Caption accuracy standards for vendor contracts: how DCMP, FCC, and WCAG 1.2 benchmarks translate into enforceable SLA clauses, audit protocols, and remediation rights

The three external standards and what each actually says

Three external frameworks have defined caption accuracy in ways that are operationally useful for contract purposes. They are not interchangeable — they emerged from different contexts, apply to different types of content, and have different legal authority over an enterprise training caption programme. Understanding each one before writing a contract clause is the difference between a clause that references a real standard and one that references a number with no definition behind it.

DCMP: 99% accuracy for educational media

The Described and Captioned Media Program is a federally funded initiative administered by the National Association of the Deaf that provides captioned educational media for students who are deaf or hard of hearing. The DCMP Caption Quality Standards, published and periodically updated, define the benchmark that captioned media must meet to be distributed through the DCMP catalogue. The core accuracy requirement is 99% word accuracy — defined as 99 correct words for every 100 words spoken in the source audio.

DCMP’s standard is operationally precise in ways most enterprise SLA clauses are not. The measurement methodology is word-error rate (WER): the number of incorrect words divided by the total words in the source transcript, expressed as a percentage of accuracy. An error is any word that is substituted, deleted, or inserted relative to the source audio. Punctuation errors are evaluated separately and do not count toward the core WER calculation in DCMP’s primary metric. The 99% WER threshold means that a 10-minute video with 1,500 spoken words can have no more than 15 errors and still meet the standard.

The DCMP standard is the most directly applicable external benchmark for enterprise L&D content because it was designed for the exact type of content L&D teams produce: educational video delivered to learners who may include deaf or hard-of-hearing participants. It has definitional clarity that the FCC broadcast standard (which targets live TV, not pre-recorded educational media) and the WCAG criterion (which sets no numerical threshold) do not provide at the same level of specificity. When an L&D compliance officer writes a vendor accuracy SLA clause and needs to reference an external standard, DCMP is the right citation — not because it has legal authority over the vendor relationship, but because it provides the definitional precision the clause requires.

FCC: 100% accuracy target with a "substantially complete" doctrine

The Federal Communications Commission’s closed caption quality rules, codified at 47 CFR § 79.1, govern broadcast television captions. The FCC’s accuracy standard is 100% — captions must be a verbatim representation of the audio. In practice, the Commission has developed a “substantially complete” doctrine through enforcement proceedings: captions that omit or inaccurately represent a material portion of the audio are not substantially complete and violate the rule, while minor deviations that do not impair understanding of the programme may be treated as non-violating.

The FCC standard is important for two reasons in an L&D contract context, even though the FCC has no authority over enterprise training content. First, the 100% target with a substantially-complete doctrine is the structural model that the most rigorous enterprise accuracy SLA clauses should follow: set a verbatim-accuracy target, then define what level of deviation is permissible without triggering a remediation obligation. Second, the FCC’s enforcement proceedings have produced a body of definitional precedent about what constitutes a material error, what constitutes a non-material deviation, and how accuracy is evaluated in context rather than by raw word count alone. That definitional work is directly transferable to enterprise SLA drafting.

The FCC enforcement model distinguishes four types of caption quality failure: accuracy (wrong words), completeness (words omitted), synchrony (captions out of time with audio), and program-completeness (captions missing entirely from portions of the programme). An enterprise accuracy SLA that follows the FCC’s four-dimension model — rather than a single WER metric — captures the full quality picture. A caption file can be 99.5% accurate by WER and still fail synchrony requirements if the captions are systematically late, or fail completeness requirements if the vendor omitted a speaker’s lines because they were difficult to transcribe.

WCAG 1.2.2: "accurate" with no percentage — and how case law filled the gap

WCAG 2.1 Success Criterion 1.2.2 requires that pre-recorded audio content include captions that are “accurate.” The criterion does not define accurate numerically. The normative text says captions must “include all dialogue and all important sounds” and must be synchronised with the audio track. There is no stated percentage threshold in the WCAG specification itself.

The operational threshold for enterprise L&D has been established through a combination of OCR resolution agreements and disability rights litigation. The consistent pattern across resolution agreements from 2017 through 2025 is that OCR has not accepted caption accuracy below 99% as meeting the “accurate” standard in WCAG 1.2.2. Resolution agreements from several major university cases (National Association of the Deaf v. Harvard University, settled 2020; NCFE v. University of California system, concluded 2022; resolution agreements with multiple community college systems 2023–2025) have all treated 99% word accuracy as the minimum standard, with several requiring 99.5% or higher for technical content with specialised vocabulary.

