CRO data integration, explained.

CRO data integration explained: six labs, six file layouts, one research database

One plasma file, from a lab's drop folder to the compound ranking, in five steps

1. The lab deliversAt 9:40 am on October 7, Quillon drops a workbook on the sponsor's server: nine plasma samples for one compound, under one schedule of one contract.
2. Someone logs itWhich CRO, which study, which contract clause, which compounds were shipped. A file with no match waits for a person.
3. The file is compared and mappedAgainst the layout last accepted from this lab, then into the sponsor's own fields and units. Most of this guide lives at this step and the next.
4. Values are checked against the contractControls, replicates, recovery, time points, compounds delivered, turnaround. Below-limit samples are kept as below limit.
5. A person decides, then it loadsAnything that stopped gets a named decision. The records reach the database with their sources, and scientists can use them.

The people, labs, compounds and figures in this example are fictional.

The file that looks fine, and three more surprises in contract lab data

Deliverable, house schema, LLOQ and layout drift

Fourteen terms are defined in the one-page summary.

Signed report, results workbook and instrument export: the layers of one CRO package

In shortOne package usually carries the same result in more than one place, in different forms, with different standing. The report says what the lab stands behind. The workbook is easiest to read. The export is closest to the instrument. The contract and the shipping manifest tell you what should be there at all. A good load cites one and can check it against the others.

How many values one package becomes, and how many need a unit changed

A single stability report for 23 compounds in the example gives 96 values, 46 of which need converting to house units. The plasma file gives only nine values, and seven need converting, because the lab's new header moved the whole column. The fictional sponsor in this guide receives 38 deliverables from six labs in one week.

A half-life in a report, a workbook and a raw sheet: which copy to cite

A half-life can read 17.9 in a report table, 17.9 on a workbook summary and two replicates on a raw sheet. When they agree, you want the citation that a second reader can open fastest. When they differ, you want to know before anyone chooses for you.

The FAIR principles, a widely cited statement of what makes research data reusable, ask for the same thing in general terms: data findable, accessible, interoperable and reusable, with machines able to act on it.4 A value re-typed into a spreadsheet with no source falls short of the last two on the day it is typed.

Discovery assay, GLP study or trial feed: which rule applies to which CRO data

In shortContract lab data falls into three rough groups. Discovery results are for your own decisions, and few regulations reach them directly. GLP studies and studies supporting a submission have firm rules on records and formats. Clinical trial feeds have their own. Many sponsors receive all three, often from the same lab, so the first question for any file is which group it belongs to.

Discovery data: your own decisions, with no format rule behind them

Part 58, FDA's good laboratory practice rule, covers nonclinical laboratory studies meant to determine a product's safety, and its definition says the term "does not include basic exploratory studies carried out to determine whether a test article has any potential utility".7 Most in vitro ADME, potency and screening work lives on that side of the line. The rule that protects you here is commercial: you pay for these results and you rank compounds on them. Your contract sets your standard.

GLP and submission studies: signed reports and archived records

When a sponsor uses a contractor for an analysis that is part of a GLP study, it must tell the contractor the service is part of a study that must comply with Part 58.10 Testing facilities keep raw data, documentation, protocols, specimens and final reports in archives, with a named person responsible, for periods set by the rule.11 So a copy in a sponsor's research database, often called a data warehouse, does not replace the record. Teams usually call it a convenience copy and label it as one.

FDA keeps inspecting this work. It gave a final classification to 44 GLP inspections in fiscal 2024, against 29 in fiscal 2023. FDA says these metrics cover inspections with a final classification date in that year, so they do not show every inspection it conducted.12 Common observations include a study director who failed to assure that all experimental data were accurately recorded and verified.12

Clinical feeds: the same discipline on a bigger scale

Sponsors also take data transfers from clinical CROs and central labs. A Tufts and TransCelerate study of 105 protocols puts the average Phase III protocol at 5.9 million datapoints, up 11% annually since 2020.13

