Cleaning has its own lifecycle.
Six stages, end-to-end. The agent runs every one against your SOP-CLN library, your toxicology dossier, and your equipment train.
Health-based limits
PDE/ADE per ICH Q3C/Q3D, toxicology dossier ingested.
MACO computed
Worst-case product, worst-case train, worst-case batch sequence.
Validation protocol
Drafted against your SOP-CLN library; sampling plan included.
Swab recovery
Recovery factors per surface material, per analyte, per cleaning agent.
Three consecutive runs
Executed under protocol; deviations triaged inline.
Cleaning Validation Report
Authored, reviewed, signed. Periodic review scheduled.
Six capabilities, one per stage. Same trace graph.
Every stage is an agent-authored, human-approved workflow. The output of one stage is the structured input of the next - no rekey, no Word round-trips.
Health-based limits, derived from toxicology - not from history.
The agent ingests the toxicology dossier and computes PDE per ICH Q3C (residual solvents) and ICH Q3D (elemental impurities), with NOEL and modifying factors recorded as structured data. Where the customer has an existing HBEL library, the agent pulls from there; where there are gaps, it surfaces them with citation back to the underlying toxicological study.
Outputs: a per-product HBEL with every input attributable, a delta against the prior PDE if one existed, and a queueable review task for the toxicology lead. EMA Q&A on Risk-based Prevention of Cross-Contamination is the operating reference; the calculation is not a black box.
MACO across the equipment train.
Worst-case product, worst-case batch sequence, worst-case shared surface. MACO is recomputed on every product introduction, formulation change, or API potency revision - not at the annual review. The platform flags whether the existing cleaning validation still holds or whether a delta study is required.
The output is a defended worst-case selection: which product sets the limit, on which train, against which following product, with the calculation auditable end-to-end. The next batch sequence never “just got missed” in the worst-case analysis.
Protocol authored against your SOP-CLN library.
The cleaning validation protocol is drafted against your existing SOPs - same template, same numbering, same voice. The agent generates the sampling plan: swab locations on the P&ID, rinse points, recovery factors, acceptance criteria traceable to the underlying HBEL.
The protocol is reviewer-assistive, never autonomous. QA owns the approval signature. The draft is graph-linked back to the limits and the MACO - change either and the protocol surfaces as impacted.
Swab recovery, per surface, per analyte.
Recovery factors are captured per surface material (SS316L, EPDM, PTFE, glass), per analyte, and per cleaning agent. The agent runs the recovery study design and assembles the recovery report per EU GMP Annex 15 §6. Where literature recovery exists for a comparable matrix, it is cited but not relied on - site-specific recovery is the default.
The recovery factor is then a live input to acceptance criteria - not a number embedded in a PDF that nobody can find next year.
Three consecutive runs, executed.
Confirmatory runs are scheduled against the validated protocol. Operators on the floor work through the executable script; results stream back to the trace graph in real time. Deviations are triaged inline by the agent - root cause drafted, immediate impact assessed, CAPA proposed.
Equipment status (Clean / In use / Dirty / Under maintenance) is exposed back to the MES so the next batch does not start in the wrong state. Operator records are ALCOA+ at write time and again at review time.
Cleaning Validation Report. Authored, reviewed, signed.
The Cleaning Validation Report is drafted against your template. Acceptance criteria are traceable to the underlying HBEL calculation; the auditor can drill from the report back to the toxicology source in two clicks. Periodic review is scheduled with the delta-study triggers maintained as live data on the graph.
What survives the inspection: every PDE input, every MACO computation, every swab recovery study, every visual inspection record - one trace graph, one audit trail, one defensible chain from API toxicology to the cleaning release decision on the floor.
Every new product shifts the worst-case. The platform shifts with it.
Every new product introduction, every formulation change, every API potency revision can shift which molecule sets the MACO across the equipment train. Solubility, daily dose, toxicology - any one of them moves the worst-case. The platform re-runs the analysis on every change and flags whether the existing cleaning validation still holds.
