Short answer
A deviation report records an instance where work departed from a written procedure, with what happened, the affected step, the reason, the impact assessment and the corrective action. Regulated manufacturers must keep them. They give text-based examples of exceptions, failure causes and recovery decisions that SOPs omit, but they rarely include images or motion data and are heavily redacted.
The buyer's problem
Procedures describe the path. Real work leaves it: a part is missing, a gauge reads out of range, a step is done out of order, a time limit is exceeded. The record of what happened next is where procedural knowledge about exceptions lives, and it is almost absent from SOP corpora. A planner or evaluation set built on procedures alone sees only the happy path.
Regulated manufacturers produce this record as a matter of compliance. In US drug manufacturing, the current good manufacturing practice rule for finished pharmaceuticals requires written procedures for production and process control to be followed and documented at the time of performance, and any deviation from the written procedures to be recorded and justified (21 CFR 211.100). Other regulated sectors have comparable nonconformance systems under different names. This page treats those reports as a text asset. For prescribed responses to abnormal situations, see exception and emergency procedures.
What a deviation report contains
Formats vary by company and quality system, but typical fields are:
- Identifier, date, area, product or batch context.
- Description of the event: what was done versus what the procedure specified, often with the step reference.
- Detection: who found it and how (operator, inspector, review of records).
- Immediate action: containment, such as quarantining material.
- Classification: minor, major or critical, according to the company's scheme.
- Investigation and root cause: categories such as human error, equipment, material, method or documentation.
- Impact assessment: whether product quality or safety could be affected.
- Corrective and preventive actions, with owners and due dates.
- Approvals, signatures and closure.
The text is partly narrative and partly coded. Root-cause fields, step references and classification give structure; narratives hold the richest detail.
What this data can support, and what it cannot
Deviation reports can support:
- Exception reasoning. Examples of why a step could not be done as written, and what the operator did instead.
- Failure cause taxonomies. Coded root causes across many events show common failure types.
- Procedure-gap analysis. Repeated deviations at one step flag unclear or impractical instructions, related to procedural drift and tacit knowledge gaps in SOPs.
- Evaluation sets for planners that must recognize when to stop, escalate or take a documented alternative.
- Training language models to read and classify quality events, an adjacent use that does not involve robots.
They cannot support:
- Physical control. A report states that a stopper was seated crooked; it contains no trajectory or force data on how.
- Perception. Images appear only if the supplier attached them, and those may be controlled records.
- Representative failure frequency. Reports capture what was noticed, classified and written up, not all that went wrong.
- Robot failure data. Human process deviations differ from robot execution failures. Datasets such as RoboFail, introduced with REFLECT, hold failed executions in simulation and real robots with multisensory data for explanation and correction (Liu et al., 2023); a deviation report is a text record of human work and is not a substitute.
- Blame-free candor. Wording is shaped by compliance review and may understate or generalize.
Illustrative deviation record
# Illustrative - synthetic values, redacted style
deviation_id: dev_000731
area: area_C
step_ref: {sop_id: sop_0118, revision: "F", step: "9.2"}
event_type: step_out_of_sequence
description: "Hold-time check completed after transfer rather than before"
detected_by: reviewer_role
classification: minor
root_cause_category: method # method | material | equipment | human | documentation
root_cause_text: "Step wording ambiguous on whether check precedes transfer"
immediate_action: "Material held; QA review requested"
impact_assessment: "No product impact identified (supplier-stated)"
capa: {action: "Clarify step 9.2 wording", status: closed}
attachments: [] # photos only if retained and licensable
removed: [product_name, batch_number, names, customer_ids]
Where reports come from
| Holder | What they keep | Access notes |
|---|---|---|
| Pharmaceutical and biologics manufacturers | Deviation and CAPA records in quality systems | Highly confidential; GMP context |
| Medical device makers | Nonconformance reports | Customer and design sensitivity |
| Food and beverage processors | Corrective action and nonconformance logs | Customer audit linkage |
| Aerospace and defense suppliers | Nonconformance and disposition records | Export-control and customer restrictions |
| Contract manufacturers | Records per client | Often owned by the client |
Platform-based execution logs, covered in connected-worker platform exports, often hold the exceptions raised at the moment of work, while formal deviation reports hold the reviewed write-up. The inspection side, such as defect imagery tied to nonconformance, is covered in the inspection hub and the catalog entries for manufacturing quality records.
Permissions and provenance
This is a sensitive asset. Reports may name customers, suppliers, products, lots and employees, and may describe quality failures the company would rather not circulate. Contract manufacturers often hold records on behalf of clients who own them. Regulatory submissions or inspections may reference some events. Ask: who owns the record, is it the supplier's or its client's; do contracts forbid disclosure; can product and customer identifiers be removed without gutting the narrative; and is there any pending investigation. Removing names does not settle who may license the record. This is not legal advice; involve counsel.
Evaluating an offered set
Check coverage of root causes: a set dominated by documentation errors teaches less than a mixed one. Check step references: do reports cite the SOP and revision, so they can be paired with the procedure? Check narrative depth: one-line entries ("operator error") carry little. Check classification consistency within and across sites. Check redaction artifacts: heavy masking can leave sentences that no longer make sense. Check time span, since procedures change; pairing reports with the right revision is key (see SOP revision histories). For regulated procedure types in general, see regulated procedures.
Limitations and additional preparation
- Text-only by default. Linking to images or batch records is extra work and extra rights.
- Domain concentration. Most accessible volume is pharmaceutical and food manufacturing; transfer to warehouses or service work is uncertain.
- Heavy redaction. Expect sections withheld; the removed-field list should accompany the delivery.
- Skewed sample. Companies may release only minor or closed events.
- Interpretation needed. Root-cause categories differ between companies; mapping to a common taxonomy requires reviewers who know quality systems.
- Small human subset. Many deviations concern paperwork or timing, not physical manipulation; filter by event type to estimate usable volume.
Questions to ask suppliers
- How many reports involve physical handling steps versus documentation or timing?
- Do reports cite the SOP, revision and step?
- Which root-cause scheme is used, and is it consistent across sites?
- Who owns the records: the supplier, a client or both?
- What fields were removed, and does the narrative remain coherent?
- Are photographs or batch records attached, and can they be licensed?
- What period and how many sites are covered?
- Are open or under-investigation events excluded?
- Do any contracts restrict use of quality records in model training?
Using reports to build exception evaluations
One practical use is an exception-handling evaluation set. Take a procedure step and a de-identified deviation narrative, and ask a planner what it should do: continue, stop, escalate, or take a documented alternative. The company's recorded disposition is the reference answer, which makes the item checkable. Keep three cautions in mind. The disposition reflects that company's quality system and risk appetite, not a universal rule. Some dispositions depend on information that redaction removed. And items from one site's reports should be held out from training when evaluating generalization, split by site or by procedure family, not by individual report, so near-identical events do not appear on both sides.
Discuss your requirements
If you want records of how and why procedures were not followed, describe the industry, event types and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. Whether regulated manufacturers will agree is not known in advance, and a request does not guarantee a match; SourceX follows up about potential sources. See the humanoid training data guide and the procedures hub for related sourcing routes.
Tell SourceX what your evaluation set needs to cover.
Sources
- 21 CFR 211.100 Written procedures; deviations — procedures followed and documented; deviations recorded and justified.
- REFLECT: Summarizing Robot Experiences for Failure Explanation and Correction — failure explanation with multisensory robot data and the RoboFail dataset.