For contract purposes, the WCAG 1.2.2 standard is best cited not as a standalone accuracy threshold but as the legal obligation that makes the accuracy SLA clause necessary. The clause structure should be: “Content delivered under this agreement must meet WCAG 2.1 Success Criterion 1.2.2. Accuracy is measured against the DCMP Caption Quality Standard of 99% word accuracy as defined in the DCMP Quality Indicators, which the parties agree represents the minimum threshold for compliance with the WCAG 1.2.2 ‘accurate’ requirement.” This language ties the contractual obligation to the legal requirement while providing the definitional precision WCAG does not supply on its own.

How the standards have been operationalised in OCR resolution agreements

OCR resolution agreements are the most useful source of operational detail for enterprise SLA drafting because they emerge from real compliance failures, are negotiated between institutions that understand the practical measurement challenges, and specify not just the threshold but the measurement methodology. Three operational elements recur across resolution agreements and translate directly into contract language.

Sample size and selection methodology

OCR resolution agreements consistently specify that accuracy is assessed against a sample of content, not the full corpus. The typical sampling requirement in agreements from the 2020–2025 period is 10% of captioned content, selected randomly, evaluated on a per-minute basis. For a university with a media library of 2,000 hours of captioned content, the sample required for an annual accuracy audit is 200 hours — a substantial but manageable measurement task.

For enterprise L&D, per-module sampling is more operationally tractable than per-hour sampling. The translation is: for each course in scope, select three random 60-second segments distributed across the beginning, middle, and end of the module. Evaluate each segment against a reference transcript. Calculate accuracy across all selected segments. This methodology is auditable, reproducible, and requires a manageable number of reference transcripts relative to the total content volume.

The per-segment selection should be truly random, not structured. A vendor who knows the auditor always samples the first minute of each segment can optimise the opening segment without improving overall quality. Contract language should specify: “Segments for accuracy evaluation shall be selected using a pseudo-random algorithm seeded by the content title and a monthly audit key, which shall be disclosed to Vendor no earlier than 48 hours before the audit evaluation begins.”

Error classification: what counts as an error

OCR resolution agreements do not always specify error classification explicitly, but the DCMP standard and the FCC enforcement doctrine together provide a defensible two-tier framework. This distinction is critical for contract drafting: if every departure from the verbatim transcript counts equally, a vendor loses points for “gonna” vs. “going to” at the same rate as for “contraindicated” being transcribed as “controlled.” The two are not equivalent from a compliance standpoint.

Critical errors are substitutions, omissions, or insertions that materially impair comprehension or that replace a word with a different word having different meaning. Examples in an L&D context: a drug name replaced with a phonetically similar but different drug name; a safety procedure step omitted entirely; a number misheard as a different number in a financial compliance training module; a speaker’s name or title replaced with incorrect text. Critical errors count at a 1:1 ratio against the accuracy score — one critical error is one word error in the WER calculation.

Non-critical errors are deviations that do not impair understanding of the substantive content: contractions written out (“can’t” vs. “cannot”), filler words omitted or added (“uh,” “you know”), minor punctuation variation, or capitalisation inconsistencies. Non-critical errors count at a 0.5:1 ratio — two non-critical errors equal one error in the WER calculation — or, in a simplified version, are excluded from the accuracy calculation entirely.

Contract language: “Word accuracy shall be calculated as: (Total Words in Sample − Critical Errors − (Non-Critical Errors × 0.5)) ÷ Total Words in Sample × 100. Critical errors are defined as any substitution, deletion, or insertion that changes the meaning of the captioned content or impairs a reasonable viewer’s comprehension of the substantive information being communicated. Non-critical errors include stylistic variations, filler word transcription choices, and punctuation deviations that do not affect comprehension.”

Measurement reference: what the caption is compared against

A word-error-rate calculation requires a reference transcript — a verbatim record of what was actually said in the audio. The reference transcript is the ground truth against which the caption file is evaluated. In most OCR resolution agreement contexts, the reference transcript is produced by the institution’s own reviewers or an independent third party, not by the caption vendor. If the vendor produces the reference transcript and uses it to evaluate their own accuracy, the measurement is circular.

