diff --git a/CONTENT_BACKLOG.md b/CONTENT_BACKLOG.md
index 72f774d..a5c8e31 100644
--- a/CONTENT_BACKLOG.md
+++ b/CONTENT_BACKLOG.md
@@ -20,7 +20,7 @@ The Module column uses `1A` (definition quality) and `1B` (data integrity) for t
| M1-05 | 1B | Agreement measurement for analytics QA | pattern | v0.2.0 | drafted | Twenty-nine teams on one dataset returned odds ratios from 0.89 to 2.93; ISO 5725's repeatability and reproducibility as the lens; LinkedIn found 60 percent of its questions had more than one right answer | 1 session |
| M1-06 | 1B | Data quality SLOs and error budgets | pattern | v0.2.0 | candidate | [TODO: prior art] | 1 session |
| M1-07 | 1B | Schema drift detection and data contracts | pattern | v0.1.0 | drafted | Avro resolution as the mechanism behind the quiet change; GoCardless's six-month report on what a producer-side programme costs; the author's addition of the assistant's memory as a drift of its own | 1 session |
-| M1-08 | 1B | The reconciliation protocol | pattern | v0.2.0 | candidate | [TODO: prior art] | 1 session |
+| M1-08 | 1B | The reconciliation protocol | pattern | v0.2.0 | drafted | The bank reconciliation's reconciling-items schedule as the lineage, from the 2025 Green Book and a state auditor's guide; Airbnb, Uber and Zoom as the two-numbers stories; the author's three causes, exclusions, calendars and as-of time | 1 session |
| M1-09 | 1B | Measuring trust, not just accuracy | pattern | v0.2.0 | candidate | [TODO: prior art] | 1 session |
| M1-10 | 1B | Data incident root cause analysis | pattern | v0.2.0 | candidate | [TODO: prior art] | 1 session |
| M1-11 | 1A | State-based retention measurement | pattern | v0.2.0 | drafted | Duolingo growth model (Gustafson, Mazal) | 1 session |
diff --git a/DECISIONS.md b/DECISIONS.md
index 5572e0c..144d968 100644
--- a/DECISIONS.md
+++ b/DECISIONS.md
@@ -2,6 +2,14 @@
One entry per judgment call: decision, alternatives, why.
+## 2026-09-18 — M1-08 argues with escalation to the senior person, imports the reconciling-items schedule as the ladder, and defaults the reference to the blessed computation
+
+**Decision:** five calls made while drafting the reconciliation pattern. (a) The Position argues with the practice of handing a disagreement to the most senior person in the room, or to whoever has time, and moving on. It is taken at its strongest: the senior person may still choose the reference, and the page only forbids choosing before the conditions are listed and the ladder closes. The critique is sourced from peer-reviewed group-decision work (motivated numeracy, confidence matching, the hierarchy meta-analysis, the many-teams finance study) and from the experimentation literature's highest-paid-person's-opinion framing, and Sources & Stories says the transfer from laboratory tasks to a metrics meeting is the framework's own. (b) The difference ladder is the reconciling-items schedule from a bank reconciliation, imported from the 2025 Green Book and a state auditor's guide and said so; the rule that it closes to zero or the number is on hold is the accounting rule carried over unchanged. The lead on reconciliation testing in data migration produced no peer-reviewed source inside the ten-year window, so the audit sources carry that lineage and the page says so in Sources & Stories rather than in the body. (c) The reference defaults to the blessed computation named by M1-01, corrected if the ladder shows it holds the defect, and the independent total keeps its job as the M1-03 control total rather than becoming the reference; where no blessed computation exists, the reconciliation is where one gets named. (d) Zoom's 2020 participants-versus-users correction is used as the definition-word example beside Airbnb and Uber, cited from the company's own progress report plus the TechCrunch report that reproduces its statement, because it is the cleanest public case of one figure travelling under two words in one week. (e) The HiPPO sentence and Twyman's law are quoted from the experimentation book's preface, read from the publisher-hosted sample chapter, which upgrades that entry from chapter-title level; the revision is noted for REFERENCES.md.
+
+**Alternatives:** argue with averaging or with taking the friendlier number, which no competent practitioner defends in print, so the Position would contest a strawman; make the independent total the reference, which would put the reference outside the warehouse and contradict M1-01; cite a vendor's reconciliation feature as the lineage, which would name a product and misplace the origin; note the missing migration-testing source in the body with a caveat, which the reader-not-literature-review rule forbids.
+
+**Why:** a Position should contest the practice the reader actually follows, and at ten people a disagreement is settled by whoever is senior or free. Importing the accounting schedule whole gives the reader a procedure their accountant already recognizes. Defaulting to the blessed computation keeps the three integrity pages consistent about which number stands and which number checks it.
+
## 2026-09-16 — M1-05 argues with the single-owner key its sibling page installed, scores values by raw agreement and verdicts by kappa, and refuses to score disputed questions
**Decision:** four calls made while drafting the agreement-measurement pattern. (a) The Position argues with the single owner writing the reference answers, the practice M1-04 itself installs. The page keeps that owner and treats the owner's answers as a gauge with an unmeasured error: a second person answers the same twenty questions blind before the assistant is scored, the human-to-human rate is the ceiling the assistant is held to, and an assistant inside that rate's band gets no model work. (b) Values within tolerance are scored by raw percent agreement, because two numbers landing within a tenth of a percent by chance is not a real risk; kappa is reserved for pass-or-fail verdicts and is reported beside the raw rate, never alone. The fixed kappa bands the quality manuals grade against are not printed, because no source inside the ten-year window could be verified for them, and the question is left for the author. (c) The band is plus and minus the square root of p(1−p)/n, about two questions at twenty, and the floor of twelve agreements in twenty below which the questions rather than the people are the problem is likewise the framework's own; both are marked as such on the page and in the template. (d) A question the two people disagree on is not scored for or against the assistant. It routes to one line in the context file and a dated new row in the key, so the assistant is scored only on settled questions, and its misses there are the only errors that are its own. (e) Three calls from the author's review, folded in on 2026-09-17: the quality manuals' kappa bands are not used, because a percentage difference can be a delay, an exclusion, a calendar, or two different metrics landing near each other, and no band separates those; the logic is checked before any statistic, through a three-bullet conditions summary under every answer (what was excluded, which calendar, when the data was pulled) that an assistant can write and a person can verify; and the per-metric tolerance is set by a person as the reporting difference the operation can live with, before the study. Nobody is the better judge; the study is run for alignment, and its working signal is that important calls spend their time on actions rather than on whose number is right.
diff --git a/REFERENCES.md b/REFERENCES.md
index 885ae0b..76e2e1b 100644
--- a/REFERENCES.md
+++ b/REFERENCES.md
@@ -19,6 +19,7 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[AVRO-SPEC]** — The Apache Software Foundation. "Apache Avro Specification," version 1.12.0. avro.apache.org, undated page, fetched September 11, 2026. The key carries no year because the page is versioned rather than dated; the same rules were spot-checked against the 1.11.1 spec and the wording is unchanged. Cited for the Schema Resolution section, which makes the structural-versus-semantic asymmetry mechanical: fields are "matched by name" rather than position; "if the writer's record contains a field with a name not present in the reader's record, the writer's value for that field is ignored"; "if the reader's record schema has a field that contains a default value, and writer's schema does not have a field with the same name, then the reader should use the default value from its field"; and "if the reader's record schema has a field with no default value, and writer's schema does not have a field with the same name, an error is signalled". Enum resolution behaves the same way: an unknown writer symbol falls back to the reader's default if one exists, and otherwise errors.