When the worst-case shifts, the platform scopes a delta study - not a full revalidation. Additional swabs at the highest-risk locations, recovery confirmation for the new analyte, one or three confirmatory runs depending on risk class. The historical cleaning validation remains valid for the products it already covers; only the gap is closed.
The trace graph shows which products are covered by which study, on which equipment, under which SOP revision. When the EMA inspector asks “under what cleaning validation does this oncology product release on this suite?”, the answer is one query - not a four-hour search across three SharePoints.
The standards we operate against. Cited, not implied.
Q&A on Risk-based Prevention of Cross-Contamination
Health-based exposure limits as the operating standard. 2014, periodically updated.
EudraLex Annex 15
Qualification & Validation (2015). Section 6 on cleaning validation as the spine of the lifecycle.
Annex 1 (2022)
Contamination control strategy and aseptic-process implications for cleaning of sterile lines.
ICH Q3C / Q3D
Residual solvents and elemental impurities. PDE derivation from toxicology.
ALCOA+
Applied per record: attributable, legible, contemporaneous, original, accurate, and the plus-four.
Part 11
Electronic signatures on every cleaning execution record and the signed Cleaning Validation Report.
Five questions a Head of Cleaning Validation will ask.
Where a public toxicology dataset is incomplete - common for early-development APIs, biologics intermediates, and orphan-indication molecules - the agent flags the gap and routes the PDE derivation to your toxicology lead with a structured brief: which endpoint is missing, what surrogate data exists in the corporate dossier, what literature is comparable. The PDE is not invented; it is assembled with citation.
If the sponsor commits to a default-uncertainty-factor approach (per EMA Q&A appendix) for an interim limit, the platform records the decision, the rationale, and the trigger to revisit when better data lands. ALCOA+ on the derivation itself, not just on the records that use it.
The MACO computation is a graph traversal across product, train, and batch-sequence permutations. At 40 products on a shared suite, the platform evaluates every adjacency the campaign plan allows and ranks them by stringency. The current worst-case is the top of that list; the second- and third-most stringent are surfaced as candidates if the worst-case product is removed from the campaign.
When a 41st product is introduced, the analysis re-runs end-to-end in minutes - not in a quarterly engineering review. If the new product becomes the worst-case, the platform scopes the delta study; if not, the existing cleaning validation is annotated to cover the new molecule and the trace graph reflects the additional coverage.
Visual inspection is a defined acceptance criterion, not a culture. The agent assembles the visual inspection record with the operator identity, the lighting condition, the time elapsed since cleaning, the surfaces inspected, the residue limit (visual residue limit per Forsyth, where the customer has qualified the limit), and the disposition.
Where the visual residue limit has not been qualified for a given surface and product, the platform flags the gap rather than treating “visually clean” as self-evident. EMA and FDA expectations on visual inspection have hardened over the last five years; the record must be defensible, not assumed.
Yes. CIP and COP are different worlds for validation. CIP is a controlled cycle - flow, temperature, contact time, conductivity - with cycle data captured by the skid and consumed as ALCOA+ records. COP relies on operator technique, parts management, and post-cleaning storage; the records are operator-attributable and procedure-bound.
The agent runs both, but the protocols, sampling plans, and acceptance criteria differ. COP protocols add operator-qualification linkage and dirty/clean hold-time controls. CIP protocols add cycle-data integration and parametric release where appropriate. The trace graph keeps the distinction explicit; inspectors do not have to ask.
Dedicated equipment removes the cross-contamination MACO problem but does not remove the validation work - degradation residues, cleaning-agent residues, and microbial limits still need defending. The platform handles dedicated equipment with a streamlined protocol scope and the same trace-graph posture.
Shared equipment trains carry the full MACO analysis. Where dedication is partial - for example, dedicated tubing on a shared vessel - the platform models the dedication boundary explicitly and computes MACO only across the shared surfaces. Half-measure dedication is the most common source of cleaning-validation findings; making the boundary an explicit graph object is the fix.