Lume Sensor ATP Validation Study Design
& Statistical Analysis Plan

1. Background & Regulatory Context

1.1 The Alternate Test Procedure Program

The EPA Alternate Test Procedure (ATP) program, codified at 40 CFR 136.4 (nationwide) and 40 CFR 136.5 (limited-use, facility-specific), provides a pathway for developers to submit new analytical methods for regulatory approval as alternatives to methods listed in 40 CFR Part 136. These are the required methods for Clean Water Act compliance reporting. This document describes the study design and statistical analysis plan for a limited-use ATP application under 40 CFR 136.5, submitted by the City of Boulder Utilities to EPA Region 8.

1.2 Colorado Regulation 93 / 303(d)

Boulder Creek is listed as impaired under Colorado Regulation 93 (the state implementation of the federal 303(d) program for impaired waters) for E. coli exceedances. The applicable water quality standard is a geometric mean of 126 CFU/100 mL over a 30-day period, with a single-sample maximum of 235 CFU/100 mL. The City of Boulder Utilities currently monitors compliance at 6 designated monitoring locations on Boulder Creek using IDEXX Colilert as the reference method. This study will determine whether the Virridy Lume tryptophan-like fluorescence (TLF) sensor produces results statistically equivalent to Colilert at these compliance monitoring sites.

1.3 The Case for Continuous Monitoring

Current Reg 93 compliance monitoring on Boulder Creek relies on discrete grab samples processed by a laboratory with an 18–24 hour delay before results are available. This creates a structural monitoring gap: contamination events driven by storm runoff, upstream discharges, or wildlife activity may occur and resolve entirely between sampling events, leaving the public exposed without warning. The Lume sensor provides continuous, in-situ E. coli estimates at 15-minute intervals, enabling real-time detection of exceedance events and substantially increasing the density of compliance-relevant data.

1.4 Relationship to Existing Data

Virridy has previously collected 856 paired Lume–Colilert observations across multiple sites and waterbodies (global dataset, 2021–2025) (R² = 0.67, Cohen’s kappa = 0.82–0.84, MAPE = 7.1% in log-transformed space). This existing dataset is used to support the pre-submission consultation with EPA Region 8 and demonstrates technical feasibility, but does not constitute the formal regulatory validation dataset. The prospective study described in this document will generate the regulatory-grade paired dataset required for the 136.5 application, under a pre-specified, agreed study protocol.

2. Study Objectives & Regulatory Claim

2.1 Primary Objective

Demonstrate that the Virridy Lume tryptophan-like fluorescence sensor, using a multivariate linear regression model with TLF intensity, turbidity, and temperature as inputs, produces E. coli concentration estimates that are statistically equivalent to IDEXX Colilert results at Boulder Creek’s 6 Regulation 93 TMDL monitoring sites across all seasons, flow conditions, and concentration ranges relevant to Reg 93 compliance.

2.2 Secondary Objectives

• Demonstrate correct binary classification of samples above and below the Reg 93 geometric mean threshold (126 CFU/100 mL) with pre-specified sensitivity and specificity targets
• Characterize method precision (sensor-to-sensor reproducibility) across the 6 monitoring sites
• Determine the method detection limit (MDL) in both TLF units and translated CFU/100 mL
• Characterize known interferences (dissolved organic carbon, humic fluorescence, turbidity, temperature) and their effect on the prediction equation
• Generate a complete method document suitable for inclusion in a 40 CFR 136.5 application

2.3 Regulatory Claim Hierarchy

The primary regulatory claim is direct replacement of Colilert grab sampling for Reg 93 compliance reporting. Fallback claims are listed in order of preference if the primary claim cannot be fully supported by the dataset:

3. Method Description

3.1 Principle of the Method

The Lume sensor measures tryptophan-like fluorescence (TLF) in situ at 275 nm excitation and 340 nm emission. Tryptophan is an amino acid present in the cell walls of bacteria, including E. coli, and its fluorescence signal is a proxy for fecal contamination. The sensor simultaneously measures turbidity (NTU) and water temperature (°C) as co-variates. A multivariate linear regression model, fit to paired Lume–Colilert calibration data and validated on a held-out test set, converts the three sensor outputs to an E. coli concentration estimate in CFU/100 mL.