- **[BALFOUR-2013]** — Balfour, Brian. "The Never Ending Road To Product Market Fit." brianbalfour.com, December 11, 2013. The flattening-curve reading of product-market fit, with his own caveat that flattening only shows fit for some market and that the next job is segmenting the curve to identify who retained. Distinct from [BALFOUR-2018].
- **[BALFOUR-2018]** — Balfour, Brian; Winters, Casey; Kwok, Kevin; Chen, Andrew. "Growth Loops are the New Funnels." Reforge, 2018. Existence verified via the authors' own announcements; the page blocked direct fetching at verification time (2026-08-04), so cite the thesis, and [TODO: verify] any direct quote against the live page.
+- **[BANG-2017]** — Bang, Dan; Aitchison, Laurence; Moran, Rani; Herce Castanon, Santiago; Rafiee, Banafsheh; Mahmoodi, Ali; Lau, Jennifer Y. F.; Latham, Peter E.; Bahrami, Bahador; Summerfield, Christopher. "Confidence matching in group decision-making." _Nature Human Behaviour_ 1, article 0117, May 26, 2017. doi:10.1038/s41562-017-0117. Read in full from the author-hosted copy; the journal record verified through Crossref because the publisher redirects automated fetching. Six behavioural experiments in the United Kingdom and Iran, with computational modelling. Verified verbatim: "When disagreement arises, opinions expressed with higher confidence tend to carry more weight"; pairs "match their communicated confidence so that certainty and uncertainty is stated in approximately equal measure by each party"; the strategy "is effective when group members have similar levels of expertise" and "is, however, sub-optimal and can cause miscommunication about who is more likely to be correct. This herding behaviour is one reason why groups can fail to make good decisions." The tasks are perceptual judgements by pairs; carrying the finding to the two owners of a metric is the framework's extension.
- **[BAVARESCO-2025]** — Bavaresco, Anna; Bernardi, Raffaella; Bertolazzi, Leonardo; Elliott, Desmond; et al. "LLMs instead of Human Judges? A Large Scale Empirical Study across 20 NLP Evaluation Tasks." ACL 2025, main conference; arXiv:2406.18403, submitted June 26, 2024, read at abstract level (v3, June 2, 2025). The JUDGE-BENCH study: "20 NLP datasets with human annotations" and "11 current LLMs, covering both open-weight and proprietary models". Verified: "Models are reliable evaluators on some tasks, but overall display substantial variability depending on the property being evaluated, the expertise level of the human judges, and whether the language is human or model-generated", and "LLMs should be carefully validated against human judgments before being used as evaluators". Cited for the validation rule; the per-dataset correlations are in the unread PDF. Twenty authors in all.
- **[BEREZOVSKY-2022]** — Berezovsky, Olga. "How to measure cohort retention." Lenny's Newsletter, August 30, 2022. N-day against unbounded retention, and the cohort-table layout; stresses that the two give completely different retention data. Possible partial paywall.
- **[BEREZOVSKY-2024]** — Berezovsky, Olga. "Measuring Non-Cohorted Retention or Blended Churn." Data Analysis Journal #188, February 21, 2024. Framing citable; article body is paywalled — [TODO: verify] specifics against a subscription copy before quoting.
@@ -73,6 +74,7 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[FISHKIN-2022]** — Fishkin, Rand. "Provable Marketing Attribution is a Boondoggle; Trust Your Gut Instead." SparkToro blog, April 4, 2022.
- **[FISHKIN-2024]** — Fishkin, Rand. "Attribution is Dying. Clicks are Dying. Marketing is Going Back to the 20th Century." SparkToro blog, July 29, 2024.
- **[FORSGREN-2018]** — Forsgren, Nicole; Humble, Jez; Kim, Gene. _Accelerate: The Science of Lean Software and DevOps: Building and Scaling High Performing Technology Organizations_. IT Revolution Press, March 27, 2018. ISBN 9781942788331. Verified at publisher-page level and cited at book level; check any benchmark number against the current [DORA-2026] page rather than the book, since the thresholds move between report years.
+- **[FOWLER-2021]** — Fowler, Martin. "Bitemporal History." martinfowler.com, April 7, 2021. Read in full. An essay distinguishing the actual history of a value from the record of that history, worked through a salary example. Verified verbatim: "There is a difference between the actual history, and our record of the history"; "I'm using the terms _actual_ and _record_ history for the two dimensions"; the second dimension "comes in as we ask what did we think Sally's salary history was on February 25?"; and "Retroactive changes are a problem when actions are based on a past state that's retroactively changed". Cited for the two-times distinction behind recording an as-of time with every number; the essay is about database design, and the framework applies it to a metric pulled before a meeting.
- **[FREAKONOMICS-2020]** — Dubner, Stephen (host). "Does Advertising Actually Work? (Part 2: Digital)." Freakonomics Radio Ep. 441, November 2020. Full transcript on the page; carries Steven Tadelis's first-person eBay account and the P&G anecdote (Tim Hwang).
- **[FREDDIEMAC-2020]** — Freddie Mac. "Mortgage Closing Cycle Time Benchmark Study, Edition One," December 2020. Existence and title confirmed; the PDF could not be parsed at verification time, so every figure in it is unverified. Do not cite numbers from it until it has been re-fetched and read.
- **[FREEAGENT-2020]** — Brown, Rob. "Can't Someone Else Do It? Automating Looker Housekeeping in Two Days." FreeAgent engineering blog, October 2, 2020. First-party account of one company's dashboard cleanup: content unviewed for 90 days was deleted, removing roughly 25 percent of dashboards and 33 percent of saved views. A measurement of one company's BI instance, not a general rate; cited as the only first-party number found near the untraceable 60-80 percent unused-dashboards claim.
@@ -81,11 +83,14 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[GALLO-2014]** — Gallo, Amy. "The Value of Keeping the Right Customers." hbr.org, October 29, 2014. The modern citation anchor for the five-to-25-times acquisition-versus-retention claim; the sentence's own hedge is "Depending on which study you believe, and what industry you're in", and no study is named. Also relays the 25-95 percent retention-profit range, crediting Reichheld at Bain. Cite the sentence together with its hedge, never the multiplier as a finding.
- **[GANZ-2026]** — Ganz, Jason; Perigaud, Benoit (both dbt Labs). "Semantic Layer vs. Text-to-SQL: 2026 Benchmark Update." dbt Developer Blog, April 7, 2026. Eleven questions run 20 times each over an insurance schema originally built by a different vendor's researchers. Verified figures: claude-sonnet-4-6 90.0% text-to-SQL against 98.2% semantic layer; gpt-5.3-codex 84.1% against 100.0%; text-to-SQL on the full set 32.7% (2023) to 64.5% (2026); the two 2026 models near-identical on both paths. Vendor conflict direct and disclosed: the authors benchmark their employer's own semantic layer. Their stated caveats matter: the whole schema was loaded as context, which they say is impractical for larger datasets; questions needing too many entity hops score 0.0% through the semantic layer, whose full-set figures are 72.7% for both 2026 models against 60.5% in 2023; and three added data models improved both approaches. Cite for direction with the out-of-scope band attached, never the headline percentages alone.
- **[GAO-2024]** — U.S. Government Accountability Office. _Federal Information System Controls Audit Manual (FISCAM)_, GAO-24-107026, 2024 revision. Text extracted and searched on September 6, 2026. Verified, control BP.01.01.02 ("Control totals are employed when practicable"): "A control total is the sum of a numerical field contained in a set of records. Control totals are used to verify the completeness of a set of records as it is processed. Control totals are verified by comparing those from a processed set of records (output) to those of the same set of records before processing (input)." Also, under system-interface reconciliation: "Control totals are agreed between the source and target systems." Cited as the audit lineage of the control-total check the framework imports; the manual is written for federal financial audits and says nothing about analytics. Supersedes GAO-09-232G.