Enterprise contract language must specify who produces the reference transcript and whose transcript controls in a dispute. The cleanest approach: “Reference transcripts for accuracy evaluation shall be produced by Customer or an independent third party designated by Customer. Vendor-produced transcripts may be submitted as supplementary evidence but shall not be used as the basis for accuracy calculation without Customer’s written consent. In the event of a dispute about the accuracy calculation, Customer’s reference transcript shall control unless Vendor provides evidence of a material transcription error in Customer’s transcript within 10 business days of receiving Customer’s audit report.”

The anatomy of an enforceable accuracy SLA clause

A clause that says “Vendor will deliver captions with 99% accuracy” has six missing elements that make it unenforceable in practice. Each element must be present for the clause to be auditable. Our caption vendor SLA contract review checklist covers the full contract framework; this section focuses specifically on accuracy clause anatomy.

Element 1: the named measurement standard

The clause must specify which standard defines accuracy. Do not write “industry standard accuracy” or “generally accepted captioning accuracy.” Write the standard name. The strongest formulation for an L&D contract: “Accuracy shall be measured in accordance with the DCMP Caption Quality Standards (current version as published at dcmp.org), which define accuracy as the percentage of words in the caption file that exactly match the reference transcript. The minimum acceptable accuracy threshold is 99.0% as defined in the DCMP standard.”

Why DCMP rather than WCAG 1.2.2 or FCC? Because DCMP is the only one of the three standards that provides a named, publicly published, operationally complete measurement methodology for pre-recorded educational media. WCAG 1.2.2 provides the legal obligation; DCMP provides the measurement tool. WCAG 1.2.2 should appear in the contract as the legal requirement being met, with DCMP cited as the measurement standard used to verify compliance. The FCC standard is useful for the error-classification doctrine but does not apply directly to non-broadcast content.

Element 2: the sample methodology

The clause must specify how many minutes of content are sampled per audit period, how segments are selected, and what the minimum sample size is per module. Recommended language: “Accuracy shall be assessed against a sample of not less than 3 randomly selected 60-second segments per module, distributed across the opening, middle, and closing thirds of the module. For audit periods covering more than 50 modules, Customer may elect to evaluate a stratified random sample of not less than 20% of modules in scope, with not fewer than 3 segments per evaluated module. Segment selection shall be documented and provided to Vendor with the audit report.”

Element 3: the error classification framework

The clause must specify how errors are classified. As described above, a two-tier framework (critical and non-critical) is more defensible than a single-tier WER. The classification definitions must be included in the contract or in an exhibit. Avoid vague language like “material errors” without defining materiality. Our error rate calculator applies these definitions to real caption files and can generate an audit-ready accuracy report.

Element 4: audit frequency and reporting

The clause must specify how often accuracy is formally measured and what the vendor must provide. A three-tier audit structure works well in practice:

• Monthly spot-check: Customer evaluates 5 randomly selected segments from content delivered in the prior month. Result is logged but does not trigger SLA remediation unless accuracy falls below 97% (a two-percentage-point buffer below the SLA threshold).
• Quarterly formal audit: Customer or designated third party evaluates the full sample methodology across all content delivered in the quarter. Accuracy results are reported to Vendor within 15 business days of quarter close. Results below 99% trigger the remediation provisions.
• Annual full-corpus review: Vendor provides a self-assessed accuracy report covering all content delivered in the calendar year, using the measurement methodology specified in the contract. Customer has the right to audit-verify the self-assessment on a 20% sample.

Reporting language: “Vendor shall provide a monthly accuracy attestation within 10 business days of month close, certifying that all content delivered in the month met the 99.0% DCMP threshold. Attestation shall identify the measurement methodology used, the sample size evaluated, the calculated accuracy rate, and any modules that were below threshold with the cause identified.”

Element 5: remediation timeline

The clause must specify what the vendor must do when accuracy falls below threshold and how long they have to do it. A reasonable remediation structure: “If accuracy for any delivered module falls below 99.0%, Vendor shall (a) notify Customer within 3 business days of identifying the deficiency; (b) deliver a corrected caption file within 5 business days of notification; and (c) provide a root-cause analysis and corrective action plan within 10 business days. Corrected files shall be re-evaluated against the measurement methodology, and the accuracy score for the module shall be updated in the audit log. Re-delivery does not reset the SLA measurement period.”

