The Attribution Gap

Executive summary

Small business accounting software has a structural blind spot. A vendor bill can post to the books without being attributed to any customer, job, or project, and the profit-and-loss statement still balances. When that happens silently and often, the margin an owner reads on any given job is overstated, not because the bookkeeping is wrong on the bills that were tagged, but because the untagged bills vanish from the denominator instead of showing up as unassigned cost. This paper names that phenomenon the attribution gap and defines it as a formal, computable metric: the share of accounts-payable records lacking a valid customer, job, or project reference, reported at both record-count and dollar-volume basis.

Measuring the gap honestly is harder than the definition suggests. A single-axis measurement can mislead in opposite directions depending on the dataset. On the City of Houston's public checkbook (2,293,699 records, USD 65.1B of AP over nine years), checking only the capital-project work-breakdown-structure code reports 92.11% untagged, a structural artifact, because most municipal spending is operating cost routed through a purchase order or contract, not a capital project [Baseline]. On the French national procurement register (FOPPA, 1,380,965 records, EUR 1.75T over eleven years), checking only the buyer field reports 0.00% untagged, an opposite structural artifact, because that field is populated on every record by construction [Baseline]. The honest measurement for each dataset, a composite of the axes that dataset's own institution treats as "this cost belongs to this unit of work," reports 20.18% untagged for Houston and 30.87% untagged for FOPPA [Baseline]. The FOPPA figure independently matches the FOPPA maintainers' own documented finding of roughly 31% missing counterpart notices, external validation from a separate research team using different methods [Baseline].

This paper defines the metric, publishes the calibration method that catches the structural-artifact failure mode, and presents the municipal and procurement baselines that quantify the method's mechanics at scale. It is explicit about what those baselines can and cannot prove: they validate that the measurement instrument works, not what the attribution gap looks like inside a small business's books. SMB-specific findings are deferred until real SMB data lands, and every deferred claim in this paper is marked [TBD] rather than estimated. The goal of publishing now is to establish the vocabulary, so the attribution gap has one name and one calibration method before someone else coins a weaker version of the same idea.

1. Abstract

Silent margin misstatement in small and medium business accounting is not a bookkeeper failure. It is a structural property of how modern accounting software records accounts payable: a bill can be posted without an attribution to a specific customer, job, project, contract, or purchase order, and the resulting profit and loss aggregates still tie. When aggregation is silent, the reported margin on any given job is systematically overstated, proportional not to how many bills went untagged, but to the dollar weight of the untagged ones. This paper introduces the attribution gap as a formal, measurable, cross-domain metric: the fraction of AP records lacking a valid department-or-project attribution, reported alongside its dollar-volume share and its per-axis calibration [Baseline].

Measuring the attribution gap honestly requires more discipline than the concept suggests. A naive single-axis measurement produces misleading headlines in structurally opposite directions across real-world datasets: on the City of Houston Checkbook (2,293,699 records, USD 65.1B of AP over nine years), attributing solely by capital-project work-breakdown-structure code reports 92.11% untagged, because most municipal AP is operating cost attributed via purchase order or contract, not by work-breakdown structure [Baseline]. On the French Public Procurement Awards dataset (FOPPA, 1,380,965 records, EUR 1.75T of lots over 2010-2020), attributing solely by buyer reports 0.00% untagged, because FOPPA's schema does not record buyerless lots at all [Baseline]. The honest measurement, ANY-OF(WBS | Purchase Order | Contract) for Houston and awardPrice-present for FOPPA, produces 20.18% and 30.87% untagged respectively [Baseline]. The FOPPA result matches the maintainers' independently documented approximately 31% missing-counterpart figure, external validation that the detection method measures the phenomenon accurately rather than an artifact of how Accordia built it [Baseline].