3.2 Prediction Equation

The regulatory prediction equation takes the following form, fit in log₁₀ space to account for the log-normal distribution of environmental E. coli concentrations:

Where TLF is in ppb tryptophan equivalents, turbidity is in NTU, and temperature is in °C. The model coefficients (β₀–β₃) will be estimated by ordinary least squares regression on the training partition of the study dataset (see Section 8). The fitted coefficients will be locked before evaluation on the holdout test set and will not be modified thereafter. The final coefficient values will be reported in the ATP application and become fixed parameters of the approved method.

3.3 Sensor Specifications

Parameter	Specification
TLF excitation wavelength	275 nm
TLF emission wavelength	340 nm
TLF units	ppb tryptophan equivalents
Turbidity measurement	Nephelometric (NTU)
Temperature measurement	°C (thermistor)
Measurement interval	15 minutes (configurable)
Deployment type	Submersible in-situ sensor
Data transmission	Cellular (primary); on-board SD card (backup)
Power	Solar or line power with battery backup
Operating temperature range	0–40 °C

3.4 Method Scope and Applicability

This method is applicable to freshwater surface water for the measurement of E. coli in the range of approximately 1–10,000 CFU/100 mL. The method is designed for continuous in-situ monitoring and is not applicable to discrete grab sample laboratory analysis. The method as validated is specific to Boulder Creek, Colorado under the 40 CFR 136.5 limited-use approval; nationwide applicability under 40 CFR 136.4 will require additional multi-site validation.

4. Study Design

4.1 Study Sites

The study will be conducted at the 6 City of Boulder Utilities TMDL compliance monitoring locations on Boulder Creek. These are the same sites at which Boulder currently collects Colilert grab samples for Regulation 93 reporting to CDPHE. Using existing compliance monitoring sites ensures that the validation dataset directly reflects the conditions and concentration ranges relevant to the regulatory claim.

Site	Description	Expected Range	Regulatory Significance
BC-1	Upstream reference / background	Low (<10 CFU/100 mL)	Establishes baseline TLF and natural organic matter background; critical for MDL characterization
BC-2	Above WWTP discharge	Low–moderate	Mixed urban/natural signal; no direct WWTP influence
BC-3	Below WWTP discharge	Variable — low to high	Critical compliance point; demonstrates performance at and above the 126 CFU/100 mL Reg 93 threshold
BC-4	Mid-creek compliance point	Moderate	Mixed WWTP effluent + stormwater signal; seasonal variability important
BC-5	Downstream compliance point	Moderate–low	Natural attenuation zone; demonstrates performance during dilution and recovery
BC-6	Lower Boulder Creek / watershed outlet	Variable	Integrates full watershed; irrigation diversions may affect flow dynamics seasonally

Note: Site descriptions and names are provisional. Exact site IDs, GPS coordinates, and Reg 93 permit associations to be confirmed with City of Boulder Utilities prior to study start.

4.2 Study Duration

The study will run for 12 months from the formal study start date (defined as the date all 6 sensors are deployed and the reference method sampling program is operating at the required frequency). Twelve months is the minimum duration required to capture all four seasons and, critically, two complete spring runoff events on Boulder Creek. Spring runoff (typically April–June) produces the highest E. coli concentrations of the year and is the period most likely to generate exceedances of the Reg 93 threshold. It must be represented in both the training and holdout partitions of the dataset.

4.3 Sensor Deployment

Task	Timing	Details
Deploy Lume sensors at all 6 sites	Pre-study	Each sensor co-located within 2 meters of the designated Colilert grab sample collection point at the same depth and cross-section position. GPS coordinates, installation photographs, and sensor serial numbers documented at deployment.
Firmware and model version lock	Study Day 0	Firmware version and prediction equation coefficients (once fit per Section 8) are locked for the study duration. Any firmware updates require a protocol amendment and notification to EPA Region 8.
Data pipeline validation	Week 1	Confirm all 6 sensors are transmitting reliably. Verify measurement interval, data completeness, and sensor-to-server latency. Automated alerts configured for signal dropout, out-of-range readings, and battery/power status.

