--- title: "FASTQ Quality Scores: Decode the Characters Before Comparing Errors" description: "Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20." canonical: "https://www.leviathansystems.co/articles/genomics-ai-fastq-quality-encoding-error-estimates" last-updated: "2026-10-20" --- # FASTQ Quality Scores: Decode the Characters Before Comparing Errors Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20. A FASTQ quality character becomes a meaningful Q-score only under the declared encoding. That score expresses an estimated per-base error probability; it does not record whether that base was actually wrong. When comparing genomics-AI input offers, establish that the required base calls, corresponding qualities and their interpretation survive delivery before drawing any conclusion about observed errors or model accuracy. ## Preserve the base call and its corresponding quality An input offer can preserve a sequence while losing the information needed to interpret its qualities. If the proposed AI consumer requires both, delivery of the nucleotide letters alone leaves part of its input requirement unanswered. Begin with that consumer's definition of the input, including whether and how it uses per-base quality. NCBI's SRA File Format Guide describes FASTQ as read-identifying text, nucleotide base calls, a second identifying line and per-base qualities. The quality string must match the sequence length. For character-encoded Phred scores, it documents two predominant ASCII offsets: 33 and 64. FASTQ has historical variations, so a file-format label alone does not settle the encoding. Position matters as well as length. The quality attached to the first called base concerns that call, and so on. Equal-length strings do not establish that a delivery preserved their original association. The buyer needs evidence connecting the delivered calls and qualities to the consumer's required positions and declared interpretation. References: [SRA](https://www.ncbi.nlm.nih.gov/sra/docs/submitformats/) ## FASTQ-ENC-01: the same I can be decoded as Q40 or Q9 Consider original fictional example FASTQ-ENC-01. Its sequence is ACGT and its four quality characters are IIII. The input is explicitly declared to use Phred+33. These invented letters and characters are arithmetic examples, not human sequence data, an instrument run or a classification test. The named Illumina BaseSpace Sequence Hub OLH009008_2 encoding page defines ASCII code = Q + 33 and lists I at code 73 as Q40. Correct decoding in this example gives 73 - 33 = 40 at each position. Applying the Phred probability relationship gives an estimated error probability of 0.0001 per base. A consumer wrongly applying a Phred+64 decoder would instead calculate 73 - 64 = Q9. Under the same Phred relationship, its interpreted probability would be 10^(-9/10), approximately 0.1259. The sequence, quality characters and positions did not change; the consumer assigned them a different meaning. For this declared +33 input, the +64 interpretation is a decoding error. It is not an alternative improvement in the data. If the encoding declaration were absent, bare IIII would not identify which historical variant produced it. A declaration should identify the quality scheme, including its score scale and offset; the examples here specifically use Phred scores. References: [IE](https://support.illumina.com/help/BaseSpace_Sequence_Hub_OLH_009008_2/Source/Informatics/BS/QualityScoreEncoding_swBS.htm), [IQ](https://support.illumina.com/help/BaseSpace_Sequence_Hub_OLH_009008_2/Source/Informatics/BS/QualityScores_swBS.htm) ## Translate Q into an estimated probability For Phred scores, Q = -10 log10(P), where P is the estimated probability of an incorrect base call. Reversing the relationship gives P = 10^(-Q/10). The BaseSpace OLH009008_2 quality-score page gives Q20 as P = 0.01 and Q30 as P = 0.001. A ten-point increase therefore divides the estimated probability by ten. It does not subtract ten percentage points. Q40 corresponds to P = 0.0001, or 0.01%, rather than an observed measurement of the fictional ACGT sequence. The distinction between prediction and observation is consequential. A character decoded as Q40 does not reveal which particular call, if any, is wrong. Observed error evidence needs a separately defined comparison and reference. Decoding the character correctly establishes what probability the score represents; it does not independently validate that prediction for the delivered population. References: [IQ](https://support.illumina.com/help/BaseSpace_Sequence_Hub_OLH_009008_2/Source/Informatics/BS/QualityScores_swBS.htm) ## Compare expected error counts without inventing read outcomes Now stipulate two separate fictional populations, each containing 1,000 called bases. Every base in population A has Q20; every base in population B has Q30. No sequencing experiment or error observation is supplied, and the bases have no stipulated grouping into reads. For A, the supplied probabilities sum to 1,000 x 0.01 = 10 expected errors. For B, they sum to 1,000 x 0.001 = 1 expected error. These are expectations implied by the supplied predictions. Neither population has a measured error count, and the calculation does not say that A contains exactly ten errors or B exactly one. This sum does not require the error events to be independent: expectation adds across positions. Its usefulness as a prediction of reality still depends on how well the supplied probabilities apply. Also, a count of base errors is different from a count of reads containing errors; multiple errors could fall within one read. Per-base probabilities alone do not establish a whole-read probability of containing no errors. That requires information or assumptions about the joint error events, together with the read's positions. Consequently, a statement that virtually all reads are perfect is unsupported by a Q30 label alone. Base-call quality also supplies no direct measurement of correct alignments, variant accuracy or a downstream model's accuracy. ## Retain the producer and quality-table applicability The score's interpretation has two layers: how the character represents a number, and how the producer assigned that number. Correctly subtracting an offset resolves the first layer. It does not establish the second. Illumina's Understanding Illumina Quality Scores, publication 770-2012-058, current as