LabOps

LabOps (laboratory operations) is the work that keeps a laboratory able to produce results that can be trusted: receiving and tracking samples, keeping instruments qualified, calibrated and booked, keeping reagents and consumables in stock and in date, capturing data so that it is attributable and unaltered, and running the quality system that proves all of this to a customer, auditor or regulator.

In one line: LabOps answers “for this result, can we show which sample it came from, who handled it, on which working instrument, with which in-date reagents, and that nobody changed the record without a trace?”

This Atlas’s emphasis is integrity from the physical sample to the digital record; facilities, safety and purchasing appear only where they touch that chain. For lab managers and LabOps staff, quality assurance and quality control (QA/QC), laboratory informatics staff, scientists, and engineers who connect instruments to software. No prerequisites.

Science often fails in the lab before it fails in the analysis, in mundane and expensive ways:

• Wrong sample, right answer. A swapped label makes a perfect measurement belong to someone else. In clinical laboratory diagnostics, pre-analytical errors (collection, transport, preparation) are estimated at up to 70% of all mistakes, varying widely by study design.^[9]
• Status believed, not proven. In June 2014 a CDC high-containment lab sent B. anthracis samples to lower-containment labs after a procedure that may not have inactivated them; colleagues handled them believing they were. CDC’s review cited a study plan not followed and no SOPs on when agents are properly inactivated. CDC judged exposure “extremely unlikely”, no one became ill, and it halted transfers out of its biosafety level 3 and 4 (BSL-3/4) labs.^[10]
• Records that cannot be trusted. In 2019 the FDA warned a contract testing lab about access to its high-performance liquid chromatography (HPLC) system without employee-specific logins, a chromatography audit trail it had not enabled (repeating a 2017 observation) and incomplete records.^[11] Such results may be correct and still unusable, and customers relying on the lab inherit the weakness.
• Quiet waste. Expired reagents, double-booked instruments, overdue calibrations found on audit day.

Before: personal spreadsheets, paper booking, calibration stickers, results typed from printouts, a yearly audit sprint. After: one sample register with custody, booking that refuses lapsed instruments, data captured with an audit trail, and a weekly exception report for QA.

• Not “doing the science”. Methods and study design belong to scientists; LabOps makes their environment reliable and their records defensible.
• Not buying a LIMS. ASTM’s laboratory-informatics guide treats LIMS, laboratory execution systems (LES), laboratory information systems (LIS, the clinical term), ELN, SDMS and CDS as a landscape with a life cycle, not one product.^[12]
• Not BioOps. Once data leaves the instrument and enters an analysis pipeline, reproducibility of computation is BioOps’s job. The handoff is a checked file plus sample metadata (see Lab 2).
• Not HealthOps. Patient flow, electronic health records (EHR) and clinical operations belong to HealthOps; this page covers the lab’s own operation, including medical labs under ISO 15189. No regulator cited here uses “LabOps”.

The practice is old; the label is new. US Good Laboratory Practice (21 CFR Part 58) was first published in 1978; today it requires documented equipment maintenance and calibration, labelled and in-date reagents, and handling that keeps test articles identified.^[13] The OECD GLP Principles and laboratory accreditation (ISO Guide 25, later ISO/IEC 17025) grew up in parallel from the 1970s.^[14] Electronic records brought 21 CFR Part 11 in the US and EU GMP Annex 11 (current revision January 2011).^[15][16] Data-integrity guidance then made ALCOA (and ALCOA+) the common vocabulary: the UK MHRA (2018), FDA (2018), the Pharmaceutical Inspection Co-operation Scheme (PIC/S, 2021), WHO (2021) and the OECD for GLP (2021).^{[17][18][19][20][21]}

The “-Ops” label comes from industry; uses found date from about 2021.^[3][1][4] It signals a shift in emphasis, not a new discipline: treating the lab as a system with services, data and metrics.

Mental model: two ledgers that must agree. The physical ledger is where the tube is and in what condition; the digital ledger is what the systems say about it. LabOps keeps them in step at every handover and proves it afterwards. Most lab failures are a disagreement between them that nobody noticed in time.

flowchart TB
  subgraph U["Inputs and support"]
    direction TB
    P["ProcurementOps: reagents, instruments"] ~~~ C["ComplianceOps: obligations"] ~~~ I["IoTOps: sensor fleet"]
  end
  U --> L["LabOps: sample to record"]
  L --> B["BioOps: pipelines"]
  L --> H["HealthOps: care"]
  B --> R["ResearchDataOps: publish"]
  classDef core fill:#e8f3f2,stroke:#16756f,color:#122029
  classDef neutral fill:#f1f4f6,stroke:#5b7a87,color:#122029
  class L core
  class P,B,R,H,C,I neutral

LabOps ends where instrument data enters the analysis pipelines: an editorial distinction of this Atlas, not an industry standard.

The following charter is this Atlas’s proposal; no official LabOps manifesto was found.

Mission: every result this lab releases can be traced to its sample, method, instrument, materials and people, and was produced on time by a process we can show to an outsider.

1. Identity first. No unique, durable identity, no test. Identity travels with the material and its data.
2. Status is evidence. “Calibrated”, “inactivated”, “in date”, “released” are records with an owner and a date, checked by the system at the point of use.
3. Record at the moment of work. Instrument capture beats typing; typing at the bench beats transcribing later.
4. Never overwrite. Corrections are new entries with a reason; GLP and Part 11 require that changes not obscure the original.^[13][15]
5. One person, one account, so every entry is attributable.^[18][17]
6. Review by exception, at the cadence of the data.^[18]
7. Independent quality. The approver is not the person whose work is approved (in US GLP, a QA unit separate from study personnel).^[13]
8. Measure flow and integrity together. A faster lab with weaker records is not better.

Non-goals: designing experiments, owning purchasing contracts, running analysis pipelines, giving legal advice.

Titles vary by lab; the separation of duties does not.