This paper defines the metric, publishes the calibration method so others can compute it and cite it, and presents the municipal and procurement baselines. It states plainly what those baselines prove and do not prove. Municipal and national procurement books are not proxies for SMB bookkeeper behavior. SMB-specific results, from a Codat aggregate, an Accordia real-shape sandbox demonstration environment, and Accordia's first live customer, are deferred to a future revision when that data lands. The purpose of publishing the metric ahead of the SMB data is to establish the vocabulary before others do: the attribution gap is real, it is measurable, and it deserves a name that means the same thing across the SMB accounting, procurement, and management-accounting research communities.

2. Introduction and problem statement

Takeaway first. An accounting-SaaS dashboard can be arithmetically correct and still mislead an owner about job profitability, because the software has no requirement that every cost be attributed to a job before it is allowed to post. That gap between "the P&L balances" and "the job margin is honest" has not had a name or a standard measurement. This paper gives it one.

The observable phenomenon. An SMB owner reads job margin from QuickBooks Online, Xero, or Sage. The reported number is arithmetically correct on the bills that were attributed to that job. But bills that were never attributed to any job silently disappear from the per-job denominator, and the aggregate margin on the jobs that did get tagged looks better than it would if the untagged cost were counted somewhere. The owner sees profitability that has not actually been earned. This is not a software bug. It is a structural consequence of a design choice: mainstream accounting software allows a bill to post without a job, customer, or project reference, and the general ledger does not care.

Why the phenomenon has been under-measured historically. Three reasons compound. First, it does not surface on any standard financial statement; the profit-and-loss statement still balances whether or not individual costs are attributed to jobs, because attribution is a sub-ledger concern, not a general-ledger one. Second, the mainstream SMB software category does not expose the ratio. Its dashboards default to a single-number KPI for margin, revenue, and cost per job, and exposing the attribution gap honestly would require showing that number alongside the axes on which it could have varied, a presentation discipline named formally as the Confidence-Bounded Presentation Principle and cited later in this paper [Confidence Principle]. Third, the research literature that studies cost-attribution completeness is scattered across procurement, management accounting, and information-systems venues, each using its own vocabulary for a version of the same phenomenon.

Why now. Two large public procurement datasets, the City of Houston's checkbook and the French national procurement award register (FOPPA), make it possible to define and calibrate the attribution-gap metric on real invoice-level AP data at a scale no SMB dataset currently offers [Baseline]. The calibration method this paper develops on municipal and procurement data generalizes to any AP-shaped dataset, including an SMB QuickBooks book, even though the prevalence the method measures does not generalize automatically. Publishing the method now, ahead of SMB-specific numbers, lets the calibration surface be reviewed and challenged before it is load-bearing for a specific business claim.

What this paper contributes. First, a formal definition of the attribution gap that is instrument-independent, computable on any AP-shaped dataset by a third party using the published rule, not only by Accordia's own detector. Second, a per-axis calibration method that catches the structural-artifact failure mode a naive single-axis measurement produces. Third, municipal and procurement baselines that quantify the method's mechanics on 3.67 million combined real records [Baseline]. Fourth, a framework for SMB-specific measurement that is ready to receive data the moment Codat, a real-shape sandbox, or a first customer supplies it.

3. Methodology

Takeaway first. The attribution gap is not one number. It is a primary figure plus a calibration table that proves the primary figure was not chosen to flatter the result. Any published attribution-gap percentage that omits the calibration table is, by this paper's own standard, unfalsifiable and should not be trusted, including numbers this paper itself publishes.

3.1 Formal definition of the attribution gap

For a set of AP records R, the attribution gap at record-count basis is the fraction of records where the dataset's primary attribution field is null or empty after whitespace stripping: count(r in R where dept_or_project_id(r) is null) / count(R). The dept_or_project_id function is the primary attribution axis, the field or composite of fields that the dataset's originating institution treats as "which unit of work does this cost belong to." For Houston, that composite is ANY-OF(WBS ID | Purchase Order Number | Contract Number). For FOPPA, it is whether an award price is present on the lot. For a QuickBooks Online SMB book, it is whether the bill line's CustomerRef is populated [Baseline].