+- **[GAO-2025]** — U.S. Government Accountability Office. _Standards for Internal Control in the Federal Government_ (the Green Book), GAO-25-107721, May 2025. Supersedes GAO-14-704G of September 2014; effective from fiscal year 2026, early adoption permitted. Text extracted from the published PDF and searched on September 18, 2026. Verified verbatim, Appendix II, "Reconciliations" (page 105): "Management performs reconciliations to confirm that transactions are processed, recorded, and accounted for completely and accurately. Reconciliations include identifying and comparing transactions from two sets of records to determine whether the transactions are recorded properly, have yet to be recorded, or were recorded improperly and require correction. Reconciliations also serve to identify unauthorized transactions and explain differences. Reconciliations may be performed by comparing internal records or by verifying the entity's records against external data, such as with a bank reconciliation." Paragraph 10.17 lists reconciliations among transaction control activities, paragraph 16.05 among ongoing monitoring activities, and paragraph OV4.14 names "checking supervisor reconciliations" as a response open to a smaller entity that cannot fully segregate duties. Written for federal entities; the transfer to a metric is the framework's.
- **[GARTNER-2021]** — Sakpal, Manasi. "How to Improve Your Data Quality." Smarter with Gartner, July 14, 2021. States "Every year, poor data quality costs organizations an average $12.9 million" with no methodology, sample, or measurement year; the research behind the figure is a paid Gartner product and is not public. The live page sat behind a bot check at verification (2026-08-26) and was read via an Internet Archive capture. Cite the sentence's existence and its missing methodology, never the figure as a measurement.
- **[GEORGE-2002]** — George, Michael L. _Lean Six Sigma: Combining Six Sigma Quality with Lean Production Speed_. McGraw-Hill, 2002. ISBN 9780071501903. Bibliography verified only; the Little's Law and process-cycle-efficiency formulations attributed to it come from secondary summaries and are unverified. Cite at thesis level, and check against a copy before quoting.
- **[GOODHART-1975]** — Goodhart, Charles. "Problems of Monetary Management: The UK Experience." In _Papers in Monetary Economics_, Reserve Bank of Australia, 1975. Citation verified via independent bibliographic sources; original not freely available. Quote only the 1975 sentence, and credit the popular "measure becomes a target" phrasing to [STRATHERN-1997].
- **[GORDON-2019]** — Gordon, Brett R.; Zettelmeyer, Florian; Bhargava, Neha; Chapsky, Dan. "A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook." _Marketing Science_ 38(2), 2019. Verified at abstract/DOI level; full text paywalled. doi:10.1287/mksc.2018.1135.
- **[GOULD-2025]** — Gould, Elliot; Fraser, Hannah S.; Parker, Timothy H.; Nakagawa, Shinichi; et al. "Same data, different analysts: variation in effect sizes due to analytical decisions in ecology and evolutionary biology." _BMC Biology_ 23, article 35, February 6, 2025. doi:10.1186/s12915-024-02101-x. Abstract verified through an indexed copy because the publisher gates automated fetching; the journal record verified through Crossref and Europe PMC. Verified: "174 analyst teams, comprising 246 analysts", two unpublished datasets, 141 and 85 usable effects; "substantial heterogeneity among results for both datasets"; for the first dataset "near continuous variation in effect size from large negative effects to effects near zero, and even effects crossing the traditional threshold of statistical significance in the opposite direction"; and "no strong relationship" between peer-review ratings of the analyses, variable choice or random-effects structure and deviation from the meta-analytic mean. Cited for the peer-rating finding. The index record carries the 2020 registered-report date; the published article is 2025.
+- **[GRANDHI-2024]** — Grandhi, Sai Sameera; Agarwal, Abhinav; Singh, Onkar; Mankala, Nagesh Kumar (Uber). "Streamlining Financial Precision: Uber's Advanced Settlement Accounting System." Uber Engineering Blog, October 24, 2024. First-party account of an internal payments system Uber does not sell, read in full. Verified verbatim: "Settlement accounting is a critical financial process that involves reconciling and recording all transactions related to settlements, refunds, and other financial activities conducted through PSPs"; "Payment reconciliation, in essence, is the process of cross-referencing and verifying financial transactions between the payment service providers and the bank records"; matching uses "a deterministic field called transactionReference shared between them"; "For events where reconciliation fails, the system marks them as unhealthy and submits them to the DLQ (Dead Letter Queue)"; and the scale figure "about 1.2 billion settlements each month". Cited for the shape of a standing reconciliation, never for the platform; no failure rates are published.
+- **[GREER-2018]** — Greer, Lindred L.; de Jong, Bart A.; Schouten, Maartje E.; Dannals, Jennifer E. "Why and when hierarchy impacts team effectiveness: A meta-analytic integration." _Journal of Applied Psychology_ 103(6), pp. 591–613, June 2018. doi:10.1037/apl0000291. Abstract verified through Europe PMC and Semantic Scholar because the publisher and PubMed refuse automated fetching; the journal record verified through Crossref. Verified verbatim: the meta-analysis draws "on results from 54 prior studies (N = 13,914 teams)"; "on net, hierarchy negatively impacts team effectiveness (performance: ρ = −.08; viability: ρ = −.11), and that this effect is mediated by increased conflict-enabling states"; and the negative relationship "is exacerbated by aspects of the team structure (i.e., membership instability, skill differentiation)". A 2019 correction addressed a coding error in one included study. Cited for direction and mechanism; the effect is small, and the page says so.
- **[GUSTAFSON-2023]** — Gustafson, Erin. "Meaningful metrics: How data sharpened the focus of product teams." Duolingo Blog, February 17, 2023.
- **[HANDY-2017]** — Handy, Tristan. "The Startup Founder's Guide to Analytics." 2017; first published on thinkgrowth.org (Medium), read at the dbt Labs republication, which is dated March 30, 2017. A consultant's stage model by employee count, recommending the first full-time analytics lead at 20 to 50 employees, stated from work with "over a dozen companies"; no survey or dataset behind it, and the author does not claim one. The author later ran dbt Labs, an analytics tooling vendor. The widely repeated hire-at-20-to-50 threshold traces here and no further.
- **[HE-2025]** — He, Horace; with others at Thinking Machines Lab. "Defeating Nondeterminism in LLM Inference." Thinking Machines Lab blog, September 10, 2025. Verified: 1,000 completions from one open model (Qwen3-235B-A22B-Instruct-2507) at temperature 0 produced "80 unique completions, with the most common of these occuring 78 times", identical through the 102nd token and first diverging at the 103rd. The stated primary cause is that inference kernels are not batch-invariant, so an answer depends on the server load the request landed in, on top of floating-point non-associativity. The fix, batch-invariant kernels released as thinking-machines-lab/batch-invariant-ops, requires owning the inference stack, which a team on a hosted API does not. The author's employer is an AI lab; cited for the measurement.