R = responsible, A = accountable, C = consulted, I = informed. The FDA places audit-trail review with the people responsible for record review.^[18] Who signs releases depends on the regime; the release row fits a GMP QC lab. An ISO/IEC 17025 lab nominates the people authorised to release results (Australia’s accreditation body, NATA, requires it),^[26] so there that signatory, not QA, is usually accountable (our reading). In a small lab one person may hold several columns; keep at least the analyst/reviewer split and the QA independence.

stateDiagram-v2
  [*] --> Expected: submission form
  Expected --> Received: arrives, logged
  Received --> Rejected: fails acceptance
  Received --> Accessioned: unique ID, label
  Accessioned --> Stored
  Stored --> InTest: checked out
  InTest --> Stored: returned
  InTest --> Consumed: used up in test
  InTest --> Reported: result released
  InTest --> Quarantined: doubt in testing
  Stored --> Quarantined: excursion or doubt
  Quarantined --> Stored: QA release
  Quarantined --> Disposed: QA decision
  Reported --> Retained: retention period
  Retained --> Disposed: authorised
  Rejected --> Disposed
  Consumed --> [*]
  Disposed --> [*]

The lab-system lifecycle around it follows Day 0 to Day 2:

• Day 0 (design): rules that apply; sample IDs, acceptance criteria, statuses, custody events; systems; requirements and a risk-based validation plan.
• Day 1 (go-live): qualify instruments, migrate or freeze legacy records, train, run in parallel.
• Day 2 (operate): calibration, weekly exception review, periodic evaluation of computerised systems,^[16] audits.

Escalation: anything that may affect a released result goes to QA the same day and may trigger customer notification; MHRA expects significant data-integrity incidents to be notified to regulators where appropriate.^[17]

A logical architecture: small labs collapse boxes into one product, large ones split them. Each fact has one home, and the boxes exchange identifiers rather than retyped values.

flowchart TB
  I1["Networked instruments"] --> D["Device interfaces: SiLA 2, LADS, APIs"]
  I2["Standalone instruments, printouts"] --> LIMS["LIMS: samples, tests, results"]
  S1["Freezer and room sensors"] --> M["Monitoring and alarms"]
  D --> CDS["CDS / SDMS: raw data, audit trail"]
  M --> QMS["QMS and asset register"]
  QMS -. "blocks lapsed instruments" .-> LIMS
  CDS --> G["Release gate"]
  LIMS --> G
  G --> BIO["BioOps pipelines"]
  LIMS --> REP["Reports to customers or EHR"]
  CDS --> ARC["Archive and backup"]
  classDef core fill:#e8f3f2,stroke:#16756f,color:#122029
  classDef neutral fill:#f1f4f6,stroke:#5b7a87,color:#122029
  classDef decision fill:#fdf3e3,stroke:#c9800f,color:#122029
  class LIMS,CDS,QMS core
  class G decision
  class I1,I2,S1,D,M,BIO,REP,ARC neutral

Component responsibilities and trade-offs:

• Integration layer as trust boundary. Instrument PCs often run old operating systems tied to validated software, where every patch needs change control and an impact assessment, and sometimes re-qualification; segment them behind the integration layer. Standards reduce custom drivers: SiLA 2 runs over HTTP/2 with Protocol Buffers and the gRPC wire format;^[27] LADS is a device-type-agnostic OPC UA information model for laboratory devices;^[28] Allotrope defines standardised laboratory data formats, including JSON-based Allotrope Simple Models it describes as human- and machine-readable.^[29]
• One home per fact (LIMS: sample status; CDS: raw data and audit trail; asset register: calibration), and keep dynamic records dynamic: archive native files, not just PDFs.^[18]
• Alternatives: one “lab platform” reduces integration work but increases lock-in and validation scope; best-of-breed tools need a stronger integration layer.

flowchart TB
  A["Sample arrives with submission form"] --> B{"Meets acceptance criteria?"}
  B -- no --> X{"Customer accepts testing under a documented deviation?"}
  X -- no --> XR["Reject: return or dispose, inform customer"]
  X -- yes --> C
  B -- yes --> C["Accession: unique ID, label, store"]
  C --> D["Test on a qualified, calibrated instrument with in-date materials"]
  D --> E["Raw data and audit trail captured"]
  E --> F{"Second-person review: complete and within specification?"}
  F -- yes --> R["Release result and report"]
  F -- "OOS or anomaly" --> G["Phase I: laboratory investigation, keep preparations"]
  G --> H{"Clear laboratory error?"}
  H -- yes --> I["Invalidate with documented justification, retest per plan"]
  I --> D
  H -- no --> J["Phase II: full-scale investigation by the quality unit"]
  J --> K["Decision and report with investigation outcome"]
  classDef core fill:#e8f3f2,stroke:#16756f,color:#122029
  classDef risk fill:#fbeceb,stroke:#b4332b,color:#122029
  classDef decision fill:#fdf3e3,stroke:#c9800f,color:#122029
  class A,C,D,E,R core
  class B,F,H,X decision
  class XR,G,I,J,K risk

The exception path follows the US FDA’s OOS guidance: every OOS result is investigated; Phase I assesses the lab’s data, if possible before preparations are discarded; a contract lab passes findings to the manufacturer’s quality unit, which runs Phase II if no clear lab error is found; retests are predefined, not adjusted to results; and “Laboratory error should be relatively rare”.^[25] Invalidation needs a documented, scientifically sound justification, and “testing into compliance” is not consistent with CGMP.^[18] Labs outside pharmaceuticals are not bound by this guidance, but its logic (decide the retest rule before seeing the retest) is a sound default.

Owners are in the RACI and the checks in the metrics; the playbooks give the steps.

1. Use identifiers unique beyond your lab. The AMP/CAP pipeline guideline notes that probably more than 100 institutions have a case number “S17-123”, and for clinical next-generation sequencing (NGS) recommends at least four identifiers (sample, patient, run, laboratory location) in each data file and its name.^[30] Elsewhere: sample ID plus run ID plus site.
2. Label at receipt with a 2D barcode and readable ID, on stock that survives your storage (cryogenic, solvent).
3. Model aliquots as children of the parent sample, so S-0001-A2 traces to S-0001.
4. Record a custody event at every handover (from, to, where, when, by whom); a handover with no event is a gap.
5. Scan, don’t read, at the instrument. Later genetic-fingerprint checks against sequencing data^[30] belong to BioOps.
6. Quarantine on doubt until QA decides.