The attribution gap at dollar-volume basis uses the same numerator and denominator restricted to summed dollar amount rather than record count. The two bases must always be reported together, because they diverge. On Houston, 20.18% of records are untagged, but 55.66% of dollar volume is untagged [Baseline]. A record-count-only summary systematically understates the dollar exposure, because untagged bills skew toward the largest amounts, a pattern documented in the amount-bucket distribution in Section 5.1.

Per-axis calibration is the load-bearing discipline. For every candidate axis considered as the primary field, work-breakdown structure, purchase order, and contract for Houston, award-price presence, buyer, and supplier for FOPPA, the attribution gap is also computed with that single axis as the sole check, and reported as a sidecar table alongside the composite headline [Baseline]. A headline that a reader can see agrees across independently-chosen axes is a materially stronger signal that the detector's mechanics are honest than a headline that lives or dies on one field's choice.

3.2 Structural-artifact prevention

The load-bearing methodological finding from the municipal baseline is that a naive single-axis detector produces misleading headlines in structurally opposite directions depending on the dataset. Houston's work-breakdown-structure axis alone reports 92.11% untagged, because that field captures capital-project work only, and most municipal AP is operating expense attributed through a purchase order or contract instead [Baseline]. FOPPA's buyer axis alone reports 0.00% untagged, because FOPPA's schema populates a buyer identifier on every lot by construction and does not record buyerless lots at all [Baseline]. Neither result is a detector defect. Both are structural properties of the underlying data model that would silently mask a real measurement error if the report showed only one axis.

This is the same discipline already load-bearing in Accordia's product, formalized as the Confidence-Bounded Presentation Principle: a metric that shapes a decision must be presented alongside the axes on which its measurement could have varied [Confidence Principle]. The whitepaper and the product share the same refusal to publish a single unqualified scalar.

The methodological ceiling this section states plainly: without per-axis calibration, any published attribution-gap figure is unfalsifiable. A reader has no basis to judge whether a headline number reflects the honest primary axis or a single-axis structural artifact chosen, deliberately or not, to produce a more dramatic or more flattering result.

3.3 Cross-domain generalization to SMB accounting

The calibration method generalizes to any AP-shaped dataset by the same three-step discipline: define the primary attribution axis in the dataset's own semantics, enumerate sidecar axes that are structurally distinct from the primary, report all of them, and only then state a headline number.

For a QuickBooks Online SMB book specifically, Accordia's production detector today checks a single axis: whether the bill line's CustomerRef is populated [Capability Register C34, C35, status per Capability Register at time of reading]. That single-axis design mirrors the exact failure mode Section 3.2 documents for Houston's work-breakdown-structure axis: a QuickBooks book that attributes cost consistently through Projects (a QuickBooks sub-customer), Class, Department, or Location, but not through the top-level CustomerRef, would report as more untagged than it honestly is.

Accordia's response is a design specification for a second-generation detector that checks all five of QuickBooks's attribution-bearing fields independently, CustomerRef, SubcustomerRef (Projects), Class, Department, and Location, and reports a per-axis breakdown plus a workspace-configurable composite primary-axis verdict, by direct analogy to Houston's ANY-OF(WBS | PO | Contract) composite [Detector v2 Spec, Section 4.2, Section 6.3]. The design refuses by construction to emit a single-axis scalar without the four sibling axis blocks alongside it, the same defense-in-depth pattern the Confidence-Bounded Presentation Principle names as its strongest concrete instantiation to date [Detector v2 Spec, Section 4.1]. As of this writing, this second-generation detector is a design specification: the Capability Register lists it PLANNED, meaning a reader should not infer that Accordia's live product currently reports multi-axis QuickBooks attribution [Capability Register C46, PLANNED].