@@ -103,6 +108,7 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[JIN-CHEN-2018]** — Jin, Li; Chen, Andrew. "The Power User Curve: The best way to understand your most engaged users." a16z / andrewchen.com, August 2018 (on-page day not independently confirmed; year solid). DAU/MAU shortcomings and the L30 histogram. (also )
- **[JONES-2021]** — Jones, Andrew (GoCardless). "Improving Data Quality with Data Contracts." Tech @ GoCardless on Medium, December 24, 2021. First-party account by the engineer who introduced the practice at the company. Read in full. Carries the diagnosis: critical data streamed out of service databases by an in-house change-data-capture service forces every consumer onto an internal data model that "has not been designed for consumption", "lacks documentation", and "will change as the service evolves, causing unexpected breaking changes for the data consumers"; his verdict is "Ultimately, we believe this data is of poor quality." Load-bearing for the practice this framework argues with: catalogs and anomaly detection "do nothing to improve the quality of your data. To do that, you need to go to the source of the data and how it is generated." Direct fetching is blocked (HTTP 403), so verification was against the rendered page. The author's own site carries a stub dated 28 December 2021; the Medium original's date is the one recorded here.
- **[JONES-2022]** — Jones, Andrew (GoCardless). "Data Contracts at GoCardless — 6 Months On." Tech @ GoCardless on Medium, May 25, 2022. The follow-up to [JONES-2021], read in full, and the most useful published source on what a producer-side contract programme actually costs. Adoption split verified verbatim: data contracts were by then "powering around 50% of these events" for asynchronous inter-service communication, while "We haven't yet seen the same kind of adoption for data primarily produced for the data teams", with most usage still driven by the legacy change-data-capture pipelines; until the planned decommissioning deadline approaches, "it can be difficult to make the case for teams to do so when compared against their other priorities". Also records that the publishing friction is intentional ("this friction is a necessary part of creating an interface to your data") and that the outbox pattern "does add some load to your database" and can "easily ... overload a database", "leading to outages and incidents". Direct fetching is blocked (HTTP 403); verification was against the rendered page.
+- **[KAHAN-2017]** — Kahan, Dan M.; Peters, Ellen; Dawson, Erica Cantrell; Slovic, Paul. "Motivated numeracy and enlightened self-government." _Behavioural Public Policy_ 1(1), pp. 54–86, 2017. doi:10.1017/bpp.2016.2. Abstract read in full from the publisher's product page. Verified verbatim: subjects highest in numeracy "did substantially better than less numerate ones when the data were presented as results from a study of a new skin rash treatment"; when the same data were presented as a gun-control study, "polarization did not abate among subjects highest in numeracy; instead, it _increased_"; and the identity-protective account "predicted that more numerate subjects would use their quantitative-reasoning capacity selectively to conform their interpretation of the data to the result most consistent with their political outlooks." A U.S. survey experiment on a politically charged question; carrying it to a metric that fits a team's story is the framework's extension.
- **[KANBAN-2025]** — _The Kanban Guide_, v2025.5, May 1, 2025. Stewarded by John Coleman, Daniel Vacanti, Colleen Johnson, Prateek Singh, Julia Wester, Christian Neverdal, Magdalena Firlit, Tom Gilb, and Steve Tendon. Free, dated, primary definitions of work in progress, throughput, work item age, and cycle time. Verified negative: the guide does not mention Little's Law.
- **[KAPLAN-1958]** — Kaplan, E. L.; Meier, Paul. "Nonparametric Estimation from Incomplete Observations." _Journal of the American Statistical Association_ 53(282), 1958, pp. 457–481. doi:10.1080/01621459.1958.10501452. Bibliography verified across independent listings; the publisher page returned 403 and a hosted PDF could not be parsed. Cite as provenance for the product-limit estimator; do not quote it.
- **[KAUSHIK-2006]** — Kaushik, Avinash. "Excellent Analytics Tip #5: Conversion Rate Basics & Best Practices." Occam's Razor, 2006.
@@ -113,12 +119,13 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[KELLOGG-2024]** — Kellogg, Dave. "Why Great Marketers Look at Pipeline Coverage, Not Just Pipeline Generation." Kellblog, September 17, 2024.
- **[KENT-2021]** — Kent, Brian Patrick. "Conversion rate modeling: worth the effort?" crosstab.io, January 9, 2021.
- **[KERR-1975]** — Kerr, Steven. "On the Folly of Rewarding A, While Hoping for B." _Academy of Management Journal_ 18(4), 1975. Citation verified via independent bibliographic sources; full text paywalled. doi:10.2307/255378.
+- **[KHISTI-2018]** — Khisti, Ninad; Betz, William (Airbnb). "Measuring transactional integrity in Airbnb's distributed payment ecosystem." The Airbnb Tech Blog on Medium, January 17, 2018. First-party engineering account, read in full from the rendered page because direct fetching is blocked (HTTP 403). Verified verbatim: "Typically, merchants reconcile all the money movement on the platform via 3-way reconciliation between platform transaction records, processor settlement files, and bank statements"; transactional integrity "represents a no-surprise, accurate money movement"; the pipeline compares "each and every transaction from day zero", whereas "reconciliation often leaves out already matched transactions"; and the work "led to large improvements in payment reconciliation, leading to more accurate financial reporting". Cited for the three-way shape and the trace-every-transaction rule; no figure for the integrity rate itself is published.
- **[KHUNE-2024]** — Khune, Abhi; Busch, Callie; Johnson, Jeffrey; Chakka, Pradeep. "QueryGPT – Natural Language to SQL Using Generative AI." Uber Engineering Blog, September 19, 2024. First-party account of an internal natural-language-to-SQL service: naive similarity search over schema and SQL samples produced "declining accuracy in the generated queries" as more tables were onboarded, and was replaced by curated per-domain workspaces plus an intent-classification step. Cite as an existence proof about retrieval degradation, not as a template for small teams. Verified negative: the widely circulated hour- and dollar-savings figures attributed to this system do not appear in the post; do not attribute them to it.
- **[KNIGHT-LEVESON-1986]** — Knight, John C.; Leveson, Nancy G. "An Experimental Evaluation of the Assumption of Independence in Multiversion Programming." _IEEE Transactions on Software Engineering_ 12(1), 1986, pp. 96–109. Cited at abstract level, read in a course-hosted copy. Verified from the abstract: 27 versions of a program written independently from one specification at two universities and run against one million tests; "the programs were individually extremely reliable but that the number of tests in which more than one program failed was substantially more than expected" under independence. Cited for the caveat that two computations share mistakes more often than chance predicts, so a second path must start from different material; it is a software-reliability experiment on student programs, not an analytics study, and it predates the framework's recency preference by design, because it is the canonical result.
- **[KOHAVI-2007]** — Kohavi, Ron; Henne, Randal M.; Sommerfield, Dan. "Practical Guide to Controlled Experiments on the Web: Listen to Your Customers not to the HiPPO." KDD 2007, author's copy.
- **[KOHAVI-2010]** — Kohavi, Ron; Longbotham, Roger. "Unexpected Results in Online Controlled Experiments." SIGKDD Explorations 12(2), 2010.
- **[KOHAVI-2017]** — Kohavi, Ron. "Twyman's Law and Controlled Experiments." Slide deck, written November 5, 2013, updated May 7, 2017, with a companion LinkedIn article of the same date. Verified: the law as "Any figure that looks interesting or different is usually wrong"; the earliest scholarly reference Kohavi found is A. S. C. Ehrenberg, "The Teaching of Statistics," _Journal of the Royal Statistical Society_ Series A 138(4), 1975, which quotes the law by name; the attribution is to W. A. (Tony) Twyman, a British radio and television audience-measurement researcher who, per Kohavi, never published it himself; a 1996 quotations dictionary lists the origin as unknown. The deck's examples include default-value artifacts (a birth date of 11/11/11; a departure city of Adak Island from an alphabetical default), a daylight-saving hour of zero traffic, and a newspaper stock table with eleven identical zero returns, corrected three days later. Cite for the law and its provenance; the examples are anecdotes from the author's experience. and
-- **[KOHAVI-2020]** — Kohavi, Ron; Tang, Diane; Xu, Ya. _Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing_. Cambridge University Press, 2020. Cited at book level; check page-level claims against a copy. Chapter 3 is titled "Twyman's Law and Experimentation Trustworthiness" and Chapter 21 "Sample Ratio Mismatch and Other Trust-Related Guardrail Metrics," both verified from the Cambridge University Press chapter listings on September 6, 2026; the pattern cites the book at chapter-title level only, and no page-level claim is made.