Under ICH E6(R3), the international guideline for good clinical practice, the sponsor may transfer activities to a service provider, yet "the ultimate responsibility" for them, including the reliability of the trial data, stays with the sponsor (section 3.6.6).9 Section 4.2.5 asks that data moved between systems keep their integrity, with documented traceability and reconciliation "to avoid data loss and unintended modifications".9 EMA's 2023 guideline on computerised systems in trials says the same in one line: "The process of data transfer between systems should be validated to ensure the data remain accurate."14

Such a transfer may carry dozens of tables, among them laboratory results and vital signs. In this guide they are all trial datasets whose columns must match the data transfer agreement. Layout checking asks whether the file arrived as agreed. What the content means for a participant belongs to the trial's own clinical processes, and this guide does not cover that work. The example in this guide stays with nonclinical data.

SEND, SDTMIG 3.4 and ADaMIG 1.3: what FDA's Data Standards Catalog requires, with dates

In shortCDISC, the Clinical Data Interchange Standards Consortium, writes the standards, and FDA's Data Standards Catalog says which versions FDA supports and requires, and from when. The catalog in force is version 11.0, posted in March 2025. CDISC has published newer versions than FDA lists, so "the latest version" and "the required version" can differ. Requirement dates differ by center and application, and for SEND they depend on when a study started.

SEND, SDTM, ADaM and Define-XML: what each CDISC standard covers

The dates a CRO data contract should name

The catalog's footnotes separate NDAs, ANDAs and certain BLAs from certain INDs. Its requirement column is the date the requirement "becomes effective"; FDA's Technical Conformance Guide reads the SEND dates as applying to studies started after each date.5,2

SENDIG, SDTMIG and ADaMIG are the implementation guides for each standard. NDA: new drug application. ANDA: abbreviated new drug application. BLA: biologics license application. IND: investigational new drug application. CDER and CBER: FDA's centers for drugs and for biologics. The Animal Rule is FDA's route for relying on animal efficacy studies.

Which studies owe SEND, and why most discovery assays never do

FDA's guide says the SEND requirement "is not limited to GLP studies": a non-GLP toxicity study submitted to support a determination of safety must include SEND.2 It also says the expectation links to the pharmacology and toxicology information FDA needs to assess proposed clinical investigations. So the same mouse study can matter in two ways: exploratory work stays outside, and a study relied on for a decision comes inside. That is a question for your regulatory team early, before the study closes.

For most discovery assays the question never arises. In vitro stability, permeability, protein binding, inhibition of CYP liver enzymes and plate-based potency have no table in the guide, which says SEND is not required for study types no FDA-supported SENDIG lists, such as primary pharmacology.2 Sponsors define their own house schema, and each lab keeps its own layout.

Two technical items are worth a line in any contract. XPT v5, the SAS transport file format, is still the exchange format in the catalog.5 FDA's resources page says CDER and CBER are testing Dataset-JSON, with "Initial results indicate potential use as a replacement for XPT v5".19 That is a test result, with no switch date set. FDA's guide also carries technical rejection criteria for study data, so a submission can fail on a missing trial summary file before anyone reads a result.2

What 21 CFR Part 11, Part 58 and ALCOA+ ask of a value copied from a contract lab

In shortPart 11 sets conditions for electronic records that another FDA rule requires, or that go to FDA. Part 58 sets the conditions for GLP records. ALCOA+ sums up what regulators expect of data in general. No software can be certified against any of them: a tool can offer controls that support your compliance, and your own validation and procedures do the rest.

Does Part 11 apply to your research warehouse?

It depends on what the warehouse is for.

Once you rely on a value from the warehouse to support a submission, the picture changes, so your quality unit should write down where that line sits. FDA's 2003 guidance counts records that a predicate rule requires and that are kept in electronic format in place of paper, or relied on to perform regulated activities.21

ALCOA+ applied to one cell in a workbook

ALCOA stands for attributable, legible, contemporaneous, original and accurate. WHO's guideline uses ALCOA+, adding complete, consistent, enduring and available.1 The UK's MHRA uses plain ALCOA and says the plus adds "no difference in expectations".28 Here is what each word asks of the concentration in cell E5.