The re-delivery timeline matters: five business days is achievable for AI-assisted re-captioning with human review; 10 business days may be required for human-only workflows on long-form content. Align the timeline to the vendor’s actual workflow, then hold them to it. The root-cause analysis requirement is valuable because it forces the vendor to identify whether the issue is a workflow problem (that will recur) or an isolated input problem (the audio quality for a specific recording was unusually poor).

This remediation structure connects to the audit rights framework covered in detail in our caption vendor audit rights and examination evidence post, which covers how to structure the audit right itself and what documentation the vendor must produce when an audit is invoked.

Element 6: consequences of non-compliance

The clause must specify what happens when accuracy falls below threshold and remediation fails or is repeated. A three-tier consequence structure:

• First instance: Vendor re-delivers at no additional cost and provides a corrective action plan. No financial penalty unless re-delivery fails the accuracy threshold.
• Systemic underperformance (3+ modules in a quarter below threshold, or 2+ consecutive quarters with quarter-average accuracy below 99%): Customer may apply a service credit of [X]% of the invoice for the affected quarter. Service credit does not limit other remedies. Vendor must provide a written remediation plan within 15 business days approved by a named senior contact.
• Material breach: If quarter-average accuracy falls below [97%], or if accuracy remains below 99% for [2] consecutive quarters after a written remediation plan has been submitted, Customer may terminate the agreement for material breach without penalty and without application of any early-termination fee.

The specific credit percentages and thresholds should be negotiated based on the volume and criticality of the content. For a compliance training programme where inaccurate captions create legal exposure, the consequences of material breach should include an indemnification provision for costs arising from ADA or Section 504 claims that name inaccurate captions as a contributing factor. Our caption vendor accuracy evaluation methodology post covers how to evaluate whether a vendor’s self-reported accuracy numbers are credible before the contract is signed.

The audit protocol: who runs it, what they produce, and how the chain of custody works

An accuracy SLA clause that creates an audit right is only valuable if the audit is operationally defined in a way that produces defensible results. This section covers the operational structure of the audit — the roles, the documentation, and the chain of custody that makes audit results usable in a vendor dispute or regulatory investigation.

Who runs the audit

The monthly spot-check is typically run by the L&D accessibility coordinator or a designated team member with caption quality training. The quarterly formal audit should be run either by the same coordinator using a documented methodology or by an independent third party. The annual review is the most appropriate occasion for third-party verification.

Third-party auditors bring two advantages: they are independent of the vendor relationship (removing any incentive to score charitably), and their audit reports carry more weight in a dispute or regulatory investigation than internal self-assessments. A third-party audit report that finds 98.2% accuracy is a stronger foundation for a service credit claim than an internal audit that finds the same number, because the vendor cannot attribute the result to the customer’s measurement error. Our QA methodology for training video teams post provides the operational framework for running accuracy evaluations internally.

The reference transcript

As established in the contract clause (element 2 above), the reference transcript is produced by the customer or a designated third party. The reference transcript must be created from the source audio, not from the vendor’s caption file. Evaluators who create the reference transcript by listening to the vendor’s captions while watching the video are measuring something different from what the SLA requires.

For content with technical vocabulary, the reference transcript should be reviewed by a subject matter expert before being used as the basis for accuracy evaluation. A compliance training module on pharmaceutical labelling requirements that uses specialised drug names and regulatory terminology should be reviewed by someone familiar with that vocabulary. The reference transcript is not just a verbatim record — it is the correct answer key against which the vendor’s work is graded. Our glossary architecture for AI captions post covers how vocabulary management at the glossary level reduces technical-term error rates before they reach the SLA measurement stage. Our Whisper accuracy benchmarks by vertical post provides context on typical baseline error rates by content type, which informs realistic SLA threshold-setting.

The audit log and chain of custody

Every accuracy evaluation should produce a standardised audit log entry that can be compiled into a quarterly report and referenced in a vendor dispute. The minimum fields for each audit log entry:

• Module identifier (course name, module number, internal content ID)
• Date of caption delivery
• Date of accuracy evaluation
• Evaluator name and role
• Sample segments evaluated (start time, end time for each segment)
• Total words in sample
• Critical errors identified (with type: substitution, deletion, insertion)
• Non-critical errors identified
• Calculated accuracy rate
• Pass/fail against SLA threshold
• Reference transcript file identifier
• Caption file version identifier (to confirm which file was evaluated)

The audit log is a compliance document, not just an internal quality record. If a regulatory investigation or civil litigation draws on caption quality records (covered in depth in our caption records e-discovery post), the audit log is evidence of the institution’s quality management programme. An audit log that shows the institution was systematically measuring accuracy, identifying failures, and enforcing vendor remediation obligations is a materially better compliance posture than one that relies on the vendor’s own attestations.

