Operational excellence library FDA inspection data Established methods Simulated and illustrative datasets

FDA's most frequent supplement CGMP citations are for missing specifications and records, the controls these methods exist to sustain

Eight operational-excellence methods, each with equations, a worked example and its limits: what each moves (yield, throughput, deviations, cost, release) and what it cannot establish.

Status. EUTECNIQ has no plant or process data. Every dataset is labeled sourced (FDA), simulated (seeded) or illustrative (invented).

Specifications, records and identity tests dominate FDA's dietary-supplement CGMP citationsFDA inspection data

1,938Form 483 citations under 21 CFR Part 111 in FY2025

702of them (36%) from six citation types, each a specification, a record or an identity test

36–39%share of Part 111 citations under Subpart E (specifications) in every year FY2021–FY2025

8methods in this library, each with equations, a worked example and its limits

U.S. FDA, inspection observations FY2021–FY2025 (Excel files, Foods tab), computed by subpart against eCFR 21 CFR Part 111; details in the first section.

Why it matters

FDA's own inspection data point at the controls this library covers

Six citation types make up 36% of all Part 111 citations, and every one is a specification, a record or an identity testGovernment data

Form 483 observations generated by FDA's electronic inspection tool for inspections ending October 1, 2024 to September 30, 2025; manually prepared 483s are not included. A citation is an inspector's observation, not a finding of product harm.U.S. FDA. Inspection observations, fiscal year 2025 (Excel file, Foods tab; page updated December 16, 2025). Retrieved October 22, 2026.

Text version of the citation counts

365 of 702

of the six most frequent citations concern specifications: 111.70(e) for finished batches and 111.70(b) for components.

Government dataU.S. FDA, inspection observations, FY2025

271

concern master manufacturing records or batch production records (111.205(a), 111.210, 111.260).

Government dataU.S. FDA, inspection observations, FY2025

concern the identity test required for each lot of a dietary ingredient (111.75(a)(1)(i)).

Government dataU.S. FDA, inspection observations, FY2025

36%

of all 1,938 Part 111 citations are those six; the other 1,236 are spread over 183 other citation types.

Government dataU.S. FDA, inspection observations, FY2025 (Foods tab)

FDA's spreadsheet covers Form 483 observations generated by its electronic inspection tool for inspections ending October 1, 2024 to September 30, 2025; manually prepared 483s are not included. A citation is an inspector's observation, not a finding of product harm.

Five fiscal years: specifications drew 36–39% of Part 111 citations every yearGovernment data

U.S. FDA, Inspection observations FY2021–FY2025 Excel files (Foods sheet), downloaded Oct 24, 2026; grouped by 21 CFR 111 subpart (eCFR, Oct 22, 2026) with src/modules/qual_six/compute/p111_years.py. Section-level view: Quality system.

Data table: Part 111 citations by fiscal year

Fiscal year	All Part 111	Subpart E	Subparts H and I	Other
FY2021	704	260 (36.9%)	103 (14.6%)	341
FY2022	1,218	459 (37.7%)	201 (16.5%)	558
FY2023	1,173	423 (36.1%)	189 (16.1%)	561
FY2024	1,583	614 (38.8%)	309 (19.5%)	660
FY2025	1,938	759 (39.2%)	351 (18.1%)	828

Explanation, details and sources

In fiscal 2025, FDA's electronic inspection tool recorded 1,938 citations of 21 CFR Part 111, the dietary-supplement CGMP rule. The six most frequent each concern a written specification, a manufacturing record or an identity test.

Business relevance

Each method answers one operating question and moves a measurable business outcome

Which business measure each of the eight methods moves, and what kind of data its worked example usesEstablished methods

Summary of the method table below; method sources are cited in each tab of the library.

Show the table (8 rows)

Method	Question it answers	Business measure it moves	What it does not establish	Data in the example
Value stream map	Where does a lot wait, and for how long?	Lead time from receipt to release, inventory held, throughput	Why defects occur or whether a step is under control	Illustrative times and yields
DMAIC	How is a recurring quality problem fixed and kept fixed?	Deviations, rework, cost of poor quality	An improvement unless the measurement system and data are valid	Illustrative example; sigma table from normal theory
Process capability	How much of the specification does the process use?	Expected out-of-specification rate, yield, release risk	Future performance if the process is unstable, non-normal or sampled too little; shelf life	Simulated (seeded) assay results
Control charts (SPC)	Has the process changed?	Earlier detection of drift, fewer out-of-specification investigations; trending of release and stability results	That product meets its specification	Simulated (seeded) lot results
Pareto and root cause	Which problems are most frequent, and what could cause one of them?	Where corrective and preventive action (CAPA) effort goes first	Severity; a cause until it is tested	FDA counts (sourced); fishbone causes illustrative
FMEA	Where can the process fail, and which controls come first?	Order of controls, testing and supplier qualification	Product safety; scores are ordinal judgments	Illustrative scores
Gap analysis	What is missing for an inspection, and who closes it by when?	Inspection readiness; supplier qualification decisions	Compliance or an inspection outcome	Illustrative scores for a hypothetical site
OEE	How much planned equipment time is fully productive?	Capacity, throughput and cost per kilogram	Product quality beyond the reject count	Illustrative inputs

Observed

From an EUTECNIQ process: none exist.