Renamed, moved or redefined: three kinds of change in a contract lab's file

In shortA renamed or moved column changes where a value sits. A redefined column changes what the number means, and it can keep its letter and position. The first two can be matched and recorded by a careful process. The third needs a person, and it is the one a fixed import misses. Compare each file with the last accepted layout before anything is mapped.

What the 1,000-fold change does to one mouse profile

Take the file from the example. Quillon's last three runs for this compound read 1,190 to 1,410 ng/mL at 0.25 hours. Today's reads 1.284 in the column now headed µg/mL. Multiply by 1,000 and it agrees with history at 1,284 ng/mL. Read it under the old header and the whole profile lies along the floor of the chart, with the 8-hour value at 0.0275, far under the method's 1.0 ng/mL limit. That last point gives the check a second way in: a reported value under the lab's own limit is impossible.

Eight points to compare before mapping, in order

1. Sender and study: is this the lab and study the contract expects?
2. Sheet names: did a sheet get renamed, added or hidden?
3. Header text: any column renamed?
4. Column order: any column moved or inserted?
5. Units written in headers: any change at all, in any column?
6. Qualifier vocabulary: BLQ, <LLOQ, ND, a blank or 0, compared with last time.

7. Value ranges: does each numeric column sit near the last few runs? A jump of ten or more deserves a look even when the header is unchanged.
8. Row count and compound list: against the manifest and the contract.

A unit mismatch cost NASA an orbiter, and spreadsheets change data as quietly

NASA lost the Mars Climate Orbiter in 1999. Its investigators gave the root cause as "the failure to use metric units in the coding of a ground software file".29 The data was in English units, the receiving software expected metric, and the board noted the error was not caught. A lab file can fail the same way: two teams, one interface, a unit both assumed.

Spreadsheets add their own version of the problem. A 2016 study found that about one-fifth of papers with supplementary Excel gene lists carried erroneous gene name conversions, because the software silently turned names into dates or numbers.30 No error message appeared there either.

Converting µg/mL, half-life to clearance and replicate CV: working CRO numbers by hand

In shortMost of the checking is simple arithmetic. Each conversion should be a named rule with one multiplier. A half-life turns into clearance only if you know the protein concentration. A replicate limit is one formula. Do each once by hand so you can judge any tool, and so you can see where a value carries an assumption.

Six unit rules from the example, each one line

The rule for 0 with flag BLQ has no multiplier at all: keep the qualifier, as section 8 explains. The nM row's 18.7 µM is from the example's fictional hERG study; the hours row is our own illustration.

From half-life to intrinsic clearance on one compound

A microsomal stability assay incubates a compound with liver enzymes and measures how fast it disappears. It is one of the assays in vitro drug metabolism good-practice papers describe.31 The lab reports a half-life, the time for half of the compound to disappear, and the sponsor usually wants intrinsic clearance, or CLint, per milligram of protein.

The usual arithmetic is the one below: the half-life gives a rate, and the incubation volume per milligram of protein scales it, as one contract lab's published method page sets out.32 It reproduces the 28.8 and <23.1 on the example's fictional report, which states 0.5 mg/mL of microsomal protein. This is arithmetic for reading a report. The assay figures here are sample report content, so take no method, condition or dose from them.

Compound NWB-4471: half-life 48.2 min, so 0.693 ÷ 48.2 = 0.01438, then ÷ 0.5 × 1,000 = 28.8 µL/min/mg. Compound NWB-4472: half-life reported as >60 min, so the clearance can be stated as <23.1 µL/min/mg, and it stays a bound. If a report omits the protein concentration and you assume 0.5 mg/mL where the lab used 0.2, your clearance reads 2.5 times too low, and nothing in the number gives it away. So a contract should ask for incubation conditions in every report.

