafting of GLP reports. The existing rules still apply to whoever signs and to the systems used. FDA's AI draft leaves out drafting that does not affect the reliability of results, EMA says your procedures extend to AI in non-clinical studies, and the European GMP draft on AI does not reach GLP. Keep the record and the signature with people.

Three limits on AI drafting that follow from the GLP rules

How long a GLP report takes: FDA's paperwork estimate, and what nobody publishes

In shortLittle is published. FDA's paperwork estimate puts reporting at about 27.65 hours per response, and half a million hours a year across sponsors. It is a planning number. No public survey times a report from data received to signature, so the useful number is your own.

The notice's averages are worth reading closely: 27.65 hours per final report, one hour per QA statement, and 32.7 hours per protocol record.12 An hour for the statement is plausible because the statement itself lists dates; the audit behind it is counted elsewhere. For the report, 27.65 hours is an average over everything from short tests to long studies, and it is an estimate made for paperwork purposes. It is no benchmark for a repeat-dose study.

One public contract shows why a late draft hurts a laboratory's cash

Report dates also touch a sponsor's calendar. Tufts researchers put a day of delay at about $40,000 in direct trial cost, and at about $500,000 in lost sales in their journal article and about $800,000 in their white paper, all from one dataset of 645 drugs.44 Those figures describe trials in Phase II and III. A study report sits earlier, and it costs a day only when a filing is waiting on it, so use them with that limit in mind.

Who assembles a GLP report today: study director, contributors, technician, report writer and QA

In shortA study director sits at the center of a small group, and may carry several studies at once. In the fictional sample team, five of 11.2 days go to re-keying and cross-checking numbers that already exist; we found no public figure for a real team. Laboratories write in house, a contract laboratory writes for its sponsor, and software fills in parts of the work.

From study director to archivist: who touches a GLP final report

Six hand-offs between a data package and a signed GLP report

Where a repeat-dose report's working days go

In house, at a contract laboratory or with software: five ways GLP reports get written today

The VLT-TOX-028 draft: 156 numbers traced, one blank left empty and recovery data still to come

In shortThis is the example from the top of the page, inside the product. Six AI agents, each a program with one job, read the signed plan, the export and the template, and draft the report overnight. Dr. Lena Ortiz, the study director, reads it beside the export. A blank and a missing report stop her signature until the people who hold those records supply them. She signs; the agents never do.

1

Each number in the draft joined to its cell in the data export

The data package arrives on October 5, with study plan v2.0 and its amendment, the data exports, a signed toxicokinetics report and repeat-dose template v5.2. On the morning of October 6, the agents have drafted nine sections. The header reads 156 traced, 1 blank, 2 gaps and 3 of 9 sections approved.

Lena points at 142.7 g in Section 4.2. The sheet behind it opens at BW_Males, cell E7 lights, and a line joins the two. Below, the screen names the file, the sheet and the three sections that use this value: 1, 4.2 and 5. The export is locked and read-only.

2

A percent change lights both of the cells it comes from

The −9.0% is calculated, so it lights two cells: E7 and B7. The footer shows the working, (142.7 − 156.8) ÷ 156.8. If QA recalculates this figure, it can see at once which inputs the report used.

This is the question from section 6. On screen, percent changes in the sample use the displayed cells, and your house rule decides which you want.

3

The Week 4 cell is empty, so the sentence says so

Cell E5 of FC_Males has no value. The sentence reads "not in raw data", and no four-week mean is stated. The screen names the evidence: empty in the export of October 5, weighing logged on September 12 by Ana Kowalski, balance file not transferred.

The Blanks and gaps view lists this next to a second blocker. Study plan section 9.4 requires microscopic examination of the recovery groups, and that file is not in the package. A minor item asks Lena to confirm one excluded weighing, recorded as spilled.

4

One request, and the blank becomes a number from a corrected file

Lena clicks "Ask Ana Kowalski". Ana sends VLT-TOX-028_FC_v2.xlsx, signed and versioned, and the first export stays on the record. Week 4 is 25.9. The section is redrafted with a mean of 25.3 against 26.6 g/animal/day in controls (−4.9%), and the consistency checker finds 0 mismatches.

Lena confirms the exclusion of the spilled weighing, and a note appears under the table: which cage, which week, why, and the date of the raw data note. That is the record OECD asks for on excluded data.11

5

Recovery pathology arrives, and three sections are redrafted

Section 4.5 shows the gap, and the draft conclusion says plainly what it waits for: "Reversibility is not yet shown". Before the recovery data, the NOAEL reads "proposed", in Section 5.

Lena requests the data from Dr. Omar Haddad, the study pathologist. MI_Recovery.xlsx arrives for 20 animals, signed by him. Sections 4.5, 1 and 5 are redrafted, and the header rises to 166 traced numbers. The conclusion's wording is still Lena's to accept or change. The section drafter proposes adversity and the NOAEL and marks them for her.

