Methodology

In Brief

Every figure in this report is read twice from the source PDFs and pinned to one named survey edition. The resulting certified dataset is then held in three independently maintained formats (a CSV catalog, a spreadsheet, and a graph database) that must agree before a figure ships, and reconciling them has caught real errors before publication. The graph itself pairs each certified figure with a context layer that records why a value changed between editions, not merely that it did.

Every quantitative claim in this report is traced to a primary source and labeled with a verification tier. This page documents the method: how a figure gets certified, how the certified dataset is stored and cross-checked, and how it is organized into a graph. For the per-data-point registry and the publishers’ own restatements between survey editions, see Data Integrity & Corrections Log.

Methodology

The report separates three kinds of content and never lets them blur together. Reported data is reproduced directly from the KBCM/Sapphire Source Materials and attributed to the publishers. Derived figures are calculations the report performs from reported data, with the formula and inputs recorded so any result can be reconstructed. Interpretation covers the framing, causal reading, and forward-looking projections, which are the report’s editorial judgment rather than survey findings. Every figure carries a tier that says which of the three it is and how far it can be trusted.

The verification itself is a fixed procedure. Each figure is read from the source PDF twice, once as text and once from the rendered page image, to catch transcription and OCR error. Each is pinned to a single named survey edition, and values from different editions are never spliced into one series, because the publishers restate prior years from one edition to the next. Derived figures are recomputed from their certified inputs. Anything that cannot be located in or reconciled to the Source Materials is not published as certified; it is held back and labeled. The result is a claim of representational fidelity, that the report faithfully reflects what the publishers printed, and explicitly not a claim about whether the underlying survey data is itself correct. The Disclaimer sets out the full scope and limits of that verification.

Multi-Format Verification

The certified dataset is not stored once. Every figure that passes verification is written into three independently maintained representations built from the same page-verified values: a CSV catalog, a spreadsheet workbook, and a graph database, each queried separately rather than generated from one another at read time. That redundancy is a defense, not documentation for its own sake: a figure that only exists in one format has no independent check against transcription drift; a figure that agrees across three does.

The same discipline extends to the report itself: the site’s own rendering data is reconciled against the certified catalog rather than trusted on its own, and that comparison is what surfaces real defects before publication rather than after. The report’s rendering data is verified value-identical to the certified catalog it draws from, and any future edit that moves the two out of alignment is a defect by definition, not a judgment call. Verified

Certified-catalog storage architecture and rendering-data reconciliation verified against the project’s own reconciliation recordVerified

Knowledge Graph & Context Graph

A further step, layered on top of the certified catalog rather than replacing it, ingests each year’s extracted and verified results into a knowledge graph paired with a context graph. The knowledge graph holds four layers: a deterministic subgraph of every certified figure with its segment cuts, source citations, and cross-edition restatements; a causal layer of roughly 440 cause-and-effect relationships extracted from the research corpus and independently red-teamed for evidentiary faithfulness; a cross-source triangulation layer marking where a KBCM figure agrees or diverges from external benchmarks such as High Alpha, Bessemer, and OpenView; and an entity-linkage layer tying KBCM metrics into SuccessCOACHING’s existing customer success management (CSMBOK) knowledge graph. Verified

The context graph sits beside it and answers a different question: not what a figure is, but why it moved. Each survey edition carries its own record of respondent count, fielding window, and any known sampling changes. Every metric restated between editions now carries a classified reason (a name-mapping change, a shift presumed from panel recomposition, a stated change in methodology, or a correction to a prior print) instead of an unexplained delta. And every year-over-year series carries an explicit continuity flag, so a trend chart cannot silently join two editions whose underlying respondent pools are not comparable. This is the added step that turned the report’s own restatement analysis, covered next, from an observed pattern into a structure that can be queried and checked rather than only narrated.

Knowledge graph layer counts and context graph structure verified against the project’s production knowledge graph documentation and semantic-layer implementation recordVerified

Querying the Graph

Structure alone does not make a graph useful; it also has to be askable. A benchmark semantic layer sits underneath the knowledge graph for exactly that reason: a governed metric registry holding one canonical definition per metric, an alias crosswalk mapping every raw spelling a survey or a query uses (“Net Dollar Retention,” “NDR,” “net dollar retention (ndr)”) to that single canonical definition without ever rewriting the source-faithful raw values underneath it, and two resolution functions that encode the vintage-citation, display-value, and forecast rules as enforced database code rather than as prose a query author has to remember correctly. The registry is derived and enforcing; it never carries authority over the certified catalog itself, which remains the source-faithful arbiter. Verified

The resolver functions this feeds are what turn a restatement from a flag into an explanation. Asked to resolve a metric-year that was restated, the graph does not return a generic discrepancy warning; it returns the specific reasoning (which edition changed the value, by how much, and the classified cause) in the same call that returns the number. Asked for a full series, it returns the observation sequence with every gap in continuity flagged, and it will not silently bridge across a break in respondent-pool comparability unless the caller explicitly acknowledges the break first. That is the mechanism, not just the intent, behind the earlier claim that a restated figure can be checked and not only narrated: the graph is built to fail honestly on a bad question rather than answer it plausibly.

Semantic layer and resolver-function behavior verified against the project’s benchmark semantic layer specification and production knowledge graph utilization guideVerified

Key takeaways

Every figure in this report is read twice from the source PDFs and pinned to one named survey edition, because the surveys restate their own history.
Certification claims representational fidelity to what the publishers printed, not that the underlying survey data is itself correct; the two are deliberately kept separate.
The certified catalog is held in three independently maintained formats that must agree; cross-checking the report’s own data against it has caught real defects before publication, not after.
Certified figures are also ingested into a production knowledge graph with four layers: certified benchmarks, causal relationships, cross-source triangulation, and entity linkage into the broader customer success graph.
A paired context graph now gives every restated figure a classified reason and every multi-year series a continuity flag, so a chart cannot silently splice incompatible survey cohorts.
A governed semantic layer resolves every metric name and alias to one canonical definition, and its resolver functions return the reasoning behind a restated figure, not just the flag that it changed.

Frequently asked questions

How was the data in The Retention Reckoning verified?

Each figure is read from the source KBCM survey PDF twice, once as text and once from the rendered page image, then pinned to a single named edition. Derived figures are recomputed from certified inputs, and anything that cannot be reconciled to the source is held back.

Does the report verify that KeyBanc's survey numbers are correct?

No. The report certifies representational fidelity, that every figure faithfully reflects what the publishers printed. It makes no claim about whether the underlying survey data is itself correct, only that the report reproduces and derives from it accurately.

Why is the certified data stored in a knowledge graph as well as a spreadsheet?

Because a flat table forces every figure to have one final value, and KBCM restates its own numbers from edition to edition. The graph keeps every version tied to its edition and links each figure to why it changed, what other sources say about it, and what the research suggests drives it, none of which a spreadsheet alone can represent.