4.4 Operators and Responsibilities

Party	Role
Virridy	Sensor deployment, maintenance, calibration verification, data management, and statistical analysis. Contracted to City of Boulder for the duration of the study.
City of Boulder Utilities	Applicant on the 136.5 ATP application. Oversight of the study. Coordinates Colilert grab sample collection at the 6 monitoring sites per existing Reg 93 SOPs.
Reference Laboratory	State-certified laboratory for E. coli analysis. Processes all Colilert samples for the study. [See Open Question B-1: confirm certified lab and willingness to increase sampling frequency]
CDPHE	State regulatory oversight. Reviews study design and provides letter of support for the 136.5 application.

5. Reference Method Protocol

5.1 Reference Method

The reference method for this study is IDEXX Colilert for E. coli quantification in freshwater, analyzed using the Quantitray/2000 MPN method per the manufacturer’s protocol and consistent with 40 CFR Part 136 Table IA. The specific Colilert formulation (standard 24-hour vs. Colilert-18 18-hour) will be confirmed with City of Boulder Utilities and will match the version currently used for Reg 93 compliance reporting. Both are EPA-approved; consistency with Boulder’s existing compliance program is required so that the ATP validation dataset is directly comparable to the regulatory baseline.

5.2 Sample Collection Protocol

Parameter	Requirement
Collection frequency	Minimum 3 times per week per site during the study period. Event-based enhanced sampling (see 5.3) increases this frequency during high-flow events.
Collection timing	Exact time of grab sample recorded to the minute by the collector. Matched against nearest Lume 15-minute reading (pairing window: ≤15 minutes; see Section 6).
Collection location	Within 2 meters of the Lume sensor, same depth and cross-section position.
Collection containers	Sterile, single-use 100 mL bottles per Part 136 / Colilert protocol.
Preservation and holding time	Per approved Colilert protocol: chilled at 4 °C, analyzed within 6 hours of collection.
Chain of custody	Formal chain of custody documentation for every sample, from collection through laboratory reporting. COC retained for minimum 10 years.
Field conditions log	At each sampling event: weather (clear/rain/snow/runoff), visual water clarity, estimated stream flow condition (baseflow/elevated/flood), any visible upstream disturbance, air temperature, and collector initials.
Laboratory duplicate	Every 10th paired sample: collect a field duplicate grab (two bottles from the same sampling event). Submitted blind to the laboratory. Provides reference method precision estimate (relative percent difference).

5.3 Enhanced Sampling During High-Flow Events

Spring runoff and storm events are the primary drivers of Reg 93 exceedances on Boulder Creek. The Lume’s continuous monitoring will automatically detect elevated TLF signals. When any Lume sensor reads above a pre-defined alert threshold (to be determined during initial deployment, approximately 2× baseline TLF), the Colilert sampling frequency at that site will be increased to daily for the duration of the event. This generates additional paired observations at high concentrations, which are the most important for demonstrating performance at and above the 126 CFU/100 mL threshold.

6. Paired Data Collection Protocol

6.1 Definition of a Valid Paired Observation

A paired observation is defined as a Colilert result and a Lume sensor reading that meet all of the following criteria:

1. The Colilert grab sample was collected within ±15 minutes of a Lume 15-minute measurement interval (i.e., the Lume reading is the single reading closest in time to sample collection, within one measurement interval)
2. The Lume sensor was operating normally at the time of the reading (no QC flags applied; see Section 7)
3. The Colilert sample was analyzed by the certified reference laboratory and has a valid COC
4. The grab sample was collected within 2 meters of the sensor at the same depth and cross-section
5. No documented maintenance activity (sensor cleaning, calibration, removal) was performed within 30 minutes of the reading