+- **[KOHAVI-2020]** — Kohavi, Ron; Tang, Diane; Xu, Ya. _Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing_. Cambridge University Press, 2020. Cited at book level; check page-level claims against a copy. Chapter 3 is titled "Twyman's Law and Experimentation Trustworthiness" and Chapter 21 "Sample Ratio Mismatch and Other Trust-Related Guardrail Metrics," both verified from the Cambridge University Press chapter listings on September 6, 2026. The Preface, "How to Read This Book", was read on September 18, 2026 from the publisher-hosted sample chapter (). Two sentences from it are verified verbatim: the authors' aim of "a data-driven culture that informs rather than relies on the HiPPO (Highest Paid Person's Opinion) (R. Kohavi, HiPPO FAQ 2019)", and "We believe in the skepticism implied by Twyman's Law: Any figure that looks interesting or different is usually wrong; we encourage readers to double-check results and run validity tests, especially for breakthrough positive results." No page-level claim is made beyond the Preface.
- **[KRAMER-2023]** — Kramer, Emily. "How to measure what marketing activities are actually driving revenue." MKT1 Newsletter, September 20, 2023. Partially paywalled — cite free-section claims only.
- **[LEI-2025]** — Lei, Fangyu; Chen, Jixuan; Ye, Yuxiao; Cao, Ruisheng; et al. "Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows." ICLR 2025. Read from the benchmark's project site and leaderboard, fetched September 6, 2026; the paper itself was not read. Verified from the site: 632 workflow problems over databases "often containing over 1,000 columns"; GPT-4o at "10.1% on Spider 2.0 tasks, compared to 86.6% on Spider 1.0", and o1-preview at 17.1 percent. On the fetch date the Snow split (547 examples) was topped by self-submitted commercial agents at 96.70, 96.53 and 94.15 percent with no linked papers, the Lite split (547 examples) at 76.23 percent, and the DBT split (68 examples) at 65.6 percent. A leaderboard entry is a vendor's own submission of a purpose-built agent with undisclosed per-database tuning; cite the launch figures as the paper's and the leaderboard rows as self-reported, never as measured capability. The leaderboard moves, so any citation carries the access date.
- **[LEI-LEXICON]** — Lean Enterprise Institute. Lean Lexicon terms: cycle time and takt time. The key deviates from the AUTHOR-YYYY convention because the pages carry no publication year and no edition marker. The lexicon's lead-time term URL returned 404; the lead-time ordering quoted in the metrics library appears on the cycle-time page itself. and
@@ -141,6 +148,7 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[MEHTA-2025]** — Mehta, Ravi. "Confidence engineering: Why your onboarding is probably too short." Ravi on Product, December 18, 2025. Free. First-party numbers for a checkout expanded from 3 steps to 25 steps, plus the decision-consequence and information-gap frame. His exact role at the company described is not stated on the page, so describe him as a product leader who worked on it.
- **[MENKVELD-2024]** — Menkveld, Albert J.; Dreber, Anna; Holzmeister, Felix; et al. "Nonstandard Errors." _The Journal of Finance_ 79(3), pp. 2339–2390, June 2024 (online April 17, 2024). doi:10.1111/jofi.13337. Abstract verified through an indexed copy because the publisher refuses automated fetching; the journal record verified through Crossref. Verified: nonstandard errors are defined as the uncertainty that variation in the evidence-generating process across researchers adds to a result, studied "by letting 164 teams test the same hypotheses on the same data"; "NSEs turn out to be sizable, but smaller for more reproducible or higher rated research. Adding peer-review stages reduces NSEs. We further find that this type of uncertainty is underestimated by participants." Cited for the underestimation finding.
- **[META-2024]** — Meta Platforms, Inc. 2023 Form 10-K, filed February 2, 2024 (accession 0001326801-24-000012). Daily active people (DAP) definition, the ML-estimate caveat for cross-app people counting, the duplicate/false-account discussion, and the announced retirement of per-app DAU, MAU, ARPU, and MAP from periodic reports.
+- **[MILLER-2020]** — Miller, Ron. "Zoom retracts statement that it has 300M daily active users." TechCrunch, April 30, 2020. Secondary source, cited because it reproduces the company's statement on the correction, which the company's blog carried only as a silent edit first noticed by The Verge. Verified verbatim, quoting Zoom: "this was first announced in our April 22 webinar as 300 million daily participants by our CEO Eric Yuan"; "When we realized this error on April 23, we corrected the wording to 'participants'"; and TechCrunch's own distinction that daily active users "is a standard industry metric measuring the number of unique users per day. Participants, on the other hand, can participate in multiple meetings per day." Pairs with [ZOOM-2020], the first-party report of the same week.
- **[MODIG-2012]** — Modig, Niklas; Åhlström, Pär. _This is Lean: Resolving the Efficiency Paradox_. Rheologica Publishing, 2012. ISBN 9789198039306. Both authors at the Stockholm School of Economics. Bibliography verified; the flow-versus-resource-efficiency content is unverified at page level, so cite the thesis rather than passages. No flow-efficiency benchmark number outside vendor blogs survived verification.
- **[MOHAMMED-2008]** — Mohammed, M. A.; Worthington, P.; Woodall, W. H. "Plotting basic control charts: tutorial notes for healthcare practitioners." _Quality and Safety in Health Care_ 17(2), 2008, pp. 137–145. doi:10.1136/qshc.2004.012047. Read in full from a hosted copy. Peer-reviewed source for the XmR arithmetic: limits at the mean ± 2.66 × mean moving range, with 2.66 = 3/1.128 and the range limit at 3.267 × mean moving range; the roughly 1-in-370 false-alarm rate of three-sigma limits under ideal conditions; the guidance that 20–25 points firm the limits up while fewer give "soft" provisional ones worth using; the three widely recommended run rules; and the warnings that day-of-week patterns and seasonality need preliminary work before charting. The examples are healthcare, and every formula is general.
- **[MONTECARLO-2024]** — MacDonald, Lindsay. "Data Quality Anomaly Detection: Everything You Need to Know." Monte Carlo blog, updated June 20, 2024. Vendor material from a data-observability company; the institutional key flags the conflict. Describes automatically applied freshness, volume, and schema monitors "without the need for any threshold setting," with baselines learned from each table's history. The post does not discuss false positives or alert fatigue, which is the cost side of what it sells; cite it for what the category does, never for effectiveness. The URL redirected from montecarlodata.com at fetch time.
@@ -171,6 +179,7 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[RYDBERG-2020]** — Rydberg, Johan. "Spotify's New Experimentation Platform (Part 1)." Spotify Engineering, October 29, 2020. Cited for the Metrics Catalog: before it, experiment analyses ran manually in notebooks and there was no consistency across the company in how experiments were analyzed; the catalog made metric pipelines managed and served their results to UIs and notebooks from one place. Verified at article level. The consistency claim is about experiment analysis specifically, not company reporting; do not stretch it further.