Full enumeration of the SMB sidecar-axis surface, and any prevalence claim about how QuickBooks books actually distribute across those five axes, is deferred to a future revision when real SMB data is available to walk the surface honestly, per Section 4.

3.4 Reproducibility

The detector-mechanics replica run against Houston and FOPPA is intentionally independent of Accordia's production QuickBooks detector, because the production detector operates on raw QuickBooks JSON payloads inside a platform service and is not importable against arbitrary tabular data [Baseline, Methodology section]. A third party can replicate the municipal and procurement baselines end to end from the public datasets, at the Houston Checkbook public data portal and the FOPPA Zenodo record, using the replica method Accordia will share on request [Baseline, Methodology section].

4. Data sources

Takeaway first. Two datasets are backed by real data today, at municipal and national-procurement scale. Three SMB-specific sources are named, scoped, and currently empty. This section exists so no reader mistakes a placeholder for a finding.

4.1 Inline data sources (backed by data now)

City of Houston Checkbook. Years 2018 through 2026, 2,293,699 records, USD 65.1B of AP volume. Public dataset landing page: data.houstontx.gov/dataset/checkbook. Primary attribution axis: ANY-OF(WBS ID | Purchase Order Number | Contract Number). Attribution-gap headline: 20.18% by record count, 55.66% by dollar volume [Baseline, Executive summary and Houston detail].

French Public Procurement Awards (FOPPA v1.1.3). Years 2010 through 2020, 1,380,965 lots, EUR 1.75T of award volume. Public dataset landing page: zenodo.org/records/10879932. Primary attribution axis: whether an award price is present on the lot. Attribution-gap headline: 30.87% by record count, 0.00% by dollar volume, because the untagged records are structurally the ones with no award amount recorded at all [Baseline, Executive summary and FOPPA detail]. Independently corroborated by the FOPPA maintainers' own technical report (reference hal-03796734), which documents an approximately 31% missing-counterpart-notice figure using their own methodology, cited here as external validation of the detector's mechanics, not as a claim that Accordia and the FOPPA maintainers used identical methods [Baseline].

Currency handling. Houston is reported in US dollars, FOPPA in euros. Amounts are reported per dataset in native currency and are never summed across datasets. No single 2010-2020 foreign-exchange rate is honest for that purpose; the euro-to-dollar rate ranged roughly from 1.05 to 1.40 across the FOPPA window [Baseline, Methodology section].

4.2 Deferred data sources (placeheld until available)

Codat aggregate SMB AP prevalence. [TBD, awaiting a reply to Accordia's outreach]. The intended contribution is anonymized SMB-shaped bill-level records spanning many small businesses, sourced from a payments-and-accounting-data aggregator, to establish an SMB-specific attribution-gap prevalence figure that municipal baselines cannot proxy. Outreach was sent 2026-07-30; status as of this writing is tracked in the outreach tracker [Outreach Tracker, Codat row].

Accordia DEMO-2 real-shape sandbox. [TBD, awaiting build]. A generated SMB-shape book designed to closely match production QuickBooks bill patterns, so the detector's calibration behavior can be measured in-domain without waiting for a live customer.

Accordia customer-1 real book. [TBD, awaiting Accordia's first paying customer]. Attribution-gap measurement on a live SMB book connected through the Intuit App Store. When available, this figure will be reported as a range with calibration context rather than a single scalar, per the Confidence-Bounded Presentation Principle, and with an explicit small-sample caveat until additional customers are onboarded [Confidence Principle].

What is not deferred. The method, the definition, and the municipal and procurement baselines are complete now. Only the SMB prevalence numbers are placeheld.

4.3 What each source can and cannot say

5. Results

Takeaway first. Both inline datasets show a materially non-trivial attribution gap once measured honestly, and both show the gap concentrated in the largest-dollar records rather than spread evenly. No cross-domain comparison table is published in this version, because a table with only two municipal rows filled in would misleadingly suggest this paper's central claim rests on procurement data rather than on SMB data that has not landed yet.