6.2 Minimum Sample Size Requirements

Metric	Minimum Target	Rationale
Total valid paired observations	≥600	6 sites × ≥100 per site. Substantially exceeds Colilert’s original multi-lab validation sample count. Provides sufficient power for equivalence testing and holdout evaluation.
Paired observations per site	≥80	Minimum for site-level performance statistics. Lower-bound sites may have fewer observations due to site-specific access constraints.
Observations above 126 CFU/100 mL	≥30 total (≥5 per site)	The Reg 93 threshold must be represented in the dataset. Classification performance at the threshold is meaningless without samples above it. Enhanced event sampling (Section 5.3) is designed to meet this requirement.
Observations above 235 CFU/100 mL	≥10 total	The Reg 93 single-sample maximum. High-end performance must be characterized.
Seasonal coverage	All 4 seasons at all 6 sites	Spring runoff, summer baseflow, fall recession, and winter low-flow represent distinct hydrologic and biological conditions. All must be represented.
Continuous data completeness	≥90% uptime per sensor per month	Demonstrates operational reliability of the method as deployed.

6.3 Data Recording and Management

All paired observations will be stored in a centralized, version-controlled database. Each record will include: site ID, sample date and time (UTC), Lume TLF reading (ppb), turbidity (NTU), temperature (°C), raw Lume E. coli estimate (CFU/100 mL), Colilert MPN result (CFU/100 mL), data quality flags, field conditions log entries, laboratory COC reference, and collector/operator ID. Raw data will be archived in original format with all transformations and QA flag applications fully documented and reproducible. Minimum retention: 10 years.

7. QA/QC Requirements

7.1 Sensor Calibration and Verification

QC Check	Frequency	Details
Field blank	Monthly (per site)	Submerge sensor in laboratory-grade DI water before and after each maintenance visit. TLF reading must be ≤[MDL to be determined]. Any reading above MDL triggers investigation and possible recalibration before data collection resumes.
Tryptophan standard verification	Monthly (per site)	Verify sensor response against a prepared tryptophan standard at a known concentration (e.g., 100 ppb). Acceptable response: within ±15% of the expected value. Out-of-range triggers recalibration.
Optical window cleaning	Bi-weekly	Clean optical window with approved cleaning solution. Verify post-cleaning blank reading before returning sensor to service.

7.2 Data Quality Flags

Flag	Trigger	Action
Fouling suspected	TLF drift >20% over 24 hours without corresponding Colilert change; post-cleaning blank >MDL	Flag all readings during fouling period. Exclude from paired dataset. Document in maintenance log.
Out-of-range turbidity	Turbidity >[upper operating limit NTU — to be determined from initial deployment]	Flag readings. High turbidity may saturate optical signal. Include flagged readings in sensitivity analysis but exclude from primary dataset.
Maintenance window	Any period during which sensor was removed, cleaned, or calibrated	Exclude all readings within ±30 minutes of maintenance activity from paired dataset.
Power interruption	Gap in data >30 minutes	Flag readings immediately before and after gap. Investigate cause. Exclude interpolated values from paired dataset.
Temperature out of range	Water temperature <0 °C or >35 °C	Flag readings. Outside characterized operating range. Include in sensitivity analysis; exclude from primary dataset.

7.3 Reference Method QC

QC Check	Frequency	Details
Laboratory field duplicate	Every 10th sample	Blind field duplicate submitted to certified laboratory. Acceptable lab RPD: ≤[to be confirmed with reference laboratory; typically ≤30% for environmental E. coli samples].
Laboratory positive control	Per laboratory QA plan	Certified laboratory runs positive and negative controls per their accredited QA plan. Control results reported with each analytical batch.
Colilert holding time compliance	Every sample	Sample analyzed within 6 hours of collection. Samples exceeding holding time excluded from paired dataset.