Vendor notification and response

The contract should specify how the vendor is notified of audit results, the timeline for the vendor to respond, and what the vendor’s response must include. Reasonable language: “Customer shall provide Vendor with audit results within 15 business days of completing the evaluation. Audit results shall include: (a) the audit log summary; (b) the segments evaluated with start/end times; (c) the list of errors identified with classification; and (d) the calculated accuracy rate. Vendor shall acknowledge receipt within 2 business days and shall provide any written objection to the accuracy calculation within 10 business days. Objections not submitted within 10 business days are waived. Vendor may request review of up to 3 specific error classifications per quarterly audit, with reasons stated. Customer will respond to objections within 10 business days, and Customer’s determination is final for SLA purposes.”

Remediation rights and what happens when the vendor cannot recover

The remediation rights clause answers the question: after the vendor fails the accuracy threshold and delivers a corrected file, what are the customer’s rights if the corrected file also fails, or if the vendor cannot explain why the error rate was so high? Our caption vendor pilot program design post covers how to stress-test a vendor’s accuracy claims before full contract commitment, but the remediation rights clause governs what happens after the relationship is underway.

Re-delivery and reset of the SLA clock

A critical contract point: re-delivery of a corrected file should not reset the original delivery date for SLA measurement purposes. If the vendor delivers a module on 1 July with 97.2% accuracy and delivers a corrected version on 8 July with 99.4% accuracy, the content was not available at SLA-compliant accuracy for seven days. That gap has real operational consequences: if a learner was assigned the module during those seven days, they consumed sub-threshold content. The SLA should track whether the original delivery met threshold, not only whether the final delivered version meets threshold.

Language: “Re-delivery of a corrected file does not constitute retroactive compliance for the original delivery. The original delivery accuracy rate shall be recorded in the audit log for the period in which the original file was delivered. Re-delivery accuracy is tracked separately and is used to assess whether the vendor cured the deficiency within the required timeline.”

Systemic underperformance: when patterns matter

A single module at 98.5% is a threshold miss but is not evidence of a systemic programme failure. Three modules at 97% in the same quarter, or a consistent pattern of barely-clearing-threshold results on technical content, is evidence that the vendor’s quality programme is not calibrated for the content type. The contract should give the customer a right to escalate based on patterns, not just individual failures.

Pattern language: “If more than [10%] of modules evaluated in any quarter receive an accuracy score below the 99.0% threshold, or if any module receives an accuracy score below [96.0%], the parties shall convene a quality review meeting within 15 business days. Vendor shall present a written Quality Improvement Plan (QIP) identifying root causes and proposed corrective actions. The QIP shall include: (a) root-cause analysis of each sub-threshold result; (b) specific workflow changes to address the identified causes; (c) a revised accuracy target for the following quarter; and (d) the name and contact information of the Vendor senior quality manager responsible for the QIP. Customer may request a QIP review at 30 and 60 days after submission.”

The feedback loop structure that makes QIP review operationally effective is covered in our caption feedback loop for iterative accuracy improvement post, which provides the workflow for systematic error tracking and vendor communication. The QIP language above creates the contractual obligation; the feedback loop framework provides the operational mechanism.

Indemnification for compliance costs

When inaccurate captions result in a regulatory complaint, an OCR investigation, or a disability rights claim, the L&D team’s institution bears the compliance costs. If the institution had an accuracy SLA with a vendor that was systematically underperforming, there is an argument that the vendor’s failure to meet the contracted standard is a contributing cause of those compliance costs. Not all vendor contracts include an indemnification provision for downstream compliance costs, and vendors will resist this language — but it is worth negotiating for in agreements covering content with known compliance exposure (required training, onboarding content, mandatory certification programmes).