Sourced Government data

Published data with its source, such as FDA counts.

Simulated Simulated

Fixed random draws (seeds 2026 and 1022) and slider settings.

Illustrative Illustrative

Invented values; no real product or site.

Explanation, details and sources

The table links every method in the library to the measure it improves and to what it cannot establish. The last column says what kind of data the worked example uses.

Interactive library

Eight methods, each with its equations, a worked example and its limits

A value stream map shows that most of a lot's lead time is waiting

Explanation, details and sources

A value stream map shows that a botanical lot spends most of its lead time waiting, not being processed. One lot of a standardized extract from receipt to shipment, in Rother and Shook's icons: process and data boxes, inventory triangles, push arrows, information flows and a timeline ladder separating waiting from processing.

Lead time43.6 days

Process-cycle efficiency3.2%

Rolled throughput yield74.9%

Current-state value stream map: standardized botanical extract, one lotIllustrative times

Show data: step times, yields and waits for both states

Illustrative values for one lot. C/T = cycle time; C/O = changeover; FPY = first-pass yield; wait = time held after the step.
Step	Value-adding	C/T current (h)	C/O current (h)	FPY current	Hold after step, current	Wait current (d)	C/T future (h)	C/O future (h)	FPY future	Hold after step, future	Wait future (d)
Receipt & sampling	No	2	0	99%	Quarantine	6	2	0	99%	Quarantine	2
Identity testing	No	8	0	96%	Released herb stock	12	8	0	99%	Supermarket (pull)	4
Extraction	Yes	10	4	98%	Extract hold	1	10	3	98%	FIFO lane	0.5
Concentration	Yes	8	2	99%	Concentrate hold	2	8	1	99%	FIFO lane	0.5
Spray drying	Yes	12	6	97%	Powder hold	3	12	3	98%	FIFO lane	1
Blending	Yes	3	2	99%	Sample queue	2	3	1	99%	Sample queue	0.5
QC testing	No	16	0	94%	Record-review queue	5	16	0	97%	Review by exception	1
QA release	No	4	0	90%	Finished goods	10	2	0	98%	Finished goods	4

Totals computed from the table
State	Processing (h)	Value-adding (h)	Waiting (d)	Lead time (d)	Process-cycle efficiency	Rolled throughput yield
Current	63	33	41	43.6	3.2%	74.9%
Future	61	33	13.5	16.0	8.6%	87.7%

Dataset: illustrative. All times and yields are invented for illustration, not measurements. Value-added time counts only steps that transform the material; testing and record review are required by CGMP but are not value-adding in lean terms.Icons and method: Rother M, Shook J. Learning to See: Value-Stream Mapping to Create Value and Eliminate Muda. Lean Enterprise Institute; 1999.

Current versus future state on three measuresIllustrative

What changes in the future state

Identity testing at receipt, a kanban supermarket of released herb, FIFO holds of a day or less and batch records reviewed by exception.

Why it mattersLead time is working capital: every day a lot waits is inventory carried and a later release to the customer.

Details of the four changes

Qualified suppliers and identity testing at receipt shorten quarantine. Part 111 still requires at least one identity test on each dietary-ingredient lot (111.75(a)(1)(i)) unless FDA has granted an exemption (111.75(a)(1)(ii)); a supplier's certificate can be relied on only for other components, after the supplier is qualified (111.75(a)(2)(ii)).
A supermarket of released herb pulled by kanban replaces a large push inventory.
FIFO holds of one day or less between extraction, concentration and drying limit degradation of labile constituents.
Electronic batch records reviewed by exception target the record failures that dominate FDA citations.

Method, assumptions and limits

What it calculates: Lead time, process-cycle efficiency and rolled throughput yield for one lot.
Equations: Lead time (d) = Σ waits (d) + Σ cycle times (h) / 24. Process-cycle efficiency = value-adding time / lead time. Rolled throughput yield = Π FPY_i.
Assumptions: One lot moves through every step; waits are averages; changeover time is shown but not added to the lot's lead time.
Does not establish: Capacity, cost or the cause of yield loss. A map describes one value stream at one time.
Verification: Totals reproduced by tests/numerics/ref/insights-operational-excellence.py.

DMAIC improves a quality problem in five gated phases

Explanation, details and sources

DMAIC turns a quality problem into a measured, controlled improvement in five gated phases. Each phase ends at a tollgate review. The example is an extract whose marker assay varies too much from lot to lot.