7.4 Data Completeness Requirements

For a calendar month’s data to count toward the study dataset, the following minimum completeness thresholds must be met: (a) ≥90% of expected Lume 15-minute readings must be present and unflagged for each site; (b) ≥80% of planned Colilert grab samples must have valid paired Lume readings within the ±15 minute pairing window. Months not meeting these thresholds will be documented and excluded from the primary dataset; if exclusion brings total paired observations below the Section 6.2 minimums, the study will be extended accordingly.

8. Model Development & Validation

8.1 Model Form

The prediction equation (Section 3.2) will be fit by ordinary least squares (OLS) regression to log₁₀-transformed E. coli concentrations. Below-detection-limit Colilert results will be treated as one-half the detection limit for regression purposes (standard left-censored data convention). The model form is additive with three predictors — TLF, turbidity, and temperature — and no site-specific or temporal covariates. No interaction terms will be included unless pre-specified below and justified by exploratory analysis on the training partition only.

8.2 Training / Holdout Split — Pre-Specified Protocol

Partition	Data Period	Purpose
Training set	Months 1–9 of the study (all 6 sites)	Fit the OLS regression coefficients (β0–β3). Model is built entirely on this partition. The coefficients are fixed after training and not updated.
Holdout test set	Months 10–12 of the study (all 6 sites)	Independent evaluation of the fixed model. All primary performance statistics (see Section 9) are calculated on this partition only. The holdout period is chosen to include spring runoff (if the study starts in January, this is the second spring; otherwise adjusted accordingly).

A temporal split is used rather than a random split because it mimics real-world deployment: the model is trained on historical data and applied forward in time. This is more conservative than a random split and eliminates any possibility of the training data "seeing" future conditions.

8.3 Model Acceptance Criteria (Training Phase)

Before the model is locked for holdout evaluation, it must meet the following minimum acceptance criteria on the training set:

• R² ≥ 0.50 (log₁₀-transformed concentrations)
• All three predictors (TLF, turbidity, temperature) must have non-zero coefficients with the expected sign (TLF positive; turbidity positive or negative depending on correlation with organic matter; temperature positive)
• No predictor with a statistically significant coefficient of opposite sign to that expected (would indicate collinearity or overfitting)
• Residuals approximately normally distributed (Shapiro-Wilk test, p > 0.05) and homoscedastic across the fitted range

If these criteria are not met on the training set, the model form will be reviewed (e.g., log₁₀ transformation of predictors, addition of one pre-specified interaction term: TLF × turbidity) with notification to EPA Region 8. The holdout set will not be examined until the model form is finalized.

8.4 Model Coefficient Reporting

The final fitted coefficients (β₀, β₁, β₂, β₃), standard errors, 95% confidence intervals, and p-values will be reported in the ATP application. The complete prediction equation with numerical coefficients will be included in the method documentation (Section 3.2) such that any laboratory can implement the method using a standard spreadsheet without proprietary software.

9. Statistical Analysis Plan

9.1 Primary Endpoint — Continuous Equivalence

Evaluated on the holdout test set only (Months 10–12). All analyses on log₁₀-transformed values unless otherwise specified.

Test	Method	Pre-Specified Threshold
Pearson correlation	r between log10(Lume) and log10(Colilert) on holdout set	r ≥ 0.80 [proposed; to be confirmed with EPA Region 8]
Coefficient of determination	R² on holdout set	R² ≥ 0.65 [proposed]
Mean absolute percentage error	MAPE in log-transformed space on holdout set	MAPE ≤ 15% [proposed]
Systematic bias	Geometric mean ratio: median(Lume/Colilert) on holdout set. Test for bias using paired t-test on log differences.	Geometric mean ratio between 0.67 and 1.50 (i.e., within a factor of 1.5×); paired t-test p > 0.05 for no significant bias [proposed]
Limits of agreement	Bland-Altman analysis: mean difference ± 1.96 SD of differences (log10 scale)	95% of differences within ±1.0 log10 unit (factor of 10×); reported for transparency [proposed]
Formal equivalence test	Two one-sided t-tests (TOST) on log10 differences. Equivalence margin: ±0.5 log10 units (≈ factor of 3×; comparable to Colilert’s own inter-lab variability of 20–30%)	Both one-sided tests significant at p < 0.05 within ±0.5 log10 equivalence margin [proposed]

9.2 Secondary Endpoint — Classification at Reg 93 Threshold

Binary classification performance at the Regulation 93 geometric mean threshold of 126 CFU/100 mL, evaluated on the holdout test set.