Language: “Vendor shall indemnify, defend, and hold harmless Customer from and against any claims, damages, penalties, and reasonable legal fees arising directly from Vendor’s failure to deliver captions meeting the accuracy standard specified in this agreement, including but not limited to costs arising from OCR investigations, ADA or Section 504 regulatory proceedings, or civil claims where inaccurate caption delivery is named as a contributing factor, provided that Customer’s exposure is attributable to Vendor’s failure and Customer has not materially contributed to the accuracy failure through its own acts or omissions.”

Using the standards framework in RFP evaluation

The accuracy SLA framework is not only useful after a contract is signed — it is a filter for evaluating vendors before you commit. Our caption vendor RFP response evaluation post covers the full RFP evaluation methodology; this section focuses specifically on the four questions that reveal whether a vendor has a real accuracy programme.

Question 1: Which external standard do you use to define accuracy?

The correct answer names a standard with a published methodology — typically DCMP, FCC WER methodology, or a named industry body. An answer of “we define accuracy as the percentage of words that match the client-approved transcript” is not wrong but raises a follow-up: how is the reference transcript produced, and how is the comparison performed? An answer of “we have our own internal standard that delivers 99% accuracy” without naming a methodology is a red flag — an internal standard with no external reference point is not auditable against an industry benchmark.

Question 2: What is your error classification framework?

The vendor should distinguish between errors that impair comprehension and errors that are stylistic deviations. A vendor who says “all word discrepancies count equally” is using a cruder measurement tool than one who can distinguish critical from non-critical errors. The better vendors can articulate how they handle proper nouns, technical terminology, and specialised vocabulary — which is the domain where most compliance training caption failures occur. Our Whisper accuracy benchmarks by vertical post provides data on where AI-assisted captioning typically fails by content type, which helps calibrate what to ask about in an RFP for specific content types.

Question 3: How do you document your accuracy results, and what can you provide to a customer for audit purposes?

A vendor with a real quality programme produces accuracy documentation as a routine output of their workflow, not as a special effort made in response to a customer audit request. The question should reveal: (a) whether they maintain per-file accuracy records; (b) what format those records take; (c) whether they can provide per-module accuracy reports on a scheduled basis without special instrumentation; and (d) whether their records include the reference transcript and the segment-level error breakdown that a customer audit would require. A vendor who says “we can provide that” but cannot show you a sample report from an existing client is probably producing that documentation for the first time for your proposal.

Question 4: Can you agree to a 99.0% DCMP-standard accuracy SLA with the clause language above?

Present the accuracy SLA clause language from this post — or a version of it adapted to your specific requirements — and ask the vendor whether they can agree to it. The vendor who says “yes, with these two specific modifications” and explains why is a vendor who has operated under SLA accuracy contracts before and knows where the friction points are. The vendor who says “we can’t agree to external measurement methodology because our quality process is proprietary” is a vendor whose quality process cannot be audited. The vendor who says “our standard contract has accuracy language, let us substitute our clause” — read their clause with the six-element framework above as a lens, and count how many elements are missing.

Draft SLA clause text and audit log template

The following language integrates the six elements of an enforceable accuracy SLA clause. It is a starting point for negotiation, not a finished contract. Legal review is required before use in a specific vendor agreement. Bracketed terms should be completed based on your specific content volume, vendor workflow, and risk tolerance.

Draft accuracy SLA clause

Caption Accuracy Standard. All captions delivered under this Agreement for pre-recorded video content shall achieve a minimum word accuracy rate of 99.0% (“Accuracy Threshold”), measured in accordance with the Described and Captioned Media Program (DCMP) Caption Quality Standards as published at dcmp.org, which the parties agree represents the minimum standard for compliance with WCAG 2.1 Success Criterion 1.2.2.

Measurement Methodology. Word accuracy is calculated as: (Total Words in Evaluated Sample − Critical Errors − (Non-Critical Errors × 0.5)) ÷ Total Words in Evaluated Sample × 100. “Critical Errors” means any substitution, deletion, or insertion that changes the meaning of the captioned content or materially impairs a reasonable viewer’s comprehension of the substantive information being communicated. “Non-Critical Errors” means stylistic variations, filler word transcription choices, and punctuation deviations that do not affect comprehension.

Audit Sample. Accuracy shall be evaluated against a minimum of three (3) randomly selected 60-second segments per module, distributed across the opening, middle, and closing thirds of each module. For quarterly audits covering more than 50 modules, Customer may elect to evaluate a stratified random sample of not less than 20% of modules, with not fewer than 3 segments per evaluated module. Reference transcripts shall be produced by Customer or a Customer-designated third party from source audio, independent of Vendor’s caption file. Vendor-produced transcripts shall not be used as the accuracy reference without Customer’s written consent.