Define, Measure, Analyze, Improve, Control: questions, tools and a worked exampleEstablished method

Phase	Question	Main tools	Illustrative example: extract assay
Define	What is the problem, for which customer, and what does success look like?	Project charter SIPOC map Voice of the customer and critical-to-quality tree	Marker assay of a standardized extract ranges 92–108% of label claim; the target is 95–105%.
Measure	How big is the problem, and can we measure it reliably?	Process map and data plan Measurement system analysis (gage R&R) Baseline capability: Cp, Cpk	Gage R&R on the HPLC assay, then baseline data from 30 lots.
Analyze	Which inputs drive the variation?	Pareto chart and fishbone Hypothesis tests and ANOVA Regression on process data	Test whether raw-material marker content and extraction temperature explain most of the variance.
Improve	Which changes reduce variation, shown with data?	Designed experiments (DoE) Mistake-proofing (poka-yoke) Pilot runs and confirmation	DoE on solvent ratio and time; standardize lots to a target marker content.
Control	How is the gain kept?	Control plan and SOPs Statistical process control Updated specifications and records	I-MR chart on the assay; a supplier specification for marker content.

Dataset: illustrative example.Montgomery DC. Introduction to Statistical Quality Control. 8th ed. Wiley; 2019 (DMAIC and the supporting tools).

Defects per million opportunities at each sigma level, with and without the 1.5σ shiftNormal theory

Computed from the standard normal distribution. Shifted: one-sided tail beyond (k − 1.5)σ. Centered: two-sided tail beyond ±kσ. Harry MJ, Lawson JR. Six Sigma Producibility Analysis and Process Characterization. Addison-Wesley; 1992. Coskun A, Ialongo C. Biochem Med. 2020;30(1):010901.

The 3.4 per million figure assumes a drift

"Six sigma" is scored as 4.5σ to the nearest limit, assuming a 1.5σ long-term drift of the mean (figure below).

Show the table (6 rows)

Sigma level	DPMO with shift	Yield with shift	DPMO centered
1σ	691,462	30.9%	317,311
2σ	308,538	69.1%	45,500
3σ	66,807	93.3%	2,700
4σ	6,210	99.38%	63.3
5σ	233	99.98%	0.573
6σ	3.4	99.99966%	0.00197

A 1.5σ drift moves the nearest limit from 6σ to 4.5σ away, which is where 3.4 defects per million comes fromNormal theory

Standard normal distribution. Harry MJ, Lawson JR. Six Sigma Producibility Analysis and Process Characterization. Addison-Wesley; 1992. Coskun A, Ialongo C. Biochem Med. 2020;30(1):010901.

Text version: the 1.5σ convention

Six Sigma practice, following Motorola-era work published by Harry and Lawson, assumes the process mean drifts by up to 1.5σ over the long term, so "six sigma" is scored as 4.5σ to the nearest limit: 3.4 defects per million. A centered process with limits at ±6σ produces about 0.002 per million.

The 1.5σ figure is a convention, not a law: Coskun and Ialongo argue that evidence is needed before applying it to the extra-analytical phases of laboratory testing. EUTECNIQ would report capability from measured within- and between-lot variation instead of assuming the shift.

Defects per million opportunitiesDPMO = 10⁶ d / (u · o)

Defects found12 Units inspected2,000 Defect opportunities per unit5

DPMO1,200

Yield per opportunity99.880%

Long-term Z (no shift)3.04

Sigma level, 1.5σ convention4.54

Where the observed defect rate falls on the sigma scaleCalculator, illustrative counts

DPMO and the sigma-level convention: Harry and Lawson 1992; Montgomery 2019. Normal quantiles computed exactly.

Method, assumptions and limits

What it calculates: Defects per million opportunities, the yield per opportunity, the long-term normal deviate Z with Φ(Z) = 1 − DPMO/10⁶, and the sigma level under the 1.5σ convention (Z + 1.5).
Assumptions: Opportunities are independent and equally likely to fail; the count of opportunities per unit is a definition the team must fix before counting, and changing it changes DPMO.
Valid range: At least one defect: with zero defects the rate is not estimable from the sample (the warning says so). Inputs are clamped to the slider ranges.
Does not establish: Capability of a measured variable (use C_pk with data) or the quality of any real process.
Verification: Independent reference: tests/numerics/ref/insights-operational-excellence.py (scipy.stats.norm).

Capability indices mean something only for a stable, normal, well-sampled process

Move the sliders: C_pk and P_pk diverge when lots drift.

Explanation, details and sources

Capability indices compare the spread of a process with its specification, but only a stable, normal, adequately sampled process earns a meaningful C_pk. A marker assay, % of label claim, specification 95.0–105.0%: 100 simulated results in 25 subgroups of four. C_pk and P_pk diverge when lots drift between subgroups, and the stability check can override the capability reading.