- **[SANDERSON-2024]** — Sanderson, Chad. "Implementing Data Contracts In The Data Warehouse." Monte Carlo blog, page states updated January 25, 2024. Double-flag: the author led the data-contract programme at Convoy, so it is a practitioner account, and it is hosted by a data-observability vendor with a commercial stake in warehouse-side monitoring, with product links and a closing call to action. Load-bearing concession, verified verbatim: "starting in the data warehouse doesn't require the help of application engineers or all data producers", and "Overall, it requires a smaller culture shift in order to succeed." His own recommendation remains both: "The best way to ensure data products can be effectively produced and maintained is to implement data contracts upstream at the service level as well as in the data warehouse." Cite the concession and the recommendation together; quoting the first without the second misrepresents him.
- **[SANDERSON-KREUZIGER-2022]** — Sanderson, Chad; Kreuziger, Adrian (both Convoy). "An Engineer's Guide to Data Contracts - Pt. 1." Data Products newsletter on Substack, October 5, 2022. The byline carries both names; Sanderson introduces Kreuziger, a principal engineer at Convoy, as the author of the body, so the key names both. First-party implementation account, part one of three. Load-bearing verbatim: the stated requirement "Data contracts must be enforced at the producer level", with the sentence that follows it, "If there's nothing enforcing a contract on the producer side, you don't have a contract. That's a handshake agreement at best"; that enforcement "must be automated as part of the existing CI/CD pipeline"; that a meaning change is a breaking change, since "changing the underlying meaning of the data being produced should break the data contract", illustrated by an entity whose length and width switch "from being stored in inches to centimeters"; and the admission that "There are simply some aspects of semantic enforcement that cannot be reliably managed prior to deployment."
+- **[SAO-2020]** — Office of the Washington State Auditor, Center for Government Innovation. _Best Practices for Bank Reconciliations_. January 2020. Read in full from the published PDF. Guidance for local governments "large or small". Verified verbatim: "A bank reconciliation is a control activity that compares banking records to accounting records, and ultimately ensures that both sets of records agree with one another"; "Perform reconciliations at least monthly"; "Document reconciliations. Documentation of reconciliations should include source documents used in the reconciliation and a schedule showing the comparisons made and adjustments or reconciling items identified"; "Ensure that your policy designates a reasonable period, such as two weeks, to research and resolve discrepancies"; "Investigate each reconciling item to the source"; "If an unknown item cannot be identified, isolate and observe it during your next reconciliation"; "Independently review reconciliations"; and "A supervisory review should always be documented by signature and date". The guide disclaims prescriptive force; the framework carries its working rules to a metric.
- **[SAPARINA-LAPATA-2024]** — Saparina, Irina; Lapata, Mirella. "AMBROSIA: A Benchmark for Parsing Ambiguous Questions into Database Queries." NeurIPS 2024, Datasets and Benchmarks track, spotlight; arXiv:2406.19073, submitted June 27, 2024, read at abstract level (v2, October 31, 2024). Verified: the benchmark covers "scope ambiguity, attachment ambiguity, and vagueness", and "even the most advanced models struggle to identify and interpret ambiguity in questions". Cited for the finding that a model does not notice a question with two readings.
- **[SCHELTER-2018]** — Schelter, Sebastian; Lange, Dustin; Schmidt, Philipp; Celikel, Meltem; Biessmann, Felix (Amazon Research); Grafberger, Andreas (University of Augsburg). "Automating Large-Scale Data Quality Verification." _PVLDB_ 11(12), 2018, pp. 1781-1794. doi:10.14778/3229863.3229867. Peer-reviewed and first-party; read from the VLDB PDF. Cited for the check vocabulary the framework's semantic half borrows: the vision that users "should be able to write 'unit-tests' for data ... analogous to established testing practices in software engineering"; Completeness as "the fraction of non-missing values in a column"; and "Consistency is defined as the degree to which a set of semantic rules are violated", where intra-relation constraints may span columns, the paper's own example being "if category is t-shirts, then the range of size is {S, M, L}". Also the syntactic-versus-semantic accuracy split, where for a product attribute colour "a value blue can be considered syntactically accurate in the given domain even if the correct value would be red". Constraints compile to aggregation queries over a distributed engine, which is why the same checks are cheap in a warehouse.
- **[SCHULTZ-2014]** — Schultz, Alex. "Lecture 6: Growth." How to Start a Startup (Y Combinator / Stanford CS183B), 2014.
@@ -200,6 +209,7 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[TAYLOR-2025]** — Taylor, Sam. "Xmrit Launches Trend and Seasonality Features for XmR Charts." xmrit.com (Commoncog's free charting tool), April 24, 2025. A feature announcement, so treat tool claims as self-description; the statistical content stands alone: significant seasonality makes an XmR chart "either miss exceptional variation, or wrongly detect seasonal variation as exceptional" by widening the limits, and the published fixes are deseasonalizing by seasonal factors or charting the period-over-period change.
- **[TENBENSEL-2019]** — Tenbensel, Tim; Jones, Peter; Chalmers, Linda M.; Ameratunga, Shanthi; Carswell, Peter. "Gaming New Zealand's Emergency Department Target: How and Why Did It Vary Over Time and Between Organisations?" _International Journal of Health Policy and Management_ 9(4), 2020, pp. 152–162. doi:10.15171/ijhpm.2019.98. Open access, fully verified. Note the year mismatch: the journal issue is 2020 and the DOI slug is 2019; the key follows the DOI.
- **[TENE-2015]** — Tene, Gil. "How NOT to Measure Latency." QCon San Francisco 2015; InfoQ recording posted March 26, 2016. Speaker is CTO and co-founder of Azul Systems. Talk title, speaker, venue, date, and abstract verified; the arguments about averages, percentile-of-percentiles, and coordinated omission are in the video and unverified at source, so cite at talk level. Latency is a different clock from cycle time; use it for the statistical point only.
+- **[TREASURY-TFM]** — U.S. Department of the Treasury, Bureau of the Fiscal Service. _Treasury Financial Manual_, Volume I, Part 2, Chapter 5100, "Fund Balance with Treasury Accounts." Undated page, fetched September 18, 2026; the key carries no year because the chapter shows none. Verified verbatim: "Treasury requires that agencies reconcile their FBwT accounts on a regular and recurring basis (at a minimum monthly)"; "Unresolved differences in these reports compromise the overall integrity and status of the government's financial position. Therefore, agencies that had remaining differences at conversion to a full CARS reporter, must continue to work to clear those differences"; and the rule that suspense-account balances include no "items or transactions more than 60 Business Days old". Federal cash accounting, cited only for the monthly cadence and the duty to clear differences; re-check the page before quoting again.
- **[TURNBULL-2015]** — Turnbull, Alex. "How We Reduced Churn by Over 71%." Groove blog, 2015, formerly at groovehq.com/blog/churn. The primary behind the widely repeated 71-percent churn story. At verification (2026-08-26) the old address redirects to the blog's new host and returns a page not found, and the Wayback Machine serves no capture of the post, so the primary is unretrievable; the figures survive only in third-party retellings, which state no baseline churn rate, period, or churn definition. Cite the disappearance and the claim's status as an untraceable first-party anecdote, never its figures. No working link exists.
- **[TWITTER-2017]** — Twitter, Inc. Q3 2017 Shareholder Letter, October 26, 2017. Exhibit 99.1 to Form 8-K (accession 0001564590-17-020095). "Note About Our MAU Adjustment": Digits SMS-authentication traffic counted as MAU since Q4 2014, roughly 1-2 million users per quarter; data-retention limits on reconciling earlier periods.