The replicate limit, on the one compound that failed it

NWB-4488 in human microsomes gave half-lives of 17.6 and 27.6 min. The mean is 22.6. The sample standard deviation is 7.07. The coefficient of variation, or CV, is 7.07 ÷ 22.6 = 31.3%, against a contract limit of 20%. Two numbers can look like one good result once averaged at 22.6. The CV is the only sign of a problem.

Recovery, and what it says about permeability

A Caco-2 assay measures how fast a compound crosses a layer of gut cells. Recovery, also called mass balance, asks how much of what you put in can be found at the end. In the example, NWB-4481 recovers 58% against a contracted 70%. The compound may have stuck to the plate or stayed in the cells, so its permeability may read low. A number below the limit does not make the result wrong, and it should not reach a chemist without that flag.

BLQ, LLOQ and the digit zero: what to load for a sample below the limit of quantification

In shortA result below the LLOQ means the method could not put a reliable number on it. Store that fact, the limit and the unit. A zero or a half-limit written into the database closes off choices the analysis should keep open, and moves results in ways that depend on how they are later used. FDA's guide gives exact instructions for SEND datasets, and they make a good house rule.

What a limit of quantification is

ICH M10 defines the LLOQ as "the lowest amount of an analyte in a sample that can be quantitatively determined using a method with predefined precision and accuracy".8 For chromatographic methods such as LC-MS/MS, calibration standards are allowed to sit within ±20% of nominal at the LLOQ and ±15% elsewhere.8

So a value just above the limit is already the least certain number in the study, and a value below it is outside the range the method was validated for. M10 also recognizes that the limit can move in a run, when the lowest standard is rejected and the run's LLOQ rises.8 That is why the limit belongs on each record: one figure in a method note will not always do.

FDA's instructions for a PC record below the limit

FDA's footnote 35 adds that its 2018 bioanalytical guidance says 'BQL', while 'BLQ', as in FDA-supported SENDIG versions, is appropriate in SEND datasets.2 Discovery data needs none of this formally. It is the clearest public model for a house schema.

What a zero does to one profile: it adds area, and pulls down an average

The area under the curve, or AUC, turns a profile into one exposure figure. Work the example's intravenous (IV) profile with the linear trapezoid, the simple sum of areas between time points. From 0.25 to 8 hours: (1,284 + 612) ÷ 2 × 0.75 = 711; (612 + 143) ÷ 2 × 3 = 1,132.5; (143 + 27.5) ÷ 2 × 4 = 341. That totals 2,184.5 ng·h/mL to the last measurable sample.

Now write 0 at 24 hours. The next trapezoid is (27.5 + 0) ÷ 2 × 16 = 220, because the straight line runs across 16 hours. That adds about 10% to the AUC, 2,184.5 becoming 2,404.5, against the expectation that a zero pulls exposure down. For comparison, a fitted tail carries on the fall seen between 4 and 8 hours, a half-life near 1.7 hours. It adds 66.7 beyond 8 hours, about a third of the area the zero added.

A zero also pulls down any average taken at that time point, so a group mean falls. Its logarithm does not exist, so a fitted slope cannot use it at all. The same digit can push one figure up and another down, and a reader of the warehouse has no way to tell which. Each team should write down how it treats samples before the first measurable one and after the last, and keep the marked observation underneath.

Population modelers have measured the cost of these shortcuts. In simulated datasets, Bergstrand and Karlsson found that omitting below-limit data "was associated with substantial bias in parameter estimates for all tested models even for seemingly small amounts of censored data", and that substituting half the limit introduced bias in some cases and was "always suboptimal to the best method".33 The best method used the likelihood of being below the limit, which needs each observation marked with its limit. A database of zeros cannot give it that.

Four states a research database must keep apart

When a CRO marks a control "Pass" under its own SOP and your contract says otherwise

In shortA lab reports against its standard operating procedure unless told otherwise. Your statement of work may set tighter ranges. A "Pass" in the report answers the lab's question, and the sponsor still has to ask its own. The check takes a minute per criterion, and the hard part is having the contract's numbers open when you read the file.