Capability indices (two-sided)C_p = USL − LSL6σ_within
C_pk = min(USL − x̄, x̄ − LSL)3σ_within
σ_within = R̄ / d₂, d₂ = 2.059 (n = 4)
P_p, P_pk: same, with overall s

Process mean (% of label claim)100.6% LC Within-subgroup standard deviation (% LC)0.90 Lot-to-lot drift, between subgroups (% LC)0.60

C_p1.76

C_pk1.53

C_pk 95% interval1.31–1.75

P_p1.60

P_pk1.39

Expected out of spec (normal model, overall s)15 ppm

Observed out of spec0 of 100

Stability check: subgroup means outside X̄-chart limits0 of 25

ReadingStable, and capable at the conventional 1.33 threshold

Assumptions

Results are normal and the process is in statistical control; otherwise no capability index predicts future lots
Expected out-of-specification rate uses the overall standard deviation and the normal distribution
Assay measurement error is part of the observed spread; a gage R&R study separates it

Distribution of 100 simulated assay results against the specificationSimulated data

Show data: the 25 simulated subgroups at the current settings

Subgroup	Result 1	Result 2	Result 3	Result 4	Mean	Range
1	101.77	99.48	101.34	99.02	100.40	2.75
2	100.90	99.69	101.96	101.49	101.01	2.27
3	100.49	98.67	100.08	99.00	99.56	1.82
4	101.03	101.89	102.23	101.62	101.69	1.20
5	100.42	100.24	102.28	101.52	101.11	2.04
6	101.07	102.39	99.69	102.73	101.47	3.04
7	103.25	101.87	101.16	99.82	101.52	3.43
8	101.68	100.38	100.76	100.52	100.84	1.29
9	100.69	101.60	99.94	99.55	100.45	2.05
10	99.56	99.59	99.18	99.64	99.49	0.47
11	102.20	100.50	101.26	99.94	100.97	2.26
12	101.06	99.34	100.28	102.94	100.91	3.60
13	100.04	101.57	100.97	100.42	100.75	1.53
14	99.96	101.71	100.39	101.92	101.00	1.96
15	100.05	100.75	101.12	101.03	100.74	1.07
16	99.12	99.63	100.06	99.46	99.57	0.93
17	100.97	100.89	98.60	98.73	99.80	2.37
18	99.78	101.03	102.76	100.58	101.04	2.98
19	100.75	99.56	99.58	98.76	99.66	1.99
20	99.75	100.63	99.45	101.03	100.22	1.58
21	100.49	100.28	100.78	100.75	100.58	0.50
22	102.03	100.27	98.93	101.10	100.58	3.09
23	99.91	100.44	101.96	99.47	100.44	2.49
24	101.20	101.36	100.93	101.52	101.25	0.59
25	100.57	100.37	101.92	101.63	101.12	1.55

Dataset: simulated. Result = mean + drift × z_b + within-SD × z_w, using fixed standard-normal draws (described in the methods panel). Not results from any product.NIST/SEMATECH e-Handbook of Statistical Methods, §6.1.6; Bissell AF. J R Stat Soc C. 1990;39(3):331–340; Montgomery 2019.

Method, assumptions and limits

What it calculates: Short-term capability (C_p, C_pk) from the within-subgroup spread, overall performance (P_p, P_pk) from the sample standard deviation, an approximate 95% interval for C_pk, the expected out-of-specification rate under a normal model, and an X̄-chart stability check.
Equations: σ_within = R̄/d₂ with d₂ = 2.059 for subgroups of four. C_pk 95% interval = Ĉ_pk ± 1.96 √(1/(9N) + Ĉ_pk²/(2(N − 1))), N = 100. Expected rate = Φ(−(USL − x̄)/s) + Φ(−(x̄ − LSL)/s). X̄-chart limits = x̄ ± A₂R̄ with A₂ = 0.729.
Variables and units: Assay in % of label claim (% LC); LSL 95.0, USL 105.0; sliders in % LC.
Valid range: Mean 96–104% LC, within SD 0.30–2.20, drift 0–1.60; inputs outside are clamped. The interval approximation holds for N ≥ 25 and 0.75 ≤ C_pk ≤ 4 (NIST); outside that range the page reports it as not valid.
Limits of the indices: Normality. The indices and the ppm estimate assume normal data; skewed assays (for example, results bounded by a detection limit) need a transformation or percentile-based indices. Stability. A process out of statistical control has no single distribution, so its index does not predict the next lot; the stability check overrides the reading here. Sampling. NIST advises about 50 or more independent values and generally 100 for a capability study; with fewer, the interval is wide. Within-subgroup σ depends on rational subgrouping. Measurement. Assay error inflates the spread; it is estimated separately by gage R&R.
Does not establish: That every lot will meet specification, product stability or shelf life, or the capability of any real process.
Seeded draws: numpy.random.default_rng(2026).standard_normal: 25 between-subgroup draws, then 25 × 4 within-subgroup draws, rounded to four decimals. The full list is in src/modules/operations/compute/noise.json.
References: NIST/SEMATECH e-Handbook §6.1.6 (indices, interval, sample size, non-normal data); Bissell 1990; Montgomery 2019 (d₂, A₂).
Verification: Independent numpy/scipy reference: tests/numerics/ref/insights-operational-excellence.py.

An individuals and moving-range chart flags a change within a few lots

Add a shift, trend or extra variation after lot 25.