• Individual accounts wherever data is created or changed; shared read-only accounts only for viewing.^[18] If the system supports individual users, use it even if that means buying licences.^[17]
• Audit trails on for creation, change and deletion.^[15][16]
• Keep originals: a balance or pH-meter printout can be the original record; a spectral file is dynamic and stays electronic.^[18] Issue numbered blank forms and reconcile them.^[18]
• Review audit trails with their records, at the same frequency, or by risk where rules are silent.^[18] MHRA accepts exception reporting: a validated search for predefined abnormal data or actions.^[17]

• Qualify before use (see below), calibrate on justified intervals and review them; ILAC G24 gives methods.^[31]
• Name the responsible person in maintenance SOPs and record routine and non-routine work, including how the defect was found and what was done (21 CFR 58.63 requires this in US GLP).^[13]
• Block, don’t remind: the LIMS or booking system refuses an instrument whose calibration has lapsed.

“Validated” means documented evidence that an instrument or system does what your intended use needs, scaled to risk.

• Computerised systems. EU GMP Annex 11 bases the extent of validation on a justified, documented risk assessment, wants user requirements traceable through the life cycle, and has users review supplier documentation for off-the-shelf products; bespoke or customised systems need formal evidence for every life-cycle stage.^[16] In practice: test your configuration, workflows, calculations and interfaces, rely on assessed supplier evidence for the standard product, and do not re-test the vendor’s whole product.
• Analytical instruments. USP general chapter <1058> covers analytical instrument qualification; its official text dates from 2017. A restructured revision, “Analytical Instrument and System Qualification”, was proposed in Pharmacopeial Forum 51(2) (comments until 31 May 2025)^[32] and was still under discussion at a USP workshop in June 2026.^[33] Check the current USP–NF.
• Manufacturing qualification. EU GMP Annex 15 defines DQ, IQ, OQ and PQ for facilities, equipment, utilities and processes used to manufacture medicines;^[23] QC labs borrow its vocabulary.
• Methods are separate. Method validation and measurement uncertainty (ISO/IEC 17025 clauses 7.2 and 7.6)^[22] are scientific work; LabOps keeps the records and the instruments they depend on.

Re-qualify or assess? Every change goes through change control, and a documented risk assessment decides the testing; Annex 15 uses quality risk management to plan any validation, verification or requalification a change needs.^[23] Editorial rule of thumb: re-test the affected functions when a change could alter what the instrument measures or what the system calculates, stores or signs (a repaired detector, a move that disturbs alignment, a new software version); otherwise record the assessment and why no testing is needed.

Label reagents with identity, concentration, storage and expiry, and never use outdated material (21 CFR 58.83 in US GLP).^[13] Record lots against the tests that used them, issue first-expiry-first-out, record opened dates, and set reorder points from consumption. Purchasing belongs to ProcurementOps.

The obligations decide the controls. Orientation only, not legal advice; check current texts for your jurisdiction.

Three cautions:

1. GLP and ISO/IEC 17025 are not interchangeable (see the comparison).
2. Revisions are in flight. On 23 October 2026 the European Commission’s GMP page still listed Annex 11 as the January 2011 revision and no annex beyond Annex 21;^[40] drafts of a revised Annex 11, Chapter 4 and a new Annex 22 on artificial intelligence (AI) were consulted on from 7 July to 7 October 2025.^[41] Re-check before relying on either.
3. Clinical labs have their own regimes (CLIA in the US, national rules elsewhere, often ISO 15189 accreditation); patient privacy is covered on PrivacyOps and HealthOps.

Choose the category before the brand. Examples were checked on 2026-10-23 and are illustrations, not endorsements.

Questions that matter more than features: can it prove who did what, when and why; can you export everything, audit trails included; and how long does a validated configuration change take?

Lab automation (liquid handlers, plate movers, schedulers) moves the integrity problem rather than removing it: the robot’s log joins the record, jam recovery becomes a procedure, and unattended runs need alarms that reach a person.

AI helps around the record more than inside it: drafting SOP text for review, summarising deviation trends, pointing reviewers at audit-trail anomalies, forecasting consumption. Anything that creates or changes a GxP record or decides a result is risky. The draft EU Annex 22 is not final (caution 2 above); until it is, treat AI tools that touch regulated data as computerised systems that need validation.

Artefacts for an illustrative 20-person testing lab; field names are examples, not a standard.

1. Submission and acceptance record. Custody starts here; acceptance criteria are written in advance.

Example status: YAML syntax-checked by the Atlas validator; fictional data.

submission:
  id: SUB-2026-0412
  customer: Example Pharma Ltd (fictional)
  submitted_by: { name: "R. Patel", signed_at: "2026-10-20T07:40:00Z" }
  samples:
    - customer_ref: BATCH-77-A
      matrix: tablet
      quantity: "20 units"
      tests: [assay-hplc, dissolution]
      storage_required: "15-25 C"
acceptance_criteria:          # decided before receipt, versioned with the SOP
  - container intact and sealed
  - label matches the submission form
  - quantity >= quantity required by the requested tests
  - temperature logger within range for the whole transit
receipt:
  received_by: a.rivera       # a named person, never "receiving"
  received_at: "2026-10-20T08:02:00Z"
  outcome: accepted           # accepted | rejected | accepted_with_deviation
  deviation: null
  lab_id: S-0001

2. Instrument register entry, the source for “may this instrument be used today?”

Example status: YAML syntax-checked; fictional data.

instrument:
  id: HPLC-02
  type: HPLC with UV detector
  location: QC lab, room 2.14
  owner: j.chen
  qualification: { iq: 2025-03-04, oq: 2025-03-06, pq: 2025-03-20 }
  calibration:
    last: 2026-06-30
    interval_days: 184          # justified and reviewed (see ILAC G24)
    due: 2026-12-31
  status: in_service            # in_service | out_of_service | quarantined
  software: { cds_version: "x.y (validated 2025-03)", audit_trail: enabled }
  maintenance_sop: SOP-EQ-014
booking_rule: refuse booking if status != in_service or today > calibration.due

3. Status rules the LIMS enforces (written as requirements for the configuration, reviewed by QA):

All three labs were executed for this page on 2026-10-23 with Python 3.8.10 and SQLite 3.31.1 (standard library only); the expected outputs are copied from those runs. Fictional data; they write only to the current directory.

Objective: see how “never overwrite, correct with a reason” and custody-gap detection work, using a database you can inspect.

Prerequisites: Python 3.8 or later. Cost and risk: none; creates one file, custody.db.

1. Save the script as custody.py.
2. Run python3 custody.py custody.db.
3. Open the database with any SQLite browser and look at the custody_event table and its triggers.

Example status: executed (Python 3.8.10, SQLite 3.31.1).