5.1 Houston Checkbook, inline results

Houston's composite primary axis reports 79.82% tagged and 20.18% untagged by record count, and 55.66% of dollar volume untagged [Baseline, Executive summary]. The largest-amount bucket, records above USD 1M, contains 9,978 of 2,293,699 records, 0.43% of the population, but carries 60.22% of that bucket's own records untagged and USD 28.57B of the workspace's total USD 36.24B in untagged dollar volume, 78.83% of all untagged dollars [Baseline, Houston detail bucket table; figure corroborated in Estimator Spec Section 2]. The pattern is direct: the rare untagged bills tend to be the largest ones, and they distort the aggregate disproportionately to their count. This is the exact concentration pattern Accordia's product is designed to catch, per the correction-cost framing in Section 7.

The per-axis calibration table shows why the composite, not any single field, is the honest primary axis:

Source for all rows: [Baseline, Houston detail, Multi-axis calibration table]. The work-breakdown-structure-alone row is the structural artifact discussed in Section 3.2. The department-alone row is a different structural artifact in the opposite direction, every record in Houston's checkbook carries a department code by construction, so a department-only measurement would report a perfect 0% untagged and hide the real gap entirely.

5.2 FOPPA, inline results

FOPPA's primary axis, award-price presence, reports 69.13% tagged and 30.87% untagged by record count [Baseline, Executive summary]. This figure independently matches the FOPPA maintainers' own documented approximately 31% missing-counterpart-notice finding, published in their technical report hal-03796734 [Baseline, FOPPA detail]. Untagged dollar volume is 0.00% by construction, because the untagged records are, definitionally, the ones lacking a recorded award amount [Baseline].

Source for all rows: [Baseline, FOPPA detail, Multi-axis calibration table]. The buyer-alone row is the structural artifact discussed in Section 3.2. A data-quality note travels with the FOPPA dollar figures: the raw award-price field contains sentinel placeholder values as large as 10^20 in a small number of rows, which the baseline's normalizer caps at EUR 1B for dollar-volume math only; the cap never changes which records are tagged or untagged [Baseline, Data-quality notes].

5.3 SMB attribution gap, Codat aggregate, deferred

[TBD, awaiting a reply to Accordia's outreach to Codat]. When populated, this section will report record-count and dollar-volume attribution gap on the aggregate, per-axis calibration on the QuickBooks bill-line axes named in Section 3.3, and a comparison to the Houston and FOPPA baselines.

5.4 SMB attribution gap, DEMO-2 real-shape sandbox, deferred

[TBD, awaiting Accordia's DEMO-2 build]. When populated, this section will report the sandbox's attribution-gap distribution as a calibration reference for what a typical SMB book looks like against the detector, with the caveat that a generated shape is not a real book.

5.5 SMB attribution gap, Accordia customer-1, deferred

[TBD, awaiting Accordia's first paying customer]. When populated, this section will report the customer's attribution-gap distribution per the Confidence-Bounded Presentation Principle, a range with calibration context, and will note explicitly that a single customer is not a population estimate.

5.6 Cross-domain comparison

Deferred until at least one SMB row from Sections 5.3 through 5.5 is populated. Publishing a comparison table with only the two municipal rows filled in would misleadingly suggest this paper's central claim rests on procurement data; it does not, and it will not until SMB evidence exists to compare against.

6. Caveats and limits of external validity

Municipal and procurement data are not SMB proxies. This is the most important caveat in this paper. The City of Houston Checkbook and FOPPA are, respectively, a large US municipality's expenditure ledger and a national procurement award registry. Their bookkeeping is subject to procurement discipline, public-records mandates, appropriations tracking, and audit regimes that no small business faces. An SMB owner running a small repair shop, a small construction firm, or a landscaping company does not operate under those constraints and does not staff a procurement office. Reading a Houston or FOPPA untagged percentage as a prediction of what an SMB book will look like is a category error. The municipal baselines quantify what the detection method reports on adversarially mostly-tagged books, in other words, whether the mechanics work at scale and whether they hallucinate on well-attributed AP. They do not answer what fraction of SMB bills are untagged.

