R&D Credit Documentation: What Examiners Ask For

An R&D credit is worth exactly what you can substantiate. That sentence gets repeated so often it has stopped carrying information, so here is the specific version: in an examination, the number on Form 6765 is not the thing being tested. What is being tested is whether you can connect named business components to qualifying activities to dollars of expense, using records that existed before anyone thought about a tax credit.

This post covers what the law actually requires, what evidence carries weight, what the revised Form 6765 now asks on the return itself, and the extra burden that attaches to claims made on amended returns.

Start with what the rules actually say

There is a widespread belief that §41 imposes a special “contemporaneous documentation” requirement. It does not, and the distinction matters when someone tries to sell you a compliance product on that basis.

§6001 - the general recordkeeping rule - requires every taxpayer to keep records sufficient to establish the items shown on the return. That is the operative statute.
Treas. Reg. §1.41-4(d) points back to it: a taxpayer claiming the credit must retain records in sufficiently usable form and detail to substantiate that the expenditures claimed are eligible.
The 2001 proposed regulations did contain a contemporaneous-documentation requirement. It was dropped when the final regulations were adopted in 2003.

So the requirement is sufficiency, not a specific format or a specific timestamp. What contemporaneity buys you is weight. Records created while the work happened are believed; reconstructions built after an examination notice are discounted, sometimes to zero. That is an evidentiary reality, not a statutory one - and it is why “capture it as you go” is still the right advice.

The unit of analysis is the business component

Nearly every documentation failure traces back to the same structural mistake: a study organized by department, cost center, or sprint rather than by business component (§41(d)(2)(B)).

The four-part test applies per component. Wages are allocated per component. Form 6765’s Section G reports per component. If your study cannot say this product, this process, this software, this year, it will be taken apart in the first round of information requests and rebuilt under time pressure - using the same records, but with the examiner watching. Our four-part test deep dive covers the test itself; the plain-English version is the shorter read.

Wage nexus: the number that actually gets adjusted

Wages are the dominant expense category in most §41 claims, so wages are where examinations concentrate. The statute recognizes three kinds of qualified services (§41(b)(2)(B)):

Engaging in qualified research - the engineers, scientists, and technicians doing the work.
Direct supervision of qualified research - first-line management of the actual research, not executives two levels removed.
Direct support of qualified research - people whose work directly supports the researchers, such as a technician machining a test fixture.

There is a helpful rule inside the same provision: if substantially all of an individual’s services for the year are qualified services, then all of that person’s wages count. “Substantially all” is treated as 80% (Treas. Reg. §1.41-2(d)(2)). This is a genuine simplification for a dedicated engineering team - and a trap if applied to a hybrid role by assumption rather than evidence.

The examinable question is always the same: what connects this person, this percentage, and this component? A spreadsheet where every engineer is 80% qualified, to the same decimal, in every year, is not evidence. It is a conclusion with no work behind it.

What good evidence looks like, by the system it already lives in

You almost certainly generate enough evidence. The problem is that it is scattered across systems nobody has ever mapped to a tax position.

Issue trackers (Jira, Linear, Azure DevOps) - epics and tickets that name the objective, the unknown, and the outcome. Component-level labeling here pays for itself.
Version control - pull requests, review comments, and branch histories showing alternatives built and abandoned. The rejected branch is often the best proof of a process of experimentation you will ever have.
Design docs and RFCs - the single strongest software artifact, because a good RFC states the uncertainty and the alternatives explicitly, which is exactly what §41(d)(1)(C) asks about.
Benchmarks, test reports, and evaluation results - the record of evaluating alternatives rather than merely listing them.
Engineering change orders, travelers, run sheets, and scrap logs - the manufacturing equivalents, and often better dated than anything in software. See our manufacturing guide.
Electronic lab notebooks, protocols, and batch records - the biotech equivalents.
Contracts, statements of work, and grant terms - not evidence of research, but the documents that decide whether the research was funded and therefore excluded under §41(d)(4)(H). Read these before the study, not after.
Payroll and project time data - the bridge from activity to dollars.

The work of a study is largely the work of connecting these into one chain per component. Our documentation and substantiation page covers the full evidence taxonomy and how it maps to the deliverable.

Estimates and allocations: allowed, but not for free

Real companies do not track engineering time to business components by the hour, and the law does not pretend otherwise. Courts have allowed reasonable estimates where there is a credible evidentiary basis for them. What has repeatedly failed is the other thing: allocations resting on nothing but after-the-fact recollection, uniform percentages applied across dissimilar roles, or interview summaries with no underlying records.