"""Lab 1: an append-only custody ledger in SQLite (Python standard library only)."""
import sqlite3
import sys

db = sqlite3.connect(sys.argv[1] if len(sys.argv) > 1 else "custody.db")
db.executescript("""
CREATE TABLE IF NOT EXISTS custody_event (
  seq        INTEGER PRIMARY KEY AUTOINCREMENT,
  sample_id  TEXT NOT NULL,
  event      TEXT NOT NULL CHECK (event IN
             ('received','stored','transferred','consumed','disposed','correction')),
  from_party TEXT,
  to_party   TEXT NOT NULL,
  location   TEXT NOT NULL,
  actor      TEXT NOT NULL,          -- a named person, never a shared account
  at_utc     TEXT NOT NULL,          -- ISO 8601, recorded when it happens
  reason     TEXT                    -- mandatory for corrections
);
CREATE TRIGGER IF NOT EXISTS no_update BEFORE UPDATE ON custody_event
BEGIN SELECT RAISE(ABORT, 'custody events are append-only: add a correction'); END;
CREATE TRIGGER IF NOT EXISTS no_delete BEFORE DELETE ON custody_event
BEGIN SELECT RAISE(ABORT, 'custody events cannot be deleted'); END;
CREATE TRIGGER IF NOT EXISTS correction_needs_reason BEFORE INSERT ON custody_event
WHEN NEW.event = 'correction' AND (NEW.reason IS NULL OR trim(NEW.reason) = '')
BEGIN SELECT RAISE(ABORT, 'a correction needs a reason'); END;
""")

events = [
    ("S-0001", "received", None, "Receiving", "Dock-1", "a.rivera",
     "2026-10-20T08:02Z", None),
    ("S-0001", "stored", "Receiving", "Freezer F-07", "F-07/R2/B4", "a.rivera",
     "2026-10-20T08:15Z", None),
    ("S-0001", "transferred", "Freezer F-07", "Chem QC", "QC bench 3", "j.chen",
     "2026-10-21T09:40Z", None),
    ("S-0002", "received", None, "Receiving", "Dock-1", "a.rivera",
     "2026-10-20T08:05Z", None),
    # Gap: S-0002 shows up in QC without a recorded handover from Receiving.
    ("S-0002", "transferred", "Freezer F-07", "Chem QC", "QC bench 3", "j.chen",
     "2026-10-21T09:41Z", None),
]
INSERT = ("INSERT INTO custody_event (sample_id, event, from_party, to_party,"
          " location, actor, at_utc, reason) VALUES (?,?,?,?,?,?,?,?)")
db.executemany(INSERT, events)
db.commit()

try:
    db.execute("UPDATE custody_event SET location = 'QC bench 4' WHERE seq = 3")
except sqlite3.DatabaseError as exc:
    print("blocked:", exc)

# Each event's from_party must equal the previous to_party for the same sample.
gaps = db.execute("""
SELECT e.sample_id, e.seq, p.to_party AS last_holder, e.from_party AS claimed_from
FROM custody_event e
JOIN custody_event p ON p.sample_id = e.sample_id
 AND p.seq = (SELECT max(seq) FROM custody_event x
              WHERE x.sample_id = e.sample_id AND x.seq < e.seq)
WHERE e.event <> 'correction' AND e.from_party IS NOT p.to_party
ORDER BY e.sample_id, e.seq""").fetchall()
for sample, seq, last, claimed in gaps:
    print(f"custody gap: {sample} event {seq}: last holder {last!r}, "
          f"event claims {claimed!r}")
print(f"{db.execute('SELECT count(*) FROM custody_event').fetchone()[0]} events, "
      f"{len(gaps)} gap(s)")

# A late entry is added as a correction with a reason; the original gap stays visible.
fix = ("S-0002", "correction", "Receiving", "Freezer F-07", "F-07/R2/B4", "a.rivera",
       "2026-10-21T10:05Z", None)
for reason in (None, "Late entry: moved to F-07 at 08:20Z per freezer door log"):
    try:
        db.execute(INSERT, fix[:7] + (reason,))
        db.commit()
        print("correction recorded for S-0002")
    except sqlite3.DatabaseError as exc:
        print("blocked:", exc)

Expected output (first run on a new file):

blocked: custody events are append-only: add a correction
custody gap: S-0002 event 5: last holder 'Receiving', event claims 'Freezer F-07'
5 events, 1 gap(s)
blocked: a correction needs a reason
correction recorded for S-0002

Validation: the update and the reason-less correction were refused; the S-0002 gap was found; the reasoned correction was stored. Query SELECT seq, event, reason FROM custody_event: event 5 is still there next to the correction. The record of the gap survives its explanation, which is what an auditor wants to see.

Troubleshooting: a second run appends again (11 events, 4 gaps); delete custody.db to restart. database is locked: another program has the file open.

Limits: the file’s owner can drop triggers or edit the file; real systems add audit trails, access control, separate administrator rights and backups.

Teardown: rm custody.db.

Objective: build the LabOps-to-BioOps handoff: only files from registered samples in test, from instruments in service and in calibration, leave the lab, with checksums.

Prerequisites: Bash and Python 3.8 or later. Cost and risk: none; writes only to ./lab2.

1. Create the practice landing zone with the setup script below (bash setup_lab2.sh).
2. Save the gate as gate.py and run python3 gate.py lab2/drop lab2/lims_samples.csv lab2/instruments.csv 2026-10-23.
3. Read lab2/manifest.json.

Example status: executed (Bash, Python 3.8.10).

#!/usr/bin/env bash
# Creates a practice landing zone with fake instrument output.
# Safe: writes only to ./lab2 in the current directory.
set -euo pipefail
mkdir -p lab2/drop
cat > lab2/lims_samples.csv <<'CSV'
sample_id,status,project
S-0001,in_test,PRJ-A
S-0002,in_test,PRJ-A
S-0003,received,PRJ-B
CSV
cat > lab2/instruments.csv <<'CSV'
instrument_id,calibration_due,status
HPLC-02,2026-12-31,in_service
PLATE-01,2026-10-01,in_service
CSV
printf 'time,signal\n0.0,0.01\n0.1,0.35\n' > lab2/drop/HPLC-02_S-0001_R001.csv
printf 'time,signal\n0.0,0.02\n0.1,0.33\n' > lab2/drop/HPLC-02_S-0002_R001.csv
cp lab2/drop/HPLC-02_S-0002_R001.csv lab2/drop/HPLC-02_S-0002_R002.csv
printf 'well,od\nA1,0.41\n' > lab2/drop/PLATE-01_S-0001_R001.csv
printf 'well,od\nA1,0.38\n' > lab2/drop/HPLC-02_S-0003_R001.csv
printf 'well,od\nA1,0.40\n' > lab2/drop/HPLC-02_S-0099_R001.csv
printf 'junk\n' > lab2/drop/export_final_v2.csv
echo "created $(ls lab2/drop | wc -l) files in lab2/drop"

Example status: executed (Python 3.8.10).