Detector mechanics validated, prevalence not validated. This paper's inline results validate the measurement instrument, not the phenomenon's prevalence in the target population. A broken thermometer that reads 20% on a Houston book still reads 20% on an SMB book, and a reader could incorrectly conclude the SMB gap is 20% for that reason alone. Distinguishing "the instrument works" from "here is the population prevalence" is the entire point of Section 4's inline-versus-deferred split. A reader must not collapse the two.

Multi-axis primary-axis choice is a judgment, not a formula. Houston's primary axis is a composite chosen after a hand-trace revealed that the pure work-breakdown-structure run's approximately 92% untagged figure was a structural artifact of municipal accounting, not a bookkeeper failure [Baseline, Methodology section]. FOPPA's primary axis, award-price presence, was chosen to reflect the maintainers' documented gap rather than the structural artifact of the buyer field. Both choices are defensible and documented, and both could be challenged by a different research team using different primary-axis choices to produce different headline numbers. The per-axis sidecar tables exist precisely so a reviewer can inspect and re-choose. This paper's methodological contribution is the calibration discipline, not the specific primary-axis choices made here.

Sampling limits. Houston covers nine years of one large US municipality's AP; it is not a sample of US municipal AP generally. FOPPA covers eleven years of French national procurement; it is not a sample of European procurement generally. Neither is a random sample; both are complete populations of their scoped domains. Statistical inference from either to a broader population is not supported. This paper's claims are population-specific.

Currency non-summability. Houston is US dollars; FOPPA is euros; no single 2010-2020 exchange rate is honest. Dollar-volume figures are reported per dataset in native currency and are never summed across datasets. Any downstream citation that combines the two into a single headline dollar figure will be misusing this paper.

FOPPA amount cap. FOPPA's award-price column contains sentinel and placeholder values, some as large as 10^20. The normalizer caps amounts at EUR 1B for dollar-volume math only; this never affects the tagged-versus-untagged attribution figures themselves. Recorded here so a reproducing researcher lands on the same numbers [Baseline, Data-quality notes].

Detector-replica versus production-detector separation. The replica used against the municipal and procurement datasets is not Accordia's production QuickBooks detector; it is an independent replica of the same rule, running against normalized tabular data. The production detector operates on raw QuickBooks JSON payloads inside a platform service and is not importable against arbitrary tabular data. Any claim about the production detector's behavior on SMB books requires SMB data, which is deferred per Section 4.

What the caveats do not caveat away. They do not caveat away the detector-mechanics validation, that stands. They do not caveat away the FOPPA external-validation match to the maintainers' approximately 31% figure, that stands. They do not caveat away the calibration methodology, that stands. What they caveat is prevalence claims, cross-population generalization, and any collapse of "the instrument works" into "the population is X." This paper's central contribution is the metric and its calibration method, and nothing in this section weakens that contribution.

7. Implications

Takeaway first. The attribution gap matters differently to five audiences, and this paper makes no claim for any audience that outruns the evidence in Sections 4 through 6. Where a claim depends on data this paper does not yet have, it says so.

7.1 For SMB owners

The practical implication for an SMB owner cannot yet be stated as a number, because Sections 5.3 through 5.5 remain [TBD, awaiting SMB prevalence data]. What can be stated now is the shape of the problem and the shape of the fix. Any SMB owner reading job margin from QuickBooks, Xero, or Sage without a per-axis attribution check is reading a number whose upside error direction is known even though its magnitude is not yet published for SMB books specifically. The correction Accordia is building is not a bookkeeping process reform; it is a measurement discipline applied at read time, without changing how bills get posted.