Audit Frequency and Reporting. Customer shall conduct quarterly formal accuracy audits. Vendor shall provide a monthly accuracy attestation within 10 business days of month close. Customer shall provide quarterly audit results to Vendor within 15 business days of quarter close. Vendor may submit written objections to specific error classifications within 10 business days of receiving Customer’s audit report. Objections not submitted within 10 business days are waived. Customer’s accuracy determination is final for SLA purposes subject only to manifest mathematical error.

Remediation. If accuracy for any delivered module falls below the Accuracy Threshold, Vendor shall (a) notify Customer within 3 business days of identifying the deficiency; (b) deliver a corrected caption file within 5 business days of notification; and (c) provide a root-cause analysis and corrective action plan within 10 business days of notification. Re-delivery of a corrected file does not constitute retroactive compliance for the original delivery period. If more than [10%] of modules evaluated in any quarter fall below the Accuracy Threshold, or if any module receives an accuracy score below [96.0%], the parties shall convene a quality review meeting and Vendor shall provide a written Quality Improvement Plan within 15 business days.

Consequences of Non-Compliance. (a) First and isolated instance: Vendor re-delivers corrected file at no additional cost. (b) Systemic underperformance (as defined above): Customer may apply a service credit of [5]% of the invoice for the affected quarter. (c) Material breach: if quarter-average accuracy falls below [97.0%], or if accuracy remains below the Accuracy Threshold for [2] consecutive quarters following submission of a Quality Improvement Plan, Customer may terminate this Agreement for material breach without penalty or early-termination fee. Service credits do not limit other remedies available to Customer.

Audit log template fields

Each accuracy evaluation entry should capture: module identifier, date of caption delivery, date of evaluation, evaluator name and role, segments evaluated (start time / end time / word count for each), total words in sample, critical errors (count and list with type), non-critical errors (count and list), calculated accuracy rate, pass/fail against 99.0% threshold, reference transcript file identifier, caption file version identifier, and notes on unusual content features (heavy technical vocabulary, multiple speakers, background noise) that may explain the error distribution. The full audit log for a quarter is the evidence base for service credit claims and the compliance record for regulatory purposes.

Connecting accuracy standards to the rest of the vendor contract

The accuracy SLA clause does not operate in isolation — it is one element of a complete vendor contract for caption services. The full contract framework is covered in our caption vendor SLA contract review checklist. The connections worth highlighting here:

Vendor accuracy evaluation before signing. The accuracy SLA clause defines the ongoing obligation. The pre-signature evaluation process — covered in caption vendor accuracy evaluation methodology — verifies that the vendor can actually meet the threshold before you commit to a multi-year agreement. Signing a contract with a 99% accuracy SLA with a vendor that historically delivers 97.5% is not a quality programme — it is a credit-generating exercise.

Pilot programme before full rollout. Our caption vendor pilot program design post covers how to structure a paid pilot with a sample of your actual content before deploying the accuracy SLA framework at scale. A pilot surfaces whether the vendor’s accuracy holds on your specific content type — technical vocabulary, speaker accents, audio quality — rather than on the vendor’s curated test set.

Glossary and terminology management. Many accuracy failures on technical content are not vendor quality failures — they are terminology gaps. A vendor who does not have your organisation’s product names, regulatory citations, and technical vocabulary in their glossary will consistently transcribe those terms incorrectly. Our glossary architecture for AI captions post covers how to build and maintain the terminology base that makes vendor accuracy claims achievable on your content. The error rate calculator in our caption quality error rate calculator allows you to quantify the accuracy gap that terminology management alone can close.

Audit rights and examination evidence. The audit protocol described in this post operates within the broader audit rights framework covered in caption vendor audit rights and examination evidence. The accuracy audit is one component of the vendor examination process; the full audit rights framework also covers deliverable standards, format compliance, turnaround time performance, and glossary update obligations. The accuracy audit log is one element of the examination evidence package that the audit rights clause entitles you to.

Feedback loop for continuous improvement. An accuracy SLA that generates service credits and QIPs but does not drive systematic improvement in vendor performance is expensive to administer and produces diminishing returns. The caption feedback loop for iterative accuracy improvement post provides the operational framework for converting audit results into vendor improvement — the structured communication protocol, the error categorisation system, and the performance review cadence that transforms monthly QA data into measurable accuracy gains over time.