"""Lab 2: a release gate between instruments and analysis (standard library only).

Usage: python3 gate.py DROP_DIR LIMS_CSV INSTRUMENTS_CSV YYYY-MM-DD
Writes DROP_DIR/../manifest.json for accepted files and prints every rejection.
"""
import csv
import datetime as dt
import hashlib
import json
import pathlib
import re
import sys

NAME = re.compile(r"^(?P<instrument>[A-Z]+-\d{2})_"
                  r"(?P<sample>S-\d{4})_(?P<run>R\d{3})\.csv$")


def load(path, key):
    with open(path, newline="") as fh:
        return {row[key]: row for row in csv.DictReader(fh)}


def main(drop, lims_csv, instr_csv, today):
    samples = load(lims_csv, "sample_id")
    instruments = load(instr_csv, "instrument_id")
    day = dt.date.fromisoformat(today)
    accepted, rejected, seen = [], [], {}
    for f in sorted(pathlib.Path(drop).iterdir()):
        if not f.is_file():
            rejected.append((f.name, ["not a regular file"]))
            continue
        digest = hashlib.sha256(f.read_bytes()).hexdigest()
        m = NAME.match(f.name)
        problems = []
        if not m:
            problems.append("name does not follow INSTRUMENT_SAMPLE_RUN.csv")
        else:
            s, i = samples.get(m["sample"]), instruments.get(m["instrument"])
            if s is None:
                problems.append(f"sample {m['sample']} is not registered in the LIMS")
            elif s["status"] != "in_test":
                problems.append(f"sample {m['sample']} has status {s['status']!r}")
            if i is None:
                problems.append(f"instrument {m['instrument']} is unknown")
            elif i["status"] != "in_service":
                problems.append(f"{m['instrument']} has status {i['status']!r}")
            elif dt.date.fromisoformat(i["calibration_due"]) < day:
                problems.append(f"{m['instrument']} calibration overdue since "
                                f"{i['calibration_due']}")
        if digest in seen:
            problems.append(f"identical content to {seen[digest]}")
        seen.setdefault(digest, f.name)
        if problems:
            rejected.append((f.name, problems))
        else:
            accepted.append({"file": f.name, "sha256": digest, "bytes": f.stat().st_size,
                             "sample_id": m["sample"], "instrument_id": m["instrument"],
                             "run": m["run"], "project": samples[m["sample"]]["project"]})
    out = pathlib.Path(drop).parent / "manifest.json"
    out.write_text(json.dumps({"gate_date": today, "files": accepted}, indent=2) + "\n")
    for name, problems in rejected:
        print(f"REJECT {name}: " + "; ".join(problems))
    print(f"accepted {len(accepted)}, rejected {len(rejected)} -> {out}")
    return 1 if rejected else 0


if __name__ == "__main__":
    sys.exit(main(*sys.argv[1:5]))

Expected output (exit code 1 because files were rejected):

REJECT HPLC-02_S-0002_R002.csv: identical content to HPLC-02_S-0002_R001.csv
REJECT HPLC-02_S-0003_R001.csv: sample S-0003 has status 'received'
REJECT HPLC-02_S-0099_R001.csv: sample S-0099 is not registered in the LIMS
REJECT PLATE-01_S-0001_R001.csv: PLATE-01 calibration overdue since 2026-10-01
REJECT export_final_v2.csv: name does not follow INSTRUMENT_SAMPLE_RUN.csv
accepted 2, rejected 5 -> lab2/manifest.json

Validation: lab2/manifest.json lists two files, each with a SHA-256 checksum, sample, instrument, run and project; every rejection names an actionable reason.

Troubleshooting and lessons:

• The duplicate (R002 identical to R001) may be an innocent re-export or a copy presented as a new run; a person decides.
• Try it: set PLATE-01’s status in instruments.csv to quarantined; the rejection reason becomes the status.
• The gate compares calibration with the gate date; a production gate should use the file’s acquisition date, since data acquired before the due date is not invalid because it was uploaded late.
• ValueError: Invalid isoformat string means a date is not YYYY-MM-DD.

Extend it: send the manifest with the files; BioOps verifies the checksums on arrival.

Teardown: rm -r lab2.

Objective: practise risk-based audit-trail review: turn rules into an exception report, then judge each flag.

Prerequisites: Python 3.8 or later; read the data-integrity playbook first. Cost and risk: none. In a real lab, an exception report used for GxP review is itself a tool that must be validated.^[17]

1. Save the extract as audit.csv and the script as review.py.
2. Run python3 review.py audit.csv.
3. For each HIGH flag, write in one line what record you would ask for and who you would ask.

Example status: fictional audit-trail extract.

ts_utc,user,action,record,reason
2026-10-20T09:12:00Z,j.chen,acquire,RUN-0412,
2026-10-20T09:55:00Z,j.chen,reprocess,RUN-0412,baseline drift; SOP-CDS-07 step 4
2026-10-20T10:03:00Z,j.chen,result_entry,RUN-0412,
2026-10-20T11:30:00Z,m.okafor,approve,RUN-0412,
2026-10-20T13:02:00Z,j.chen,result_change,RUN-0412,
2026-10-20T23:41:00Z,qclab,reprocess,RUN-0413,
2026-10-21T00:05:00Z,qclab,delete,RUN-0413-inj2,
2026-10-21T08:20:00Z,j.chen,acquire,RUN-0414,

Example status: executed (Python 3.8.10).

"""Lab 3: exception report for an audit-trail extract (standard library only).