Positive controls are compounds with known behavior, run alongside your compounds to show the test system was working. In the example, verapamil in mouse microsomes on plate 2 runs at 17.9 minutes. The lab's SOP accepts 3 to 20 minutes, so its report says Pass. The sponsor's contract, schedule 4.2.3, says 3 to 15 minutes.

On page 11, in a separate section, the report also notes a second freeze-thaw of the microsome lot used on that plate. Nobody at the lab has done anything improper. Their report answers against their own procedure. The two facts sit five pages apart, and neither sits beside the contract.

Hold the plate, flag the value, record the gap or call the steward: four decisions

These are decisions a sponsor's team might reach on the example's fictional facts. Your procedures and the science of the assay decide yours.

Hand-copying errors, unseen layout changes and rejected CRO studies: what the record shows

In shortPublished error rates by method are large and come from clinical research. Nobody has measured this for discovery files, so treat them as a range to test against your own. Separately, regulators have forced repeat studies when a CRO's data could not be trusted, and the sponsors carried that cost. Layout changes and qualifier mistakes leave little public trace because they are found, or missed, inside each company.

Reading a CRO's report and writing values into a template resembles record abstraction more than typing from a form: a person finds a number in a page laid out for reading, then decides which field it belongs in.

Set those rates beside the example's 96-value report, as a rough illustration. At 0.29% you expect about 0.3 wrong values. At 6.57% you expect about 6. Across 31,482 values over eight weeks, those rates give about 91 and about 2,068. These are clinical figures applied to a lab setting, so treat them as a prompt to measure your own rate.

When FDA rejects a CRO's data: one recent case on the public record

The sharpest consequence comes when a regulator rejects a CRO's work. After a general correspondence letter of August 6, 2024 and two replies from the firm, FDA's untitled letter of March 27, 2025 told Raptim Research, a CRO in India, of its conclusion on in vitro bioequivalence studies: "your firm created and caused the submission of falsified data to FDA".36 FDA's request to generic applicants says "all study data from all in vitro studies conducted at Raptim must be rejected" and asks for a reply within 30 days.37

FDA's notice of March 28, 2025 says in vitro studies essential to approval must be repeated at sites without data integrity concerns, handles in vivo studies case by case, and moves certain generics relying on Raptim's bioequivalence data to a BX therapeutic equivalence rating.36,38 BX is FDA's code for products whose data are insufficient to determine therapeutic equivalence.39 A law firm's summary notes no allegation of falsified in vivo data, and that sponsors could address those studies through re-analysis or a new study.40

That case concerns conduct at a lab, a different problem from a changed column. It shows who pays, though. FDA's April 2024 draft on data integrity in bioavailability and bioequivalence studies says responsibility "resides with applicants", even when the work is contracted.26

Eight ways a value goes wrong between the lab and your database

How FDA classified its GLP inspections, fiscal 2020 to 2024

These counts measure FDA inspections of GLP facilities, with no split for data transfer to sponsors. We found no public count of findings about a sponsor's own warehouse, which is not usually an inspection target unless it feeds a submission.

Reading CRO files with AI: where FDA's draft, the EU and data integrity rules reach

In shortReading a PDF table and matching a column are tasks AI can help with, if every output links to its source and a person can check it. FDA's January 2025 draft leaves out drug discovery. Other parts of the work may fall inside once a value supports a nonclinical result used for a regulatory decision. Decisions on what a changed unit means stay with people, whatever any rule says.

What FDA's draft leaves out, in its own words

FDA's draft "does not address the use of AI models (1) in drug discovery or (2) when used for operational efficiencies (e.g., internal workflows, resource allocation, drafting/writing a regulatory submission) that do not impact patient safety, drug quality, or the reliability of results from a nonclinical or clinical study".41

A tool loading discovery results into a research warehouse looks like the first. A tool shaping a nonclinical value later relied on in a submission may affect "the reliability of results", and so may come within it. That line is worth settling with your regulatory team before any such use, and the safest design keeps a GLP value's record at the source and labels any copy.