Frequently asked questions

Does DCMP's 99% standard have legal authority over our vendor contract?

No. DCMP is a federally funded programme that defines quality standards for media it distributes through its own catalogue. It does not have regulatory authority over private caption vendor agreements. Its authority in your contract is the authority you give it by referencing it — it provides definitional precision and an external benchmark, not legal compulsion. The legal obligation in your vendor contract flows from your institution’s ADA, Section 508, and WCAG 2.1 AA obligations. DCMP provides the measurement methodology that verifies compliance with those obligations.

Can we use FCC accuracy methodology even though we're not a broadcaster?

Yes. The FCC’s measurement methodology — particularly the four-dimension quality framework (accuracy, completeness, synchrony, programme-completeness) and the critical/non-critical error distinction — is applicable to pre-recorded enterprise training content. The FCC’s jurisdiction is limited to broadcast; its measurement tools are public domain and can be adapted for any caption quality programme. The accuracy methodology in this post incorporates elements of the FCC framework adapted for the enterprise L&D context.

Our vendor says their AI model inherently delivers 99% accuracy. Do we still need a SLA clause?

Yes, and the answer is in the question. “Our model delivers 99% accuracy” is a claim about a general average across the model’s training distribution. Your content may have a different accuracy profile than the training distribution, particularly if it contains technical vocabulary, specialised speakers, or non-standard audio quality. The accuracy SLA clause is the mechanism by which you measure whether the model’s general accuracy claim holds for your specific content, and the mechanism by which you enforce the obligation if it does not. A vendor who resists a measurable accuracy SLA is a vendor who does not want their general accuracy claim tested against your content.

What happens if our audio quality is poor and the vendor claims they can't meet 99% because of our content?

Audio quality is a legitimate contributing factor to accuracy performance. The contract should address this explicitly: “If Vendor determines that a specific module’s source audio quality is insufficient to achieve the Accuracy Threshold through reasonable captioning effort, Vendor shall notify Customer within [5] business days of receiving the file, identify the specific audio quality issues, and propose a remediation path (e.g., audio enhancement, human-only transcription, extended review time) before captioning proceeds. Modules captioned without such notification are subject to the standard Accuracy Threshold regardless of audio quality.” This language requires the vendor to flag the problem before captioning, not use audio quality as a post-hoc excuse for a failed accuracy audit.

Should the 99% threshold apply to live captioning as well as pre-recorded content?

No — live captioning has different standards because real-time transcription cannot achieve pre-recorded accuracy rates. CART (Communication Access Realtime Translation) services have their own accuracy standards, and the ADA reasonable accommodation framework for live events applies a different standard than WCAG 1.2.2 (which covers pre-recorded content only; live captioning is covered by WCAG 1.2.4, which also does not specify a percentage). Our real-time CART captioning for live training post covers live captioning quality standards and procurement separately. The accuracy SLA framework in this post applies to pre-recorded video content only.

How long should we retain audit logs?

Retain accuracy audit logs for at least as long as your standard document retention period for compliance records, and for at least three years after the last content in the audit was retired from active delivery. The audit log is part of your accessibility compliance documentation and may be relevant in OCR document requests or civil litigation e-discovery if captioning accessibility is challenged. The specific retention obligation for caption-related compliance records is covered in our caption records in civil e-discovery post. When in doubt, retain longer — audit logs are small and the cost of over-retention is low relative to the cost of not having them when needed.

Can we reference DCMP and WCAG 1.2.2 in the same contract clause without creating a conflict?

Yes, and the recommended approach is to use them for different purposes: WCAG 1.2.2 as the legal obligation (the requirement that the SLA clause is designed to meet), and DCMP as the measurement standard (the methodology used to verify compliance with the requirement). The clause language in this post uses this structure explicitly: “the DCMP Caption Quality Standard of 99% word accuracy, which the parties agree represents the minimum threshold for compliance with the WCAG 1.2.2 ‘accurate’ requirement.” This framing ensures that if WCAG standards evolve or OCR enforcement tightens the de facto compliance threshold above 99%, the contract language is not obsoleted — it references the legal requirement and the current measurement tool, not a fixed percentage decoupled from the underlying obligation.