Usage: python3 review.py audit.csv
Flags are prompts for a reviewer's judgement, not findings of misconduct.
"""
import csv
import sys

GENERIC = {"admin", "qclab", "lab", "service", "instrument"}
NEEDS_REASON = {"reprocess", "result_change", "integration_change"}


def review(rows):
    approved = set()
    for r in rows:
        hour = int(r["ts_utc"][11:13])
        rec, act, who = r["record"], r["action"], r["user"]
        if who in GENERIC:
            yield "HIGH", r, "generic account: not attributable"
        if act == "delete":
            yield "HIGH", r, "deletion: check data kept + reason"
        if act in NEEDS_REASON and not r["reason"].strip():
            yield "HIGH", r, f"{act} without a reason"
        if act in NEEDS_REASON and rec in approved:
            yield "HIGH", r, "changed after approval"
        if hour >= 22 or hour < 5:
            yield "INFO", r, "outside staffed hours"
        if act == "approve":
            approved.add(rec)


def main(path):
    with open(path, newline="") as fh:
        flags = list(review(list(csv.DictReader(fh))))
    for level, r, why in flags:
        print(f"{level:4} {r['ts_utc'][5:16]} {r['user']:7} {r['action']:13} "
              f"{r['record']:13} {why}")
    high = sum(1 for f in flags if f[0] == "HIGH")
    print(f"{len(flags)} flags ({high} HIGH) for reviewer follow-up")


if __name__ == "__main__":
    main(sys.argv[1])

Expected output:

HIGH 10-20T13:02 j.chen  result_change RUN-0412      result_change without a reason
HIGH 10-20T13:02 j.chen  result_change RUN-0412      changed after approval
HIGH 10-20T23:41 qclab   reprocess     RUN-0413      generic account: not attributable
HIGH 10-20T23:41 qclab   reprocess     RUN-0413      reprocess without a reason
INFO 10-20T23:41 qclab   reprocess     RUN-0413      outside staffed hours
HIGH 10-21T00:05 qclab   delete        RUN-0413-inj2 generic account: not attributable
HIGH 10-21T00:05 qclab   delete        RUN-0413-inj2 deletion: check data kept + reason
INFO 10-21T00:05 qclab   delete        RUN-0413-inj2 outside staffed hours
8 flags (6 HIGH) for reviewer follow-up

Validation and judgement: six HIGH flags, two INFO. Now judge them, because the script cannot:

• RUN-0412 changed after approval, without a reason: ask analyst and approver for the justification; the approval no longer covers the record.
• RUN-0413 reprocessed and an injection deleted under a generic account at night: not attributable. Check the deleted data still exists, open a deviation, remove the account.
• The 09:55 reprocessing cites an SOP step and is not flagged: reprocessing is not wrong, unexplained reprocessing is.

Troubleshooting: KeyError: 'reason' means the CSV header differs (exports vary by system; map columns first). Convert times to UTC before applying “outside hours” rules.

Teardown: rm audit.csv review.py.

All four scenarios are illustrative; the real cases and standards they point to are sourced.

Scenario A, step by step. (1) Trace one test from receipt to report, listing hand-overs. (2) Barcode items at receipt against written acceptance criteria; record a custody event at each hand-over. (3) Make the register block lapsed instruments; review exceptions weekly. (4) When a calibration is found late, list the results since the last good calibration and compare the as-found calibration data with the tolerance: within it, record “no impact” and why; outside it, handle the results as nonconforming work (clause 7.10)^[22] and decide which customers to tell (step 4 is our method, not the standard’s text).

Pitfalls: in A, auditing your own lab but not your testing suppliers; in B, reading a high invalidation rate as “careful analysts” rather than as a sign of a weak method, instrument or training.

Why median and 90th percentile? TAT distributions are right-skewed, so the TAT literature prefers the median plus a tail measure (outlier rate, or the time by which 90% of results are complete) over the mean; definitions differ in start and end points, so state yours.^[47]

Leading vs lagging: calibrations due in 30 days, audit-trail backlog and low stock lead; TAT, RFT and audit findings lag.

• Least privilege. Analysts without administrator rights on systems holding their data; access matched to job roles, with a history of access levels kept (MHRA).^[17] Contemporaneous records also need synchronised clocks.
• Instrument PCs. Segment them (see the reference architecture), restrict removable media, back up native data, plan replacement. The OECD has GLP documents on IT security (2024) and cloud computing (2023).^[38]
• Retention. Periods come from regulation and contract (US GLP: 21 CFR 58.195, including calibration and maintenance records).^[13] Keep native files readable; test restores (BackupOps).
• Privacy. Label human samples with lab IDs, keep the identity mapping only where needed, and let consent restrictions travel with the sample (PrivacyOps; not legal advice).
• Safety. Biosafety and EHS are separate disciplines; LabOps owns records such as the verified status of material moving between containment levels (the 2014 CDC incident).

Cost drivers are mostly people’s time: transcription, searching, rework, audit preparation, investigations; system costs are licences, configuration, validation and interfaces. Ignore ROI claims you cannot check; measure your own transcription and rework hours before and after a change.

Capacity is usually limited by a few bottleneck instruments and the people qualified to run them; add capacity where the queue forms, not where utilisation merely looks high.

Sustainability is material. An NIH newsletter says lab space can use up to five times the energy of office space per square foot; Cornell cites Harvard’s “more than three times”.^[48][49] Two levers sit with LabOps:

• Cold storage. The NIH newsletter reports that −70 °C instead of −80 °C can cut ultra-low temperature (ULT) freezer energy use by 30%, and that most samples are fine at −70 °C;^[48] confirm stability for your materials first. My Green Lab’s International Freezer Challenge has run since 2017; the organiser reports 137 million kWh saved.^[50]
• Fume hoods, which the same source says can be up to twice as energy-intensive as ULT freezers: closed sashes are measurable.^[48]

Audit your own audit: each quarter, reconstruct five random released results end to end from records alone. If that takes days, the next external audit will hurt.