Which steps suit a machine, and which need a person's name

Four things to plan around when AI reads CRO files

• Accuracy is yours to measure. We found no public benchmark of extraction accuracy on CRO deliverables. Test any tool on your own finished studies, against values a person has checked.
• Transcription rules still apply to machines. OECD's advice on GLP data allows a validated system in place of second-person verification.25 For GLP records, that phrase means your own validation work, and reading a vendor's slide will not cover it.
• Draft Annex 22 sits in another field. It concerns critical GMP uses, so it does not govern a discovery warehouse.43 Its idea of keeping qualified personnel responsible for outputs in non-critical uses is a sensible model.
• New data types are coming. FDA's March 2026 draft on new approach methodologies, as the Regulatory Affairs Professionals Society (RAPS) reports it, covers how to characterize such non-animal methods and names organs-on-chips and in silico modeling among them.46 Expect more such outputs from labs, in layouts nobody has agreed yet.

Study monitors, data stewards and spreadsheets: how external research teams handle CRO files today

In shortMost teams use a mix of a shared drive, a template per assay, some scripts and a lot of careful people. The work falls between systems built for other jobs. No public study measures it, so the time split below is an estimate to replace with your own. The market is large and growing, and each new lab adds a new layout.

Monitor, steward, scientist, research head, IT and quality: six roles around a contract lab's file

About 4.25 working days in all for one in vitro ADME deliverable, by our own planning figures. Time your own.

LIMS, ELN, data management system and warehouse: what each one does with an outside file

Our own categorization. Boundaries between them blur from one product to the next, so ask each what it does with a file it did not create.

Templates, a required format or a parser: how sponsors keep up with lab files

How big the preclinical CRO market is, and why new labs mean new layouts

Each firm's scope differs, so treat the gap as a sign of how soft the numbers are. We found no reliable public figure for the share of discovery work that sponsors outsource, or for how many labs a typical sponsor uses.

One more force could add layouts. The BIOSECURE provisions of Public Law 119-60, enacted December 18, 2025, will limit federal contracting with listed biotechnology companies of concern once their limits apply. A list from the Office of Management and Budget is due within a year of enactment, and we found none published on November 3, 2026.51 Federal contractors may review their labs, and each lab a sponsor moves to brings a new file format.

One plasma file in seven screens: a unit change stops it, and nine cited values load in 11 minutes

In shortPriya Ranvelle, the study monitor, sees every lab on one conveyor. Quillon's plasma file arrives at 9:40 am and stops at the Mapped station because a unit changed. She converts it, saving the new layout as template v4. At the next station, two zeros stop it again. She loads them as below the limit, and nine values reach the warehouse at 9:51 am, each linked to its cell.

1

Every lab on one conveyor

The Intake screen has a lane for each of six labs and four stations: Received, Mapped, QC-checked and Loaded. This week shows 38 deliverables: 30 loaded, 1 ready, 4 stopped for a decision, 1 waiting on a lab and 2 on the way in. On the right, Drift watch lists each lab's accepted layout and how many changes were caught this quarter. Quillon's new file sits at Received.

2

The layout is compared before a value is read

The deliverable classifier matches the workbook to Quillon, study QB-1180 and contract NWB-SOW-047 §2. The template drift checker then compares the header with Quillon's template v3, last used on QB-1174 on September 29, and finds four changes. One of them is the concentration unit, which puts the values 1,000 times below the last three runs.

3

What would have loaded, shown before it does

Sample became Sample ID and Flag became BLQ, both matched by meaning. Route and Time (h) swapped, matched by name. Column E went from "Conc (ng/mL)" to "Conc (µg/mL)", and that stops the file. The panel shows M1-0.25 IV would have loaded at 1.284 ng/mL against a correct 1,284 ng/mL. Dr. Ranvelle's three choices are to convert and save the template, query Quillon, or hold the file. She converts, and Sam Patel, the template owner, is informed.