Metric	Definition	Pre-Specified Threshold
Sensitivity (true positive rate)	P(Lume ≥126 | Colilert ≥126)	≥0.85 [proposed — protective; missed exceedances are the higher public health risk]
Specificity (true negative rate)	P(Lume <126 | Colilert <126)	≥0.80 [proposed]
Cohen’s kappa	Agreement beyond chance on binary classification	≥0.70 [proposed — "substantial agreement"]
Overall accuracy	(TP + TN) / N	≥0.85 [proposed]
Area under ROC curve	AUC from continuous Lume predictions vs. binary Colilert threshold outcome	≥0.80 [proposed]
False negative rate	P(Lume <126 | Colilert ≥126)	≤0.15 [proposed — complement of sensitivity; protects against under-reporting exceedances]

Classification performance will also be evaluated at the Reg 93 single-sample maximum (235 CFU/100 mL) and reported as supplementary analyses. Confusion matrices for each site and season will be reported in the full comparability report.

9.3 Precision Analysis

Analysis	Data Source	Metric
Temporal within-sensor precision	Readings during stable baseflow periods (identified post-hoc by <5% CV over 24 hours)	Coefficient of variation (CV) of consecutive readings; characterizes instrument noise floor vs. real environmental signal.
Reference method precision	Field duplicate Colilert samples (every 10th sample)	RPD between field duplicates. Documents Colilert’s own measurement uncertainty at Boulder Creek, providing context for the equivalence margin.

9.4 Stratified Performance Analyses

All primary and secondary endpoint analyses will be repeated for the following stratifications, reported as supplementary analyses in the comparability report:

• By site (BC-1 through BC-6): identifies site-specific performance differences or outliers
• By season (spring/summer/fall/winter): documents seasonal performance; spring runoff reported separately
• By flow condition (baseflow / rising limb / peak / recession): defined using USGS stream gauge data from the nearest Boulder Creek gauge
• By concentration range (<10 / 10–126 / 126–235 / >235 CFU/100 mL): performance at each regulatory decision threshold
• By turbidity range (<10 / 10–50 / >50 NTU): documents turbidity interference effects

9.5 Statistical Software and Reproducibility

All statistical analyses will be performed in R (version ≥4.3) or Python (scikit-learn, scipy, statsmodels). Analysis code will be version-controlled and provided to EPA Region 8 upon request. All results will be reproducible from the archived raw data using the provided code. No analyses will be performed in Excel alone; Excel may be used for data entry and visualization only.

10. Method Detection Limit

10.1 TLF Detection Limit

The instrument detection limit for TLF will be determined per 40 CFR Part 136 Appendix B: minimum 7 replicate measurements of a tryptophan standard near the expected detection limit, run in both laboratory (controlled tryptophan solution in DI water) and field (low-concentration upstream reference site BC-1) conditions. MDL = t-value (α = 0.01, n−1 degrees of freedom) × standard deviation of replicates.

10.2 E. coli MDL (Translated)

The TLF MDL will be translated to an E. coli MDL in CFU/100 mL by applying the fitted prediction equation. The method reporting limit (MRL) will be set at 3× the MDL, consistent with standard analytical chemistry practice. Colilert results below the Lume MRL will be reported as <MRL in the paired dataset and treated as left-censored in regression analyses.

10.3 Relationship to Regulatory Thresholds

The MDL and MRL will be compared against the Reg 93 thresholds (126 CFU/100 mL geometric mean; 235 CFU/100 mL single-sample maximum) and the existing literature MDL of approximately 10 CFU/100 mL for TLF-based methods. The method must demonstrate that the MRL is substantially below the regulatory threshold to support the compliance monitoring claim.