Illustrative scenario. At 02:10 the monitoring system alarms: freezer F-07 (set point −70 °C) reads −52 °C and rising. It holds 9 racks of customer and study samples.

flowchart TB
  A["Alarm: F-07 above alert limit"] --> B{"On-call acknowledges within 15 min?"}
  B -- no --> C["Escalate to second contact and lab manager"]
  C --> D
  B -- yes --> D{"Temperature still rising?"}
  D -- yes --> E["Move racks to the reserved backup freezer, record times"]
  D -- no --> F["Check door, power, alarm sensor; keep monitoring"]
  E --> G["Mark affected samples quarantined in the LIMS"]
  F --> G
  G --> H["Record peak temperature and duration from the logger"]
  H --> I{"Within stability data for each material?"}
  I -- yes --> J["QA releases samples, note on future results"]
  I -- no --> K["QA decides: discard, test with caveat, inform owners"]
  J --> L["Deviation and CAPA: cause, fix, check effectiveness"]
  K --> L
  classDef risk fill:#fbeceb,stroke:#b4332b,color:#122029
  classDef decision fill:#fdf3e3,stroke:#c9800f,color:#122029
  classDef core fill:#e8f3f2,stroke:#16756f,color:#122029
  class A,C,E,K risk
  class B,D,I decision
  class F,G,H,J,L core

What the diagram does not show: acknowledgement is recorded by the monitoring system (the 15-minute limit is this example’s choice); the backup freezer is reserved in advance, with time out and time in recorded per rack; quarantine happens in the system, not just the room, so nobody tests a sample tomorrow before QA decides; and the logger’s own calibration is checked before its peak temperature is trusted.

• “The LIMS will fix it.” A LIMS automates the process you configure; a messy process becomes a fast messy process with an audit trail. Map the sample lifecycle first.
• Shared “lab” accounts and printed-then-deleted chromatograms. Both save minutes and destroy the record (see the data-integrity playbook).
• Measuring only speed. TAT targets without RFT and invalidation rates reward shortcuts.
• Myth: “ALCOA+ is a different standard from ALCOA.” MHRA says there is no difference in expectations whichever acronym is used.^[17]
• Gold-plating a research lab. Full GxP validation for exploratory work drives scientists to private spreadsheets, worse for integrity than proportionate controls.

Invest in a deliberate LabOps practice when several of these hold: others rely on your results (customers, regulators, clinicians, publications); samples pass through several people or locations; instruments are shared; you are or will be inspected or accredited; or you already lose time to searching, rework and audit preparation.

Keep it light, or do not build a dedicated function, when a small team runs exploratory work whose results never leave the group, with few shared instruments and no external obligations. One accountable lab manager, a sample spreadsheet with unique IDs, an ELN and a calibration list may be enough. Over-building costs validation effort for every change, licence fees, and an administrative layer scientists route around.

Exit conditions: obligations ended, volume dropped, or a partner lab took over testing. Records keep their retention period.

flowchart TB
  A{"Do results leave the group or support regulated decisions?"} -- no --> B{"Shared instruments or many sample handovers?"}
  B -- no --> C["Light: lab manager, ELN, sample sheet with unique IDs"]
  B -- yes --> D["Registry: barcodes, inventory, booking, ELN"]
  A -- yes --> E{"GLP, GMP, clinical or accreditation in scope?"}
  E -- no --> D
  E -- yes --> F{"Repeated, standard tests at volume?"}
  F -- yes --> G["LIMS with audit trails, validated to the rules that apply"]
  F -- no --> H["Validated ELN and registry with GxP controls"]
  classDef decision fill:#fdf3e3,stroke:#c9800f,color:#122029
  classDef core fill:#e8f3f2,stroke:#16756f,color:#122029
  classDef neutral fill:#f1f4f6,stroke:#5b7a87,color:#122029
  class A,B,E,F decision
  class G,H core
  class C,D neutral

For a tool choice within a branch, weight criteria explicitly. An illustrative weighting for a regulated QC lab:

Assumptions: regulated work, 5–10 year horizon, a small informatics team. A research lab would move weight from validation to usability.

This is this Atlas’s model, not an industry standard.

Dependencies: QA sign-off, informatics time, and a sponsor who protects analysts’ time. LIMS selection and validation usually run beyond 90 days; this quarter builds their foundation.

Quick start (first week)

• Named accounts for everyone who creates or changes data; audit trails on wherever available
• Every stored sample has a unique ID and a known location
• Every instrument listed with qualification and calibration status
• Every reagent in use shows identity, concentration, storage condition and expiry

System go-live readiness

• Validation approach approved by QA; requirements traced to executed tests
• Legacy data migrated and verified, or frozen read-only
• Administrators separate from analysts; SOPs updated; users trained; parallel run reviewed
• Backup, restore and native data export tested

Audit and inspection preflight

• Five random results reconstructed end to end from records alone
• Training records match current method versions; periodic system evaluations current
• CAPAs from the last audit closed with effectiveness checks

One-page LabOps charter

LABOPS CHARTER - <lab name>                         Version <n>, approved <date>
Mission     Every released result is traceable to sample, method, instrument,
            materials and people, and is produced on time.
Scope       Receipt to release; storage; instruments; inventory; lab records.
Out         Study design; analysis pipelines; purchasing; EHS programme.
Rules       <GLP / GMP / ISO/IEC 17025 / ISO 15189 / clinical / none> - owner: QA
Owner       Lab manager <name>; QA <name>; informatics <name>
Metrics     TAT median and P90; RFT; custody completeness; calibration on time;
            audit-trail backlog; OOS invalidation rate
Cadence     Daily huddle; weekly exceptions; monthly ops review;
            quarterly management review
Escalation  Anything that may affect a released result -> QA the same day

SOP skeleton: sample receipt and accessioning

SOP-SR-001  Sample receipt and accessioning                     Version <n>
1 Purpose     Establish identity and custody of every sample at receipt.
2 Scope       All samples delivered to <lab>; excludes <...>.
3 Roles       Receiving analyst (R); lab manager (A); QA (C).
4 Materials   Label printer, label stock rated for <storage>, scanner.
5 Procedure   5.1 Check form and container against acceptance criteria.
              5.2 Record receipt time, receiver and condition in the LIMS.
              5.3 Assign lab ID; print and apply label; scan to verify.
              5.4 Store at the required condition; record location.
              5.5 If a criterion fails: reject, or record a deviation and
                  inform the customer before testing.
6 Records     Receipt record, custody events, deviation (if any).
7 References  <rules, customer agreements>

Deviation record

DEVIATION <id>   Opened <date/time> by <name>
What happened    <facts only: what, where, when, which samples/instruments>
Immediate action <containment, quarantine>
Impact           Results affected: <list or "none, because ...">
Root cause       <method used; evidence>
CAPA             <action> - owner <name> - due <date>
Effectiveness    Check on <date>: <criterion and result>
Approved by QA   <name, date>

Quality frameworks side by side. This is the page’s one comparison of GLP and ISO/IEC 17025: the OECD calls them documents with different purposes, and applying one to meet the purposes of the other is “impractical, and in many cases would be inappropriate”.^[14]

Sources for the table: OECD’s position paper on GLP and ISO/IEC 17025,^[14] 21 CFR 58,^[13] the ISO pages,^[34][35] and the FDA’s guidance.^[25][18] Details vary by country.