4

Two zeros, each traced to its cell

The result extractor reads nine samples with their cells. The unit and identity normalizer converts seven values to ng/mL. The study-spec checker stops at two samples, M1-24 and M4-24, written as 0 with flag BLQ in cells E9 and E13. It suggests loading both as "below LLOQ (1.0 ng/mL)".

The card says a zero "would pull exposure down". On these sample numbers that holds only for an average at that time point. The IV profile's total goes the other way: written at 24 hours, the zero adds about 10% to AUC, 2,184.5 becoming 2,404.5 ng·h/mL (section 8). Either way the answer is the same.

5

Every stop decided, with a name on each

Dr. Ranvelle loads both as below LLOQ. The panel now reads that every check passed or is decided: nine values mapped to house units, each citing its cell or page. Both decisions show the person who made them. The button offers to load nine values, and nothing loads until she presses it.

6

Loaded at 9:51 am, 11 minutes after it arrived

One click on "Load 9 values". The warehouse loader checks the records against house schema v4, writes them with their units, qualifiers and flags, links each to its page or cell and refreshes the compound view for the Cedar program. QB-1180 moves to the Loaded station. On the sample data this took 11 minutes, with two human decisions inside it.

7

Any value opens its source cell

Dr. Lena Ortiz, a DMPK scientist, opens the review. The workbook sits beside the record grid. She clicks NWB-5104, IV at 0.25 hours, and a line runs to cell E5. As reported 1.284 µg/mL, in house units 1,284 ng/mL, confidence 97, converted by the rule approved today. The 24-hour samples read <LLOQ (1.0 ng/mL). The export's run line gives the limit as 0.0010 µg/mL, which agrees.

One stability study from Merrow: a control the lab passed, a contract it failed and 96 values each with a page or cell

In shortMerrow Biosciences sends a metabolic stability study: 23 of 24 compounds, 96 values, 46 converted to house units. Its report marks every control Pass. The study-spec checker reads the same page against the contract and raises three exceptions. Dr. Ranvelle decides each, approves a drafted query, and 74 values load while 22 wait for a repeat plate.

1

The report and the records, side by side

Page 6 of the report sits on the left, with the controls table. Verapamil on mouse microsome plate 2 reads 17.9 minutes, beside an acceptance range of 3 to 20 and the word Pass. On the right, each value has its compound, matrix, result, unit, confidence and source. The plate 2 line carries a red cross. Qualified values keep their form, such as >60 min and <23.1 µL/min/mg.

2

Three exceptions, each with its evidence

E-311 quotes the report's table, contract §4.2.3 and the note at §6.2 about a second freeze-thaw of lot ML-2291. It affects 22 values across 11 compounds, and it suggests holding them and asking for a repeat. E-312 is NWB-4488 at a CV of 31.3%, with its two replicates, against a contract limit of 20%. It suggests loading with a flag. E-313 records NWB-4490, which precipitated and was never incubated. Each evidence line opens the source.

3

A query that names the page and the clause

The CRO query writer drafts a short note to the lab's study director: what was found, page 6, Table 2, schedule 4.2.3 of the contract, the freeze-thaw note, what is being asked and which values are held. Dr. Ortiz is copied. Dr. Ranvelle edits if she wishes and presses approve. Nothing goes out before she does.

4

Contract and SOP, row by row

Study specs lists each contract's acceptance criteria beside the lab's own SOP. For microsomal stability, four rows match. Two read "Contract wins": verapamil in mouse microsomes, 3 to 15 against 3 to 20, and replicate CV, 20% against 25%. A strip shows the last 10 deliverables. From here Omar Haddad, head of external research, can propose an amendment.

A swapped tube label, a leaking well and a GLP report: three more decisions on the conveyor

In shortNot every stop concerns units. A name off the manifest goes to the data steward. A leak in a cell layer and a low recovery go to the study monitor. A GLP study loads as a labeled convenience copy, with the signed report staying the record. Each has its own reason, its own owner and a proposed answer to accept or change.