of April 23, 2014, describes quality predictors mapped through a table calibrated with empirical data. It explains that changes to hardware, software or chemistry can require a different quality table. This historical explanation supports the distinction between a base-call prediction and a direct error observation; it supplies no current instrument performance claim. Retain the producer's applicable platform, software and quality-table information where identified, along with any declared quality transformation. If scores were binned, establish the mapping applicable to that output. A bin's representative score cannot reveal the original distinctions that the mapping merged. The historical note's example bins are not a universal or current table, and seeing a small set of characters cannot identify the applicable mapping by itself. References: [I2014](https://www.illumina.com/content/dam/illumina-marketing/documents/products/technotes/technote_understanding_quality_scores.pdf) ## Decide what the input offer actually establishes Suppose two fictional offers both promise to preserve ACGT. Offer A supplies IIII too, but gives no encoding declaration or explanation of the quality source. It preserves visible text while leaving its intended numerical interpretation unresolved. The buyer cannot substantiate the requested quality interpretation by selecting whichever offset produces the more attractive number. Offer B identifies the required calls and corresponding qualities, declares Phred+33 and explains how the receiving consumer interprets the retained values. Its supporting record identifies the producer and applicable version or table information, and describes any re-encoding or binning. If that evidence holds for the actual delivered input, B addresses the representation question that A leaves open. A proposal alone is still different from evidence of what reached the consumer. A documented re-encoding need not retain identical quality characters: changing the offset while preserving the intended Q values changes the text. Conversely, retaining identical characters under a different decoder can change every interpreted probability. Judge semantic preservation against the declared scheme and any stated transformation, rather than using visual similarity as the acceptance argument. This resolves a specific purchase question: can the offer supply the input meaning the selected consumer requires? Observed error rates, alignment correctness, biological truth and downstream model performance each need evidence appropriate to their own claim. The input comparison does not prescribe a quality cutoff, laboratory method or clinical use. The related S3 integrity guide addresses a separate question: whether transfer evidence concerns the same bytes and checksum calculation. Byte agreement can complement this review, but it cannot supply a missing FASTQ encoding declaration. ## Scope a separate physical connection requirement If the selected computing and storage equipment also requires a physical cabling installation, Leviathan's Structured Cabling & Fiber Installation service provides a route to discuss agreed fiber and copper connections, endpoint labels and physical handoff records. Use the related service and project inquiry with a nonconfidential equipment outline, approved endpoints and routes, material responsibilities and required connection records. Scope and availability need confirmation. That inquiry concerns the physical installation; data engineering, bioinformatics and validation of the AI input remain with the appointed software and specialist teams. References: [L1](https://www.leviathansystems.co/services/structured-cabling) ## Primary sources - [SRA — NCBI SRA: File Format Guide, FASTQ files](https://www.ncbi.nlm.nih.gov/sra/docs/submitformats/). Checked 2026-10-20. FASTQ fields, matching sequence/quality lengths and predominant Phred ASCII offsets 33 and 64. Historical variants require their own declared interpretation. - [IE — Illumina BaseSpace Sequence Hub OLH009008_2: Quality Score Encoding](https://support.illumina.com/help/BaseSpace_Sequence_Hub_OLH_009008_2/Source/Informatics/BS/QualityScoreEncoding_swBS.htm). Checked 2026-10-20. The named guide, copyright 2020, defines ASCII = Q + 33 and I = code 73 = Q40. This does not define every historical FASTQ encoding. - [IQ — Illumina BaseSpace Sequence Hub OLH009008_2: Quality Scores](https://support.illumina.com/help/BaseSpace_Sequence_Hub_OLH_009008_2/Source/Informatics/BS/QualityScores_swBS.htm). Checked 2026-10-20. The named guide, copyright 2020, defines estimated error probability and gives the Q20 and Q30 examples. The article's populations and arithmetic are fictional. - [I2014 — Illumina: Understanding Illumina Quality Scores, publication 770-2012-058](https://www.illumina.com/content/dam/illumina-marketing/documents/products/technotes/technote_understanding_quality_scores.pdf). Checked 2026-10-20. Historical note current as of April 23, 2014: prediction, calibrated quality tables and configuration applicability. Its example bins and performance language are not adopted as current specifications. - [L1 — Leviathan Systems: Structured Cabling & Fiber Installation](https://www.leviathansystems.co/services/structured-cabling). Checked 2026-10-20. Agreed physical cabling, endpoint labels and connection records support a separately scoped installation inquiry. ## Related Leviathan resources - [S3 dataset integrity: identify the bytes and checksum calculation](/articles/ai-dataset-s3-etag-versus-checksum-integrity) - [Perplexity: identify the reference tokens behind a model score](/articles/language-model-perplexity-tokenization-evaluation-basis) - [Structured Cabling & Fiber Installation](/services/structured-cabling) - [Discuss a separate physical storage or network cabling project](/contact?service=structured-cabling&from=%2Farticles%2Fgenomics-ai-fastq-quality-encoding-error-estimates) - [Mapping Quality and Base Quality Predict Different Errors](/articles/genomics-ai-sam-mapping-quality-base-call-error-events) ## Does the equipment require a physical cabling installation? Share a nonconfidential equipment, endpoint, route and physical handoff outline to discuss installation scope, requirements and availability. [Discuss the scope](/contact?service=structured-cabling&from=/articles/genomics-ai-fastq-quality-encoding-error-estimates)