11. Interferences & Limitations

11.1 Known Interferences

Interference	Mechanism	Mitigation in This Method
Dissolved organic carbon (DOC) / humic-like fluorescence	Humic and fulvic acids can fluoresce at wavelengths overlapping the TLF signal, causing false positive E. coli estimates	Partially mitigated by turbidity correction in the prediction equation. The magnitude of humic interference at Boulder Creek is characterized during the study by comparing TLF at low-E. coli / high-DOC conditions (e.g., late fall leaf litter input) against Colilert results.
High turbidity	Very high turbidity (>[threshold NTU]) can attenuate fluorescence signal (inner filter effect) or scatter light into detector	Turbidity is included as a model predictor. Readings above the identified interference threshold are flagged (Section 7.2). Performance at high turbidity conditions is characterized separately.
Temperature	TLF signal intensity is temperature-dependent; fluorescence efficiency decreases with increasing temperature	Temperature included as a model predictor. Study spans all four seasons, ensuring temperature effects are represented in the calibration dataset.
Algal fluorescence	Phytoplankton can produce fluorescence signals in the TLF spectral region, particularly during algal blooms	Algal events documented in field conditions log. Performance during identified algal bloom periods analyzed separately. Boulder Creek turbid spring runoff typically limits algal growth.
Biofouling	Biological growth on the optical window attenuates signal over time	Bi-weekly cleaning protocol (Section 7.1); fouling detection via daily drift monitoring (Section 7.2); maintenance rotation minimizes gaps.
Chlorinated effluent	WWTP effluent may carry chlorine residual that partially inactivates E. coli but does not immediately destroy TLF signal, causing over-prediction	Site BC-3 (below WWTP) receives residual-chlorinated effluent intermittently. Performance at BC-3 analyzed separately; any systematic bias at this site documented and disclosed.

11.2 Method Limitations

• The prediction equation is calibrated to Boulder Creek water chemistry. Performance at sites with substantially different DOC, humic content, or conductivity is not validated by this study and should not be assumed without site-specific evaluation.
• The method measures total TLF, not E. coli specifically. The prediction equation uses turbidity and temperature to correct for non-E. coli TLF sources, but cannot distinguish E. coli fluorescence from all potential interfering substances.
• The method does not distinguish viable from non-viable E. coli. Post-chlorination over-prediction (documented at BC-3) reflects this limitation.
• Sensor performance during ice formation or under ice cover has not been characterized. Data collected when water temperature is <1 °C should be considered provisional.

12. Deliverables & Schedule

12.1 Pre-Study Prerequisites

• All 6 Lume sensors deployed and transmitting at Boulder Creek TMDL monitoring sites
• Certified reference laboratory confirmed and under contract for increased Colilert sampling frequency
• Colilert version (standard / Colilert-18) confirmed and consistent with Boulder’s existing Reg 93 program
• This study design document submitted to and acknowledged by EPA Region 8 (study design pre-registration)
• Statistical analysis plan (Section 9) accepted by EPA Region 8 in writing, including pre-specified thresholds
• Chain-of-custody and field conditions logging protocols in place and tested

12.2 Study Timeline

12.3 Deliverables

Deliverable	Recipient	Timing
Study design document (this document)	EPA Region 8, CDPHE, City of Boulder	Pre-study
Quarterly interim data reports	EPA Region 8, CDPHE	Months 3, 6, 9
Model fitting report (locked coefficients)	EPA Region 8	Month 9
Full statistical analysis report (holdout)	EPA Region 8, CDPHE, City of Boulder	Month 13
Boulder Creek comparability report	EPA Region 8, CDPHE	Month 14
Complete method documentation (EPA format)	EPA Region 8	Month 15
40 CFR 136.5 application package	EPA Region 8	Month 16

13. Open Questions for Regulators & Partners

The following questions require answers before the study clock can start. They are directed to EPA Region 8, CDPHE, and City of Boulder respectively. Answers will be incorporated as amendments to this document.

13.1 Questions for EPA Region 8

13.2 Questions for CDPHE

13.3 Questions for City of Boulder Utilities