Detection alone does not move an owner to act. Accordia's response is a design for converting a detection ("your AP has an attribution gap") into a quantified consequence ("this untagged bill represents an estimated dollar range of margin exposure on this specific job"). The design specifies a per-bill payload shaped as a lower bound, a point estimate, and an upper bound, traveling with a confidence label as one atomic unit, so no downstream screen can compose a bare, unqualified dollar figure from it [Estimator Spec, Section 4]. As of this writing this estimator is a design specification only; it is not yet in customer hands, and the Capability Register lists it as PLANNED [Capability Register C43, PLANNED]. When built, an illustrative (not measured) example of the shape it will produce is a range such as "18,000 to 27,000 dollars of unattributed margin exposure, driven by 6 large bills, with 12 more bills held for insufficient signal," never a single unqualified number [Estimator Spec, Section 3, marked illustrative in the source document].

7.2 For auditors and lenders

The attribution gap is a measurable, replicable metric that can be included in review-stage or lender-due-diligence checklists as a signal of how reliable a business's reported per-customer or per-job profitability actually is. It is distinct from, and complementary to, standard control-testing metrics. This paper makes an explicit non-claim: nothing here is evidence that any specific SMB book is unreliable. It is a measurement tool for auditors and lenders to apply to the books they review, with the caveats of Section 6 attached to every application.

7.3 For accounting-SaaS category incumbents

The category default of single-number margin dashboards is measurably unsafe when the underlying data carries structural ambiguity, as Sections 3.2 and 5 demonstrate on real municipal and procurement data. The Confidence-Bounded Presentation Principle offers a positive path forward: multi-axis measurement by default, with progressive disclosure so the primary view stays calm for an owner who does not need to see a breakdown [Confidence Principle, Section 5]. This paper's implication for incumbents is that the current default should be revisited, not that incumbents have been negligent; the single-number default made sense when the underlying aggregation was assumed clean. The municipal baseline evidence in this paper is now on record that the aggregation cannot be assumed clean.

7.4 For management-accounting and information-systems researchers

The attribution gap is offered as a common name for a phenomenon that has been described under scattered terminology across the procurement, management-accounting, and information-systems-controls literatures. Standardizing the name and the calibration method allows cross-study comparison that scattered terminology has prevented. This paper's methodology has not been peer-reviewed at time of writing. A research collaboration is being pursued: outreach was sent 2026-07-30 to accounting and finance faculty at Emory University's Goizueta Business School, and to accounting faculty at Georgia State University and Kennesaw State University, with the current state of that outreach documented in the outreach tracker [Outreach Tracker]. Any academic co-authorship on a future revision requires the collaborator's explicit agreement; nothing in this paper implies an endorsement derived from outreach or from being contacted.

7.5 For regulators

This is not a call for regulatory action. The metric exists, is computable on public datasets by any third party using the published method, and is available for reference should a procurement-transparency or SMB-lending regulator find it useful. Nothing more is claimed here.

8. What this whitepaper will not do

An explicit self-limitation list, so a reader knows what is deliberately left out and why, rather than discovering a gap and wondering whether it was missed.