Explanation, details and sources

An individuals and moving-range chart flags a process change within a few lots, before results reach the specification limit. Limits come from the first 20 lots (Phase I) and apply to new lots. A shift, trend or extra variation can be introduced after lot 25; with no change, rules can still fire by chance, and each added rule raises the false-alarm rate.

I-MR limits from Phase I (lots 1–20)UCL, LCL = x̄ ± 3 MR̄1.128
UCL_MR = 3.267 MR̄

Mean shift after lot 25 (σ)1.00σ Trend per lot after lot 25 (σ)0.00σ Variation after lot 25 (× σ)× 1.0

Phase I mean (% LC)99.83

σ̂ = MR̄/1.1280.98

Upper control limit102.78

First signal after Phase Ilot 32 (WE4)

Run rules applied

WE1: one point beyond 3σ
WE2: 2 of 3 consecutive points beyond 2σ on one side
WE3: 4 of 5 consecutive points beyond 1σ on one side
WE4: 8 consecutive points on one side of the center line
Nelson 3: 6 consecutive points steadily rising or falling

Individuals (top) and moving-range (bottom) charts, marker assay of 40 lotsSimulated data

Show data: 40 simulated lots at the current settings

Lot	Result (% LC)	Moving range	Signals
1	100.00		none
2	99.08	0.92	none
3	101.41	2.33	none
4	99.84	1.57	none
5	99.63	0.21	none
6	100.72	1.09	none
7	99.32	1.40	none
8	100.11	0.80	none
9	99.86	0.25	none
10	99.98	0.12	none
11	99.84	0.14	none
12	98.58	1.27	none
13	100.98	2.41	none
14	99.60	1.38	none
15	100.37	0.77	none
16	98.44	1.94	none
17	100.03	1.59	none
18	99.80	0.23	none
19	98.73	1.06	none
20	100.28	1.55	none
21	100.03	0.25	none
22	100.49	0.47	none
23	99.33	1.16	none
24	99.46	0.13	none
25	100.01	0.54	none
26	100.59	0.58	none
27	101.42	0.83	none
28	99.95	1.47	none
29	101.97	2.02	none
30	101.50	0.47	none
31	100.66	0.84	none
32	101.96	1.30	WE4
33	101.31	0.65	WE3, WE4
34	101.31	0.00	WE3, WE4
35	100.72	0.59	WE4
36	100.99	0.28	WE3, WE4
37	103.31	2.31	WE1, WE3, WE4
38	100.72	2.59	WE4
39	101.42	0.70	WE4
40	100.14	1.27	WE4

Dataset: simulated. Lot result = 100 + z for lots 1–25; afterwards 100 + shift + trend × (lot − 25) + variation × z, with fixed standard-normal draws z (seed 1022). Units of σ are % of label claim.Western Electric Co. Statistical Quality Control Handbook. Indianapolis; 1956. Nelson LS. J Qual Technol. 1984;16(4):237–239. Constants d₂ = 1.128 and D₄ = 3.267 for moving ranges of two (Montgomery 2019).

Method, assumptions and limits

What it calculates: Center line and 3σ limits for individual results and their moving ranges from Phase I, then the Western Electric and Nelson run rules for all 40 lots.
Equations: σ̂ = MR̄/d₂, d₂ = 1.128; limits x̄ ± 3σ̂; UCL_MR = D₄MR̄ with D₄ = 3.267; zone lines at ±1σ̂ and ±2σ̂.
Valid range: Shift 0–2.5σ, trend 0–0.2σ per lot, variation ×1.0–2.5; inputs outside are clamped.
Assumptions: Independent, approximately normal results; Phase I is in control. If Phase I itself signals, the page warns that limits should be re-established after investigation.
False alarms: For a normal in-control process, rule WE1 alone fires on about 0.27% of points (about one in 370). Adding rules shortens the time to detect real shifts and also raises the false-alarm rate.
Does not establish: That results meet specification. A process can be in control and still incapable, or capable and out of control.
Business use: The same chart logic trends release assays, impurity results and stability-study time points, so a drift is investigated before it becomes an out-of-specification result.
Verification: Independent numpy reference: tests/numerics/ref/insights-operational-excellence.py.

Two specification citations lead; a fishbone organizes the causes

Explanation, details and sources

Two specification citations account for half of the most frequent supplement CGMP citations; a fishbone then organizes the possible causes of one failure. FDA's fiscal 2025 counts for the six most frequent Part 111 citations; percentages are shares of those 702, not of all 1,938. The fishbone shows how an out-of-specification potency result would be investigated.