Is LabOps a job title or a department? Both, where it exists (lab operations managers and associates; “Lab Ops” teams).^[1][5] Often the work is split between the lab manager, QA and informatics without the label.

Do we need a LIMS to be ISO/IEC 17025 accredited? The standard names no product; its clauses cover control of data and information management (7.11) and technical records (7.5).^[22] Editorial view, not from the standard’s text: controlled paper and spreadsheets can work at low volume, but controlling them becomes expensive as volume grows; ask your accreditation body how it assesses them.

Is ALCOA+ a regulation? No. It is a widely used summary of data-integrity expectations that appears in regulators’ guidance (MHRA, PIC/S) and is built on legal requirements such as 21 CFR 58.130(e) and Part 11 in the US.^{[17][19][13][15]}

Where exactly does LabOps hand over to BioOps? At the release gate: a checked instrument file plus sample metadata (Lab 2). Sample identity is checked on both sides.

Are research labs subject to GLP? US GLP defines a nonclinical laboratory study as one that determines the safety of a test article, and excludes basic exploratory studies.^[13] Many research labs fall under no GxP rule, though funders and partners may still expect traceable records.

• Established, unevenly adopted: standard instrument interfaces. SiLA 2 version 1.1 added a server-initiated connection for cloud scenarios;^[27] LADS brings lab devices to OPC UA.^[28] Vendor coverage of your instrument mix is the limit.
• Maturing: standardised data formats (Allotrope Simple Models in JSON, the .eln exchange format) that aim to keep records usable beyond one product.^[29][43] One interface for mixed fleets, as vendors describe it, is a claim to test on your instruments.^[46]
• Pending: EU Annex 11, Chapter 4 and Annex 22 (caution 2); USP <1058> (qualification).
• Speculative: autonomous “self-driving” labs in routine regulated testing. Attribution, custody and review do not disappear when the operator is software.

Communities such as LabOps Unite run webinars and chapters;^[1] credentials are in the note on the label.

Situation (illustrative). A 40-person biotech runs research assays and plans GLP studies in about 18 months. Samples live in spreadsheets, instruments are booked on a whiteboard, audit trails are off on two instruments, and there is budget for “a LIMS this year”.

Option A: select and configure a LIMS now. Attractive: one project, ready for GLP. Risks: GLP requirements are not yet known, so today’s research workflow is configured and later re-configured and re-validated; the project absorbs the two people who would fix basic controls; scientists resist a heavy system for exploratory work.

Option B: fix the foundation now, select the LIMS in about 9 months. Named accounts and audit trails now; IDs, barcodes and custody events in a registry; an instrument register; an ELN. Then write LIMS requirements from the GLP study plans. Risks: two migrations; the registry may become a permanent workaround; the GLP start may slip if selection starts late.

Assumptions that decide it: (1) GLP work will differ from current research workflows; (2) basic attribution failures are the biggest current risk; (3) the informatics team is two people; (4) the GLP date is uncertain by ±6 months.

Reasoning. Under these assumptions, B reduces the largest risk (unattributable, unaudited records) within weeks, and makes the LIMS project cheaper because requirements will be known. If the GLP studies were starting within six months, or the registry work could not be staffed, A would win, with a narrower first release.

Uncertainty. No sourced benchmark prices reconfiguration in a given product; estimate it with vendors on your own workflows.

After this page you should be able to:

• Explain ALCOA+ and chain of custody with examples from your lab.
• Select controls, and a validation depth, from the rules and risks that apply.
• Design a sample lifecycle, a release gate and an exception review.
• Operate a freezer excursion or OOS investigation without losing the record.
• Evaluate a lab’s maturity and choose the next three improvements.

1. A colleague says: “We print the HPLC chromatogram, sign it and file it, so we can delete the files to save space.” What is wrong?

2. Your TAT median is 3 days against a 5-day target, but customers complain. What do you measure next, and why?

3. 11 of 18 OOS results last quarter were invalidated as laboratory error. Is that a sign of careful investigation?

4. A sample appears in the QC lab with no record of leaving receiving. The result is already released. What do you do?

5. Your lab is ISO/IEC 17025 accredited. A client asks whether that means you can run their GLP study. How do you answer?

6. Two vendor patches arrive for your validated LIMS: one changes report fonts, the other changes how a calculated result is rounded. Do both need re-testing?

LabOps is a small word for an old discipline. The label belongs to a community of practitioners and a few vendors; the practice belongs to every laboratory whose results someone else relies on. Its core is unglamorous: sample identity and custody, status the system checks, data never overwritten, review by exception. Name the rules that apply, validate in proportion to risk, and do not gold-plate the rest.

The test of a laboratory’s operations is not how fast it produced a result, but how quickly and convincingly it can show, months later, where that result came from. A lab that can reconstruct a result in an hour has good LabOps, whatever it calls them.

All sources were opened and checked on 2026-10-23. Where a site refused automated access, the archived copy that was read is named. Evidence ledger: research/labops-ledger.md.

The label and its community

Data integrity, records and computerised systems

Laboratory quality systems and metrology

Integration standards and tools (change-sensitive)

Measurement, errors and sample identity

Cost and sustainability

Documented incidents

Further reading: the full texts of ISO/IEC 17025, ISO 15189 and ISO 20387 (from ISO or national standards bodies), and OECD GLP documents No. 17 and No. 22 via the BfR index above.

Changelog

• 2026-10-23 — r0: first draft (term status confirmed niche; IT test-lab collision added; three labs executed; 43 sources checked). Awaiting independent review.
• 2026-10-23 — r1: revised after independent review (reviews/labops-r1-response.md): proportionate qualification and validation added; core acronyms defined; repetition cut; Allotrope and Freezer Challenge re-sourced; RACI, Lab 2 gate and three diagrams corrected; credentials, MAD adherents and ISO 20387 revision added; labs re-run.
• 2026-10-23, post-score polish: OECD No. 18 quoted exactly; ISO/IEC 17025 release accountability (NATA); scenario A method with a late-calibration impact check; impact window (“since the last good calibration”) made consistent across the page.