- **[TWITTER-2019]** — Twitter, Inc. "Twitter Announces Fourth Quarter and Fiscal Year 2018 Results," press release, February 7, 2019. Exhibit 99.2 to Form 8-K (accession 0001564590-19-002218). The mDAU definition, the stated rationale and non-comparability admission, MAU's announced retirement, and the transition quarter with both numbers.
@@ -225,3 +235,4 @@ _The full Phase 0 seeding is pending; entries below were verified individually a
- **[ZHAO-2022]** — Zhao, Xinshu; Feng, Guangchao Charles; Ao, Song Harris; Liu, Piper Liping. "Interrater reliability estimators tested against true interrater reliabilities." _BMC Medical Research Methodology_ 22, article 232, 2022. Read from the PubMed Central copy. A simulation study that tests chance-corrected indices against known true reliabilities. Verified: "Almost all agree that percent agreement (ao), the oldest and the simplest index, is also the most flawed because it fails to estimate and remove chance agreement"; in the tested conditions the kappa family (Scott's pi, Cohen's kappa and Krippendorff's alpha) "were crude predictors of reliability and evenness, and 31-point under-approximators of reliability"; and Gwet's AC1 was "the most accurate approximator". Cited for the under-estimation of kappa under skew. The framework's use of raw agreement for numeric values within a tolerance, where chance agreement is negligible, is its own reading and not this paper's.
- **[ZHENG-2023]** — Zheng, Lianmin; Chiang, Wei-Lin; Sheng, Ying; Zhuang, Siyuan; et al. "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena." NeurIPS 2023, Datasets and Benchmarks track; arXiv:2306.05685, submitted June 9, 2023, read at abstract level (v4, December 24, 2023). Verified: "strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80% agreement, the same level of agreement between humans", and the paper names "position, verbosity, and self-enhancement biases". Cited for the human-to-human agreement level as the ceiling a model judge is measured against, and for the biases; the setting is chat-assistant responses rather than numeric answers. Thirteen authors in all.
- **[ZHONG-2020]** — Zhong, Ruiqi; Yu, Tao; Klein, Dan. "Semantic Evaluation for Text-to-SQL with Distilled Test Suites." EMNLP 2020; arXiv:2010.02840, submitted October 6, 2020, read at abstract level. Verified: the Spider benchmark's exact-set-match metric "leads to a 2.5% false negative rate on average and 8.1% in the worst case", meaning correct queries marked wrong; the proposed alternative executes queries over a distilled suite of databases to bound semantic accuracy. The rates belong to exact-set match on academic benchmarks, measured with 21 leaderboard models and manually verified examples, not to execution comparison on a company warehouse.
+- **[ZOOM-2020B]** — Zoom Video Communications. "90-Day Security Plan Progress Report: April 22." Company PDF, April 22, 2020, subtitled "Zoom Surpasses 300M Daily Users, Announces Zoom 5.0 with AES 256-Bit GCM Encryption." Read in full. A first-party document carrying one figure under three descriptions: the subtitle's "300M Daily Users"; the body's "surpassing 300 million daily Zoom meeting participants"; "more than 300 million people around the world are using Zoom"; and "even with more than 300 million daily users." Cited for the wording alone, as the public case of one number travelling under two definitions in the same week; the correction is documented in [MILLER-2020].
diff --git a/docs/modules/01-ai-data-quality/README.md b/docs/modules/01-ai-data-quality/README.md
index 5d042d3..7d072fc 100644
--- a/docs/modules/01-ai-data-quality/README.md
+++ b/docs/modules/01-ai-data-quality/README.md
@@ -45,7 +45,7 @@ Whether the number the pipeline returns is the number the definition asks for. R
| M1-07 | [Schema drift detection and data contracts](schema-drift-and-data-contracts.md) | v0.1.0 |
| M1-05 | [Agreement measurement for analytics QA](agreement-measurement.md) | v0.2.0 |
| M1-06 | Data quality SLOs and error budgets | v0.2.0 |
-| M1-08 | The reconciliation protocol | v0.2.0 |
+| M1-08 | [The reconciliation protocol](reconciliation-protocol.md) | v0.2.0 |
| M1-09 | Measuring trust, not just accuracy | v0.2.0 |
| M1-10 | Data incident root cause analysis | v0.2.0 |
diff --git a/docs/modules/01-ai-data-quality/reconciliation-protocol.md b/docs/modules/01-ai-data-quality/reconciliation-protocol.md
new file mode 100644
index 0000000..cb259d4
--- /dev/null
+++ b/docs/modules/01-ai-data-quality/reconciliation-protocol.md
@@ -0,0 +1,179 @@
+---
+id: M1-08
+title: The reconciliation protocol
+module: 1B
+type: pattern
+status: drafted # full sourced draft; awaiting author voice pass per AGENTS.md constraint 6
+sources:
+ - BANG-2017
+ - CHANG-2021
+ - COLDEWEY-2016
+ - DICKEY-2017
+ - FOWLER-2021
+ - GAO-2024
+ - GAO-2025
+ - GRANDHI-2024
+ - GREER-2018
+ - KAHAN-2017
+ - KHISTI-2018
+ - KOHAVI-2020
+ - MILLER-2020
+ - SAO-2020
+ - TREASURY-TFM
+ - TWITTER-2017
+ - TWITTER-2022
+ - YU-2021
+ - ZOOM-2020B
+template: ../../../templates/01-ai-data-quality/reconciliation-record.md
+summary: >
+ When two numbers for the same thing disagree, write both down with the
+ time each was pulled, list the conditions behind each, bring both to the
+ finest grain they share, and build a ladder from one number to the other
+ with one line per cause. The ladder ends at zero or the number is on
+ hold. Which number stands is decided in writing before the meeting ends,
+ and the settled condition becomes a dated line in the definitions and
+ the context file, so the assistant stops reproducing the disagreement.
+keywords:
+ - reconciliation
+ - reconciling items
+ - difference ladder
+ - as-of time
+ - bitemporal
+ - control total
+ - reference number
+ - restatement
+ - fiscal calendar
+ - metric disagreement
+---
+
+# The reconciliation protocol
+
+## Problem
+
+Two numbers for the same thing are on the table and they do not match. The warehouse says one revenue figure and the billing system says another. From that moment the meeting is about whose number to believe, and the decision it was called for waits.
+
+At Airbnb the chief executive could ask which city had the most bookings last week and get diverging answers from data science and finance, built from slightly different tables and definitions.[^chang-2021] At Uber one rider metric read 6.53 million in one tool and 6.20 million in another for the same city and window, because of a stale filter in one tool's private query.[^yu-2021] Zoom's blog said the company had passed 300 million daily users, changed the word to participants the next day, and its progress report for that week carried the one figure under three different words.[^zoom-2020b] [^miller-2020]
+
+Most of the executive escalations I have watched begin this way. Two departments are moving fast, and nobody has time to say how their number was made. When I put the two side by side afterwards, the cause is nearly always one of three things. One side left out records the other side counted. The two sides ran on different calendars, a fiscal week against a calendar week or a fiscal year against a calendar year. Or one side pulled its number hours before the meeting and the live source had moved by the time the meeting started. The library names three more elsewhere: a join that changed the grain, two meanings of one word such as customer, and a unit or currency converted on one side only. Every one of these is a condition somebody left unstated, and arithmetic never finds a condition.
+
+
+
+## When this applies
+
+Reach for this page when the disagreement already exists. The problems that bring teams here look like these:
+
+- Finance and the warehouse give two figures for last month's revenue, and the board deck is due on Thursday.
+- This quarter's deck does not match the version of the same number that went out last quarter.