Pareto chart: the six most frequent Part 111 citations, FDA fiscal 2025Government data

Show data

Section	Requirement (FDA short description)	Citations	Cumulative share
111.70(e)	Specifications, finished dietary supplement	188	26.8%
111.70(b)	Specifications, components	177	52.0%
111.260	Batch record, required information	103	66.7%
111.210	Master manufacturing record, record requirements	93	79.9%
111.205(a)	Master manufacturing record, prepared	75	90.6%
111.75(a)(1)(i)	Component, verify identity of a dietary ingredient	66	100.0%

Dataset: sourced. FDA counts each citation type (Cite ID) separately: 111.205(a) also has 4 citations for a master record not followed, and written quality-control procedures (111.103) total 67 across two citation types. The data cover Form 483 observations generated by FDA's electronic inspection tool, not every observation issued.U.S. FDA. Inspection observations, fiscal year 2025 (Excel file, Foods tab; page updated December 16, 2025). Retrieved October 22, 2026.

Reading it

The two specification citations make up 52% of the 702; adding the three record citations brings the total to 91%. A new plant should write specifications and master records before the first batch, and design batch records so an incomplete one cannot be released.

Ishikawa placed the Pareto diagram and the cause-and-effect diagram among the basic quality tools (Guide to Quality Control, 1982). A count ranks frequency, not severity.

Cause-and-effect (Ishikawa) diagram for a potency result below specificationIllustrative example

Show the causes and how each would be tested

Category	Possible cause	How it would be tested
Materials	Chemotype differs from reference	Compare marker ratios and the HPTLC fingerprint with reference lots
Materials	Harvest timing or plant part	Review supplier harvest records; botanical microscopy
Materials	Adulteration or substitution	Orthogonal identity methods, such as HPTLC plus LC-MS
Materials	Moisture above limit	Loss on drying of the retained raw-material sample
Methods	Extraction time or temperature drift	Batch record and temperature data-logger review
Methods	Solvent ratio charged incorrectly	Weigh tickets and second-person verification
Methods	Sample preparation incomplete	Re-extraction recovery study on the retained sample
Methods	Assay not stability-indicating	Forced-degradation specificity check
Machines	Evaporator overheating	Temperature trend; look for degradant peaks
Machines	Dryer outlet temperature low	Dryer log and residual moisture
Machines	HPLC column degradation	System-suitability history
Machines	Balance out of calibration	Calibration and check-weight records
Measurement	Reference standard assignment	Reference-standard certificate; retest against a second lot
Measurement	Integration or calculation error	Audit trail and independent recalculation
Measurement	Sample not representative	Sampling plan; test retained composite samples
Measurement	Analyst-to-analyst bias	Gage R&R results for the method
People	Training on a revised procedure	Training records against the procedure revision date
People	Transcription error	Compare raw data with the batch record
People	Shift handover gap	Batch-record time stamps and signatures
Environment	Light or oxygen exposure	Packaging, headspace and hold-time records
Environment	Humidity during blending	Room monitoring data for the blending date
Environment	Storage temperature excursion	Warehouse temperature log

Dataset: illustrative. An out-of-specification investigation first checks the laboratory result (measurement) before attributing the failure to the material or process. Causes are examples, not findings.Ishikawa K. Guide to Quality Control. 2nd rev. ed. Tokyo: Asian Productivity Organization; 1982.

An FMEA ranks where an extract process can fail

Change a score to re-rank the failure modes.

Explanation, details and sources

A failure mode and effects analysis ranks where an extract process can fail, so controls go first where severity and likelihood are high and detection is weak. Changing a severity, occurrence or detection score updates the risk priority number (RPN = S × O × D), the ranking and the risk map. Scores are illustrative judgments for a generic extract process, not a real plant.

An FMEA is not proof of safety. A low RPN, a team's ordinal judgment, shows neither that a product is safe nor that a control works. Safety rests on hazard analysis, specifications, validated methods, test results and, for a new dietary ingredient, a documented safety evaluation.

Highest RPN: 144 (failure modes 1 and 2)

Show the table (12 rows)

		Failure mode	Effect	Current control		Recommended action
1	Raw-material receipt	Wrong species or adulterated material	Identity failure; possible safety hazard	Identity test (HPTLC) on each lot	144	Add an orthogonal identity method; qualify suppliers by audit
2	Raw-material receipt	Heavy metals or pesticides above limit	Consumer exposure; batch rejection	Supplier certificate review	144	Risk-based testing of each lot from high-risk regions
10	QC testing	Integration or calculation error	Wrong release decision	Second-person review	120	Validated calculation templates; audit-trail review
12	Storage	Oxidation of actives	Potency loss before expiry	Sealed liners	108	Nitrogen flush; stability-indicating method in the stability program
9	Blending	Non-uniform blend	Potency varies between containers	Blend uniformity at validation only	105	Routine stratified sampling for new lots
11	Batch record	Incomplete or missing entries	Release delay; FDA observation	QA review before release	90	Electronic batch record with required fields
6	Concentration	Evaporator overheating or fouling	Thermal degradation; off-color	Temperature monitoring	84	Cleaning frequency based on fouling trend
3	Milling	Particle size outside range	Variable extraction yield; filter blinding	Sieve check per batch	80	Screen inspection schedule; laser-diffraction check
7	Spray drying	Residual moisture high	Caking; microbial growth risk	In-process loss on drying	72	Control outlet temperature within a validated range
4	Extraction	Temperature overshoot	Degradation of marker compounds	Jacket temperature alarm	63	Interlock on heating; trend batch temperature profiles
5	Extraction	Solvent ratio charged incorrectly	Low yield and potency	Second-person verification	54	Mass-flow metering with batch-record capture
8	Spray drying	Residual solvent above limit	Specification failure; safety	Headspace GC at release	32	Keep; trend results for early drift

Risk map: severity against occurrence for each failure modeIllustrative scores

Dataset: illustrative scores.ICH Q9(R1) Quality Risk Management, adopted 18 January 2023 (FMEA among the tools in Annex I); AIAG & VDA FMEA Handbook, 2019 (Action Priority).