A defensible allocation has three properties. It is derived from something - tickets, project codes, release history, sprint assignment. It is specific - different people and different components get different answers, because that is how work actually distributes. And it is consistent - the same method, applied the same way, across years, so the pattern itself corroborates.

Form 6765 put substantiation on the return

The revised Form 6765 changed the posture from “keep records in case we ask” to “report the detail with the claim.” Section G asks for business-component-level information: the component name and type, whether software is internal-use, the information sought to be discovered (currently required only on amended returns), and the QREs assigned to that component split by category. We walk through the whole section, column by column, in Form 6765 Section G: what changed and who must file it.

As the Instructions for Form 6765 read at the time of writing:

Section G is optional for tax years beginning before 2026 and applicable for tax years beginning after 2025.
Two groups are excepted: qualified small businesses under §41(h)(3) that elect the payroll tax credit, and taxpayers with total QREs of $1.5 million or less (measured at the controlled group level) together with average annual gross receipts of $50 million or less for the prior three tax years, claiming the credit on an originally filed return.
When it applies, you report at least 80% of total QREs by business component, in descending order, capped at 50 components.

Treat those dates and thresholds as a snapshot: the phase-in has already been revised more than once. Confirm against the current-year instructions before you rely on an exception.

The practical consequence is the same either way. A study that already resolves to components populates Section G directly. A study built on department totals has to be dismantled first - and the component detail it produces under deadline is exactly the detail an examiner will test.

Amended returns carry an extra burden

A research credit claimed on an amended return is a refund claim, and a refund claim must meet specificity requirements to be valid at all - a threshold question that arrives before anyone looks at your engineering.

As the IRS has set them out, a valid research credit refund claim must:

Identify all the business components to which the §41 claim relates for that year;
Identify all research activities performed for each business component; and
Provide the total qualified employee wage expenses, total qualified supply expenses, and total qualified contract research expenses for the year.

Two earlier items - naming the individuals who performed each activity and describing the information each sought to discover - were waived for claims postmarked as of June 18, 2024. As of this writing that waiver stands, though the IRS notes the information may still be requested during an examination. The separate transition period giving taxpayers 45 days to perfect a timely but deficient claim has been extended through January 10, 2027 (IR-2025-99).

This is guidance that has been revised repeatedly since 2021. Check the current IRS position before filing - the authoritative page is IRS: required information for a valid research credit claim for refund. A study built to component level satisfies the substance of these items as a by-product; a study built on totals cannot satisfy them at all.

What an examination actually looks like

Expect information document requests aimed at the chain rather than the conclusion: the component list, the basis for the wage percentages, the underlying records for a sample of components, and the contracts for anything customer- or grant-funded. Expect interviews with the engineers themselves, not only with finance - examiners are looking for whether the technical story in the study matches how the people who did the work describe it. Expect sampling: a handful of components tested hard, with the result extrapolated.

Every Ricerca study includes Audit Protection: we prepare the audit response, the documentation package, and the technical defense of the study we issued, and we work alongside your tax preparer. Representation before the IRS, where it is needed, is performed by credentialed practitioners. Coverage is limited to the R&D study and the related federal credits and is subject to your engagement agreement.

A six-item readiness check

Can you produce a list of business components for the year, by name?
For each, can you point to records showing what was uncertain at the outset?
For each, can you show alternatives evaluated and how?
Can you explain every wage percentage by reference to something other than memory?
Have you screened every customer-funded, grant-funded, and SBIR/STTR engagement for the funded-research exclusion?
Could someone other than the author follow the chain from a component to a dollar figure without asking you a question?

If the answers are yes, the credit is a computation. If they are not, the credit is a position - and positions get tested. Start with the R&D tax credit guide, or read why automation-only studies fall short for what this looks like when it is done badly.

Sources

IRC §41 - Credit for increasing research activities (U.S. House, Office of the Law Revision Counsel)
IRC §6001 - Notice or regulations requiring records, statements, and special returns
Treas. Reg. §1.41-4 - Qualified research, including the recordkeeping rule at paragraph (d) (Cornell LII)
IRS - Instructions for Form 6765 and About Form 6765
IRS - IRS sets forth required information for a valid research credit claim for refund
IRS - Research credit