The deliverables list, the exceptions board, the drift log and the scorecard that renews a contract

In shortAround the conveyor sit the views for the rest of the team. The deliverables list shows every package and its turnaround clock. The exceptions board groups what needs a person. File templates keep each lab's versions and every layout change. The dashboard turns all of it into a scorecard, and settings hold your own rules.

Seven agents, one person deciding each stop, and what goes in and comes out

In shortSeven agents each do one step and hand on. None accepts a unit change, marks a failed control as passed, sends anything to a lab or loads with a decision open. Named people decide. Files come from your existing channels, and cited records go to your existing warehouse.

Classifier to loader: each agent's job, and what it never does

Priya, Sam, Lena, Omar and Dana: who appears in the Northwind Bio screens

A named person decides every stop, and a maker can be separated from a checker by role: the study monitor approves what a query writer drafts, and the steward owns what a monitor accepts into a template. Each action is on an audit trail with who and when. These controls support your compliance with Part 11 and GLP, and your validation and procedures establish it. Botminds is SOC 2 Type II audited, and manages information security to ISO 27001.

Reports, workbooks and contracts in; cited records and a scorecard out

The systems it works beside: drop folders, portals, registry and warehouse

2.1 hours, every file compared and a 70–90% estimate: the label on each number

A wrong unit, a late load, a re-keying error: pricing the damage to a discovery program

In shortStart with your own warehouse. Published figures frame the question: how often hand entry goes wrong, and what a day costs later in development. What a bad value did to one compound decision is a number only your program history can give you, and we found nobody who has published it.

Counting your own re-keying hours, wrong values and rework

A dropped compound, lost trust, weak models: five costs beyond the calculator

Re-keying hours, avoided repeats and days to the warehouse: the return from automated CRO data loading

In shortPut in your own numbers. The calculator counts the hours spent getting values from files into your warehouse, and any repeat work you expect to avoid, against what the tool costs. It shows days saved separately, because a faster load only has value if someone was waiting for it. Every default marked illustrative is an assumption to replace.

From re-keying hours to a reusable history: what a cleaner load is worth

A calculator for CRO files: values a year, seconds per value, cost per hour

A sponsor with one or two labs and a few studies a month may not justify software on hours. For that sponsor, an agreed template, a kept copy of each lab's last accepted file and a second reader may be the right answer.

Numbers on our own solution: 2.1 hours from receipt to warehouse and every file compared before loading are design targets. A 70 to 90% cut in transcription effort is our own estimate. None of the three is a published result.

Before you automate CRO data: eight questions on labs, contracts, house schema and who decides

In shortPick the answer closest to your team today. Eight questions score 0 to 2; 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.

House schema and contracts first, a two-lab replay next: reading your score out of 16

Fourteen questions that test CRO data software on unit changes, qualifiers, citations, GLP copies and who decides

In shortAsk every vendor the same questions, us included. The answers show what happens when a lab changes a header, whether a below-limit value survives, how any value traces to its source, how a GLP copy stays labeled, who decides, and how you could test all of it on your own past files.

A ten-week pilot on two labs' past files: unit changes caught, values checked and hours to load

In shortStart with your own warehouse check, which is useful whatever you decide. Then replay two labs' past files with changes you know about, and run alongside live deliverables for one full cycle with your team deciding as usual. Decide with four numbers: values wrong, changes caught, stops that were false alarms and hours per deliverable.

1. Sample 200 values from your warehouse against their source reports. That count gives you a starting error rate, whatever you decide about software.
2. Write your rules on one page: house fields and units, unit conversions, qualifier handling, who decides what.
3. Choose two labs and collect a year of their files, as they arrived, with the contracts. Mark any layout change you know of, and plant one you do not tell the vendor about.
4. Run alongside live deliverables for one full cycle, with your team deciding and loading as usual.
5. Decide with four numbers: values wrong in a checked sample, layout changes caught, false stops and hours per deliverable.

Changes caught, values checked and false stops on your own files

What people ask about loading CRO data: units, BLQ, SEND, Part 11, GLP copies and AI

In shortShort answers to the questions people type most often. Each stands alone, and the sections above carry the detail and the sources.