Limits of the method

Equal RPNs can hide very different risks (figure below), scores stay subjective, and an extract operation still needs the written hazard analysis of 21 CFR 117.130.

Text version of the three limits

RPN multiplies ordinal scores. Very different risks can share a number: S10·O1·D1 and S1·O10·D1 both give 10. The 2019 AIAG & VDA handbook introduced Action Priority tables to address this.
Subjectivity cannot be removed. ICH Q9(R1) says it can be controlled by addressing bias and assumptions and using data; it also asks that formality grow with uncertainty, importance and complexity.
Extract operations need a hazard analysis. Making a bulk dietary ingredient for further processing is generally subject to 21 CFR Part 117, including a written hazard analysis (117.130). The exemption in 117.5(e) covers only manufacturing of a dietary supplement in compliance with Part 111 and FD&C Act § 761.

Equal risk priority numbers can describe very different risksIllustrative scores

AIAG & VDA FMEA Handbook (2019), Action Priority; ICH Q9(R1), adopted 18 January 2023.

Every RPN on the severity × occurrence grid, at a chosen detection scoreComputed from RPN = S × O × D

Detection score D

Each cell is S × O × D for the chosen D. Many different cells share the same number, which is why the 2019 AIAG & VDA handbook introduced Action Priority tables to replace the RPN.AIAG and VDA, FMEA Handbook (2019); AIAG announcement, June 3, 2019.

Data table: the grid at the chosen detection score

S \ O	1	2	3	4	5	6	7	8	9	10
Each cell equals S × O × D; the table fills in when the page's scripts run.

Method, assumptions and limits

What it calculates: Risk priority number for each failure mode and the ranking by RPN.
Equation: RPN = S × O × D, each scored 1–10 (range 1–1,000). Severity: effect on the customer or product; occurrence: likelihood of the cause; detection: chance that current controls miss it (10 = almost certainly missed).
Assumptions: Scores come from a cross-functional team using written scales; the values here are invented.
Does not establish: Product safety, regulatory compliance or the effectiveness of a control. Equal RPNs do not mean equal risk.
References: ICH Q9(R1) (2023); AIAG & VDA FMEA Handbook (2019).
Verification: RPN values checked by tests/numerics/ref/insights-operational-excellence.py.

A gap analysis turns each gap into evidence, an owner and a date

Explanation, details and sources

A gap analysis scores current capability against what an inspection requires, then turns each gap into evidence, an owner and a date. A hypothetical contract manufacturer scored on the subparts of 21 CFR Part 111, using a 0–5 maturity scale (0 absent, 1 ad hoc, 2 documented, 3 implemented, 4 with records and trend review, 5 sustained and verified by internal audit).

Readiness for a Part 111 inspection: required versus current maturityIllustrative scores

Show data

Part 111 area (subpart)	Required	Current	Gap
Personnel (B)	4	3	1
Plant and grounds (C)	4	3	1
Equipment (D)	4	2	2
Specifications (E)	5	1	4
Testing (E)	5	2	3
Quality control (F)	5	2	3
Components (G)	4	2	2
Master record (H)	5	2	3
Batch record (I)	5	2	3
Laboratory (J)	4	2	2
Complaints (N, O)	4	3	1
Records (P)	4	2	2

Dataset: illustrative scores for a hypothetical site.Axes: 21 CFR Part 111 subparts B–P (eCFR, current). The maturity scale is an EUTECNIQ proposal.

How the gaps are prioritized

Largest gaps first, weighted by FDA citation frequency, so specifications, identity testing and the two manufacturing records lead. A gap closes only with evidence an inspector could review, not with a procedure never used.

Records must be kept for 1 year past the shelf-life date, or 2 years beyond distribution of the last batch if shelf-life dating is not used (111.605).