+- The assistant's answer and the analyst's query are a few percent apart, and both look right.
+- Two dashboards show two totals for one metric, and each owner is sure of theirs.
+- The last meeting ended with an agreement to look into it, and this is the third time for this number.
+
+It needs the two numbers, the queries or exports behind them, and one hour with both owners in the room. No analyst, no purchase.
+
+[Detecting plausible-but-wrong outputs](plausible-but-wrong.md) checks one number against something outside its query before it leaves the room. [Agreement measurement](agreement-measurement.md) measures how often two people agree on twenty questions. [Single point of metric computation](single-point-of-metric-computation.md) names the reference, and [schema drift and data contracts](schema-drift-and-data-contracts.md) explains an input that changed shape or meaning. Whether a settled number is a restatement that has to be disclosed belongs to [Module 4](../04-governance-and-financial-reporting/README.md). The deeper fix once the number is settled is data incident root cause analysis.
+
+Whenever another page in this library has two numbers already in the room and disagreeing, this is the page it points to.
+
+## The pattern
+
+Two numbers that disagree are reconciled in a fixed order, and the order is the pattern. Write both numbers down, each with the time it was pulled. Under each, list the conditions that produced it: what was excluded, which calendar, the as-of time, the grain, the unit and currency, and the definition of the word being counted. Bring both to the finest grain they share and compare them there, because the segment that carries the gap is where the cause lives. Then build the difference ladder. Start at one number, apply one condition at a time, write the amount each explains, and arrive at the other number. The ladder ends at zero. A remainder that is not zero is named as unexplained and the number is on hold. Before the meeting ends, one person writes which number stands and why, with name and date. The settled condition becomes a dated line in the metric definition and in the [context file](../03-ai-agent-integration/context-file.md), so the assistant stops reproducing the disagreement.
+
+```mermaid
+flowchart LR
+ TWO["Two numbers,
each with its as-of time"] --> COND["Conditions listed
under each"]
+ COND --> GRAIN["Compared at the
finest common grain"]
+ GRAIN --> LADDER{"Ladder closes
to zero?"}
+ LADDER -->|yes| REF["Reference named,
dated line written"]
+ LADDER -->|no| HOLD["Number on hold,
remainder named"]
+```
+
+Accounting has run this procedure for a long time, and I import it whole. A reconciliation compares transactions from two sets of records, sorts them into recorded properly, not yet recorded, and recorded improperly, and explains the differences.[^gao-2025] A state auditor's guide for small governments adds the working rules: reconcile at least monthly, investigate each reconciling item to its source, and allow a fixed period, such as two weeks, to resolve a discrepancy.[^sao-2020] Payment systems run the same shape at scale, Airbnb across its own records, the processors' settlement files and bank statements,[^khisti-2018] and Uber across about 1.2 billion settlements a month, with every failed match queued to be resolved.[^grandhi-2024] The ladder on this page is the reconciling-items schedule from a bank reconciliation, applied to a metric.
+
+## Position
+
+When two numbers disagree, reconcile the conditions before the values and settle in writing which number is the reference before the meeting ends, rather than averaging them, taking the friendlier one, or agreeing to look into it.
+
+The practice this argues with is competent and nearly universal. The disagreement goes to the most senior person in the room, who picks, or to whoever has time to dig, and the meeting moves on. I see why it appeals. It is fast, and the meeting has other business. The experimentation literature calls this the highest paid person's opinion, and built its case on informing that opinion rather than relying on it.[^kohavi-2020]
+
+My reading of the evidence is that a room under time pressure picks badly. Numerate people use the skill to land on the figure that fits what they already believe, and the most numerate subjects were the most polarized.[^kahan-2017] In pairs the more confident opinion carries more weight, and people match each other's confidence even when their expertise differs.[^bang-2017] Across fifty-four studies of nearly fourteen thousand teams, hierarchy's net effect on performance was negative, small, and ran through conflict.[^greer-2018]
+
+I would still let the senior person choose the reference, after the conditions are on the table. A choice made before then is a choice between two unstated definitions, and it comes back next quarter under a new name.
+
+Three things follow. A ladder that ends at zero is the only proof that the cause is known. The reference is the blessed computation from [single point of metric computation](single-point-of-metric-computation.md), and where none exists, the reconciliation names one. The settled condition is a dated line in the definitions and the context file, so the assistant stops reproducing the disagreement.
+
+## Implementation
+
+The [reconciliation record](../../../templates/01-ai-data-quality/reconciliation-record.md) carries the two numbers with their as-of times, the conditions table, the ladder, the reference decision, a machine-readable core, and the agent prompts. One disagreement takes an hour with both owners present.
+
+1. **Freeze both numbers with the time each was pulled.** The record of a value at 9:00 and the record of it at 11:30 are two facts, and systems that must answer what they believed at a past time keep both.[^fowler-2021]
+2. **List the conditions under each number.** Excluded records, calendar, as-of time, grain, unit and currency, the definition of the word counted, and the join path. An assistant can draft the list from its own query.
+3. **Compare at the finest common grain.** Bring both sides to the same grain and compute the difference by segment. A gap spread evenly across segments usually means a definition or a unit. A gap in one week points at the calendar, and one in a handful of rows at the join or an exclusion.
+4. **Build the ladder.** One line per cause. Start at the first number, apply the condition, write the amount, and carry the running total. The last line is the second number and the remainder is zero. When it is not, the remainder is named as unexplained, and nobody signs.
+5. **Decide the reference in writing.** Which number stands, why, who decided, and the date. A blessed computation stands by default, unless the ladder shows it holds the defect. Then the corrected version stands. Where none exists, one gets named now, and the losing copy gets a retirement date or a drift check.
+6. **Write the dated line.** The settled condition goes into the metric definition and into the [context file](../03-ai-agent-integration/context-file.md) the assistant reads, as a new dated line and never an edit in place.
+7. **Keep the record.** File it where the next person can find it. A second disagreement of the same pair for the same cause is evidence that the fix did not hold, and it goes to root cause analysis. How long a number may sit on hold is a question for data quality SLOs.
+
+The rows below are illustrative: revenue for a closed month, the warehouse's blessed view against the billing system's statement.
+
+| Step | Cause | Amount | Running total |
+|---|---|---|---|
+| Start | Warehouse view, pulled 09:00 | | 412,380 |
+| 1 | Refunds issued in the month; billing nets them and the view did not | −5,940 | 406,440 |
+| 2 | Calendar; the view used a fiscal period that ran two days into the next month | −7,180 | 399,260 |
+| 3 | Join to payments fanned out; 41 orders counted twice | −3,690 | 395,570 |
+| 4 | As-of; three payments captured between 09:00 and 11:30, in billing only | +580 | 396,150 |
+| End | Billing statement, pulled 11:30 | | 396,150 |
+| | Remainder | | 0 |
+
+The gap of 16,230 closes in four lines. Three of the four are defects in the view: the definition says revenue is net of refunds on the calendar month, and the join added rows, so the view is corrected and stands as the reference. The fourth is timing, recorded with its size. The billing statement keeps its job as the control total that [detecting plausible-but-wrong outputs](plausible-but-wrong.md) checks the view against. Stopped after line 2 with 3,110 open, the ladder would have named the two largest causes and proved nothing.
+
+## How you know it is working
+
+- Every disagreement that reaches a meeting leaves it with a record: two numbers, two as-of times, a ladder, a name and a date.
+- The remainder is zero more often than it is unexplained.
+- The same pair of numbers does not disagree twice for the same cause.
+- The assistant cites the dated line on the question that