6

She signs with a stated meaning, and the draft goes to QA

The dialog shows what she is signing: study report version 1.0, nine sections, 166 numbers traced, no open blanks or gaps, and a warning that six sections are not yet marked as reviewed. She enters her password and chooses a meaning. The fields follow 21 CFR 11.50 and 11.70: printed name, date and time, meaning, and a link to this exact version.36

The sample uses the meaning "release to Quality Assurance for audit", and the draft goes to Dr. Maya Chen. So read this signature as the study director's release of the draft. The page stamp reads "FINAL V1.0", yet QA has not audited this version, so it marks the release of a draft. The GLP final report signature under 58.185(b) belongs after QA's audit, on the report QA's statement describes, and the same dialog offers "final report" and "report amendment" as separate meanings. Your procedures decide the order.2,3,4

The solution's design targets are 1.4 days from data received to a signed draft, 100% of numbers linked to a data cell with none estimated, and 0.3 transcription findings per report at QA audit. They are targets, to be measured in the live solution and on your own studies. The fictional team's by-hand baselines are about 11 days and 4.6 findings. The sample study's own numbers are the ones on these screens.

Number trace, data package, studies list and rules: the rest of the report writing screens

In shortOne study is the story. These screens show how the same discipline runs across a team: the trace an auditor can download, the package checked against the plan, every blank across studies, a genetic toxicology report waiting on a historical range, and the settings a QA lead would ask about.

Select any screen to open it larger. The study, people and figures are fictional sample data. Rule text, section numbers and dose labels on the screens are the sample's own; for the real clauses, see Part A.

Six report agents, one study director and the systems a GLP report draws on

In shortEach agent has one job and a line it cannot cross. People decide what the report concludes and sign it: contributors review their sections, the study director signs the report, and QA audits on its own and signs its own statement. The solution reads the files you already make and hands back a report in your template.

From study plan reader to consistency checker: the six report agents

Who signs a VLT-TOX-028 report, and under which control

The agents' work supports the people who answer for the report. Controls such as roles, versioning and an audit trail support your compliance with 21 CFR Part 58, OECD GLP and Part 11. Whether a system is validated for your use is your facility's determination, on your procedures.35 Botminds' own credentials are SOC 2 Type II audited and information security managed to ISO 27001.

From signed plan and export to a report with its trace

Fits the study systems and quality records you already run

The solution works from what a study already produces: study data system and laboratory system exports, spreadsheets, signed PDFs and Word templates. Those systems stay where they are, and their records remain the records. It serves large programs with many concurrent studies, with roles and separation of duties, maker and checker on each section, and a trace and history you can export for audit.

What hand-keyed tables and late drafts cost your study directors and QA unit

In shortStart from your own studies. We found no published figure for what a GLP report costs or how many errors one carries, so the numbers that matter are in your timesheets, QA audit records and study calendars. A few published figures frame them.

Pulling a baseline from your last six study reports

The study report costs that no timesheet shows

What GLP report automation returns once you count keying, checks and data queries

In shortPut in your own numbers. The calculator values two things you can measure: hours per report and the hours QA findings cost. Calendar value appears on its own line and stays out of the total, because it counts only when something is waiting on the report. Every default is an assumption to replace.

Keying, number checks and queries: where report hours come back

Work out your return from reports a year, hours per report and QA findings

At these defaults, a group writing 40 reports a year pays for the tool on hours alone. A group whose reports look more like FDA's 27.65-hour average may save too little per report, and should work that out first. Short reports with few tables gain less per report than repeat-dose studies full of them.

How ready your study reporting group is: raw data, templates, QA and study director sign-off

In shortPick the answer closest to your group today. Eight questions score 0 to 2, and the last is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like.

Three readiness bands: settle your raw data, replay a closed study, or plan the rollout

What to ask a toxicology report software vendor about cells, blanks, contributors and signatures

In shortAsk every vendor the same questions, us included. The answers show how a draft is built: whether a number can exist without a source, what happens when a cell is empty, whether contributors keep their own reports, who signs what, and how you would measure it on your own study.

A twelve-week GLP report pilot: replay a signed study with planted faults, run one live report, then decide

In shortStart with a study whose report is already signed, so you know the right answers. Plant errors in a copy of its export, run the tool and count what it finds. Then run one live study in parallel, with your own process still responsible for the record. The plan below covers one study type in twelve weeks.

Catch rate, false alarms, hours and QA findings for each planted fault

GLP final report FAQ: the 14 items, signatures, QA, amendments, SEND and AI drafting

In shortThe questions people search for most, each answered on its own.