Show the table (8 rows): Gap register for the hypothetical site (illustrative)

Gap register for the hypothetical site (illustrative)
Gap	Part 111 reference	Evidence that closes it	Action	Owner	Weeks
Component and finished-batch specifications missing for identity, purity, strength and composition	111.70(b), 111.70(e)	Approved written specifications with limits for each attribute	Draft from supplier data, literature and method capability; QA approval	QA	6
Identity testing of each dietary-ingredient lot not performed	111.75(a)(1)(i)	Scientifically valid identity method run on every lot, with reference material	Qualify an HPTLC or microscopy method; obtain reference materials	QC laboratory	8
Supplier certificates accepted without qualification	111.75(a)(2)(ii)	Records confirming supplier results, with periodic re-confirmation	Confirm supplier results by testing; set a re-confirmation schedule	Supplier management	10
Master manufacturing records incomplete	111.205, 111.210	Approved master record for each formulation and batch size	Write a controlled template; review against the regulation	QA	6
Batch production records missing entries	111.255, 111.260	Complete batch records with each step documented and verified	Redesign the record; train operators; second-person checks	Operations	4
Equipment calibration not documented	Subpart D	Calibration and maintenance records for equipment that affects quality	Calibration program and schedule	Operations	6
No laboratory control or out-of-specification procedure	Subpart J	Written laboratory control processes and investigation records	Procedures, method verification, analyst training	QC laboratory	8
Complaint and returns handling informal	Subparts N, O	Written procedures with quality-control review of each complaint	Procedures and a complaint log reviewed by quality control	QA	4

Gap-closing roadmap, weeks from kickoffIllustrative plan

Show data

Task	Owner	Start week	Duration (weeks)	End week
Specifications	QA	0	6	6
Equipment calibration program	Operations	0	6	6
Identity methods and reference materials	QC laboratory	2	8	10
Supplier qualification	Supplier management	2	10	12
Master manufacturing records	QA	4	6	10
Laboratory controls and OOS procedure	QC laboratory	6	8	14
Batch record redesign and training	Operations	8	4	12
Complaint and returns procedures	QA	10	4	14
Records control and retention	QA	12	4	16

Decision points: internal audit (week 18), mock inspection (week 22), qualification decision (week 26).

Dataset: illustrative plan. Bars are colored by owner. Diamonds are decision points. Durations are examples; a real plan is sized from the assessment. A gap score is not a compliance finding or a prediction of an inspection outcome.

OEE splits lost capacity into availability, performance and quality

Adjust downtime, changeovers, output and rejects.

Explanation, details and sources

Overall equipment effectiveness splits lost capacity into availability, performance and quality losses, so the biggest loss is fixed first. The example is one week of a spray dryer producing extract powder. Adjust downtime, changeovers, output and rejects; the loss chart shows where productive time goes.

Overall equipment effectivenessOEE = A × P × Q
A = run timeplanned time Q = good outputoutput
P = outputideal rate × run time

Unplanned downtime (h)18 h Changeovers and cleaning (h)10 h Total powder output (kg)3,800 kg Rejected or reworked (kg)150 kg

Availability76.7%

Performance82.6%

Quality96.1%

OEE60.8%

Assumptions

Planned production time 120 h (five 24-hour days); ideal rate 50 kg of powder per hour
Changeovers and cleaning count as availability loss, as in Nakajima's loss structure
All inputs are illustrative, not a measured dryer

Where the planned 120 hours goModel output

Time category	Hours	Lost in this step (h)
Planned production time	120.0	none
Run time (after availability loss)	92.0	28.0
Net run time at the ideal rate (after performance loss)	76.0	16.0
Fully productive time (after quality loss)	73.0	3.0

Dataset: illustrative inputs.Nakajima S. Introduction to TPM: Total Productive Maintenance. Cambridge, MA: Productivity Press; 1988.

Method, assumptions and limits

What it calculates: Availability, performance, quality and their product, OEE, for one week of one machine, and the hours lost at each step.
Equations: Run time = planned − downtime − changeovers. A = run/planned. Net run time = min(output/ideal rate, run time). P = net/run. Q = (output − rejects)/output. Fully productive time = net × Q = good output/ideal rate.
Variables and units: Hours (h), kilograms (kg), ideal rate 50 kg/h, planned time 120 h.
Valid range: Downtime 0–40 h, changeovers 0–30 h, output 1,000–6,000 kg, rejects 0–600 kg. Rejects above output are capped at the output. Output above what the ideal rate allows in the run time is physically impossible; performance is then capped at 100% and a warning appears.
Does not establish: Product quality beyond the reject count, the cost of each loss or a benchmark for botanical dryers.
Business use: Converts lost hours into lost kilograms and capacity, so investment goes to the largest loss.
Verification: Independent reference: tests/numerics/ref/insights-operational-excellence.py.

Explanation, details and sources

Choose a method. Calculators show their equations, guard their inputs and have a Reset button and a methods panel. Every chart has a data table or a text description.

Read with care

Six statements these methods are often stretched to support

C_pk ≥ 1.33≠Every future lot will meet specification
A process in statistical control≠A process that meets its specification
A low FMEA risk priority number≠A safe product
An FMEA≠The written hazard analysis 21 CFR 117.130 requires
A gap-analysis score≠A compliance finding or an inspection outcome
Stability testing with trended results≠A validated shelf life

See how these methods fit the proposed quality system

The quality-system page describes the proposed Part 111 and Part 117 controls; process engineering covers scale-up and technology transfer.

Quality system Process engineering