• Will not claim an SMB attribution-gap prevalence figure. No SMB-shaped inline data exists at the time of this version. The Codat aggregate, DEMO-2, and customer-1 rows in Section 5 remain [TBD] until at least one of those sources is populated with real data.
• Will not sum dollar volumes across the Houston (USD) and FOPPA (EUR) baselines. No single 2010-2020 exchange rate is honest; native-currency reporting only, per dataset.
• Will not propose a threshold below which an SMB book is deemed reliable. Threshold-setting requires SMB-specific distributional evidence that is deferred; any threshold published now would be arbitrary.
• Will not claim customer wins, engagement metrics, revenue signal, or product-market fit. No such data exists at the time of this version. Any claim of this kind in a future revision requires standalone evidence.
• Will not claim that closing the attribution gap causally improves margin, pricing decisions, or business survival. This paper measures the gap. It does not claim an intervention effect; that requires studies outside this paper's scope.
• Will not re-derive management-accounting theory of cost attribution. This paper stands on the existing literature and does not re-argue foundational cost-attribution theory.
• Will not make an Accordia-specific product claim beyond what the Capability Register supports at the time of reading. Every capability referenced in Section 7.1 cites its current status; a claim that outruns the cited status is a defect in this document, not a feature of it.
• Will not be excerpted, cited, or reproduced without the caveats in Section 6. The caveats are load-bearing and travel with any reproduction, citation, or excerpt of the metric or the results.
• Will not publish co-authorship without a collaborator's explicit agreement. The byline is Accordia. Any academic co-authorship requires explicit agreement to co-author, not an implied endorsement derived from outreach or citation.
• Will not name a specific SMB business as a design partner in this version. Accordia is in design-partner discovery with owner-operated QuickBooks Online businesses that price work by job, project, repair, install, order, or contract; a specific business is not named here pending an explicit decision to do so in a future revision.

Appendix: methodology, data provenance, and limitations

A.1 Methodology summary

The attribution gap is measured by checking whether an AP record's primary attribution axis, a field or composite of fields chosen in the dataset's own semantics, is populated after whitespace stripping. Every primary-axis choice is accompanied by a sidecar table showing what each candidate axis would report alone, so a reader can see whether the headline number survives independent calibration or is a single-axis structural artifact. Full definition in Section 3.1; full structural-artifact discussion in Section 3.2. The replica method used to produce every Houston and FOPPA number in this paper is available from Accordia on request [Baseline, Methodology section].

A.2 Data provenance

Source: [Baseline, header and Methodology section]. Both figures were generated on the same run and are not stale relative to this paper's publication; a future revision that reuses these numbers without re-running the baseline should note the original generation timestamp above.

Houston per-year row counts (2018 through 2026): 255,781; 247,944; 233,939; 240,674; 235,142; 268,192; 267,983; 272,348; 271,696 [Baseline, Methodology section]. FOPPA sub-table row counts: 1,380,965 lots, 1,497,632 lot-buyer records, 1,371,535 lot-supplier records, 301,096 agent records [Baseline, Methodology section].

A.3 Limitations (pointer)

The full limitations discussion is Section 6 of this paper and is load-bearing; it is not repeated here in full to avoid drift between two copies of the same text. The single most important limitation, restated for a reader who jumps straight to the appendix: municipal and national procurement data validate that the measurement instrument works, they do not establish what the attribution gap looks like inside a small business's books, and no claim in this paper should be read as if it did.

Help extend the evidence

For a design partner. Accordia is looking for a small handful of owner-operated QuickBooks Online businesses that price work by job, project, repair, install, order, or contract, willing to share anonymized books under an appropriate NDA. In return: free margin visibility on your own book while the partnership runs, and first look at the product decisions your data helps shape. Every design partner materially shortens the honest gap between what this paper's municipal baseline proves and what Accordia's SMB findings will show. Contact Accordia through the contact page.

For academic researchers. If you study cost attribution, procurement completeness, or SMB financial reporting, and you have anonymized SMB-shaped accounting data, particularly QuickBooks bill-level records with customer or project attribution, Accordia would welcome a conversation about a research or NDA framework, or about co-authoring a future revision of this paper. Outreach to Emory Goizueta, Georgia State, and Kennesaw State accounting faculty is underway and documented in the outreach tracker [Outreach Tracker]. Contact Accordia through the contact page.

For Intuit and enterprise partners. Accordia is pursuing richer QuickBooks test-company access and anonymized aggregate attribution data through the Intuit App Partner Program to extend this paper's baseline into the QuickBooks Online population directly. Enterprise buyers and platform partners with a stake in AP attribution completeness are welcome to reach out through the contact page.

Version

Version 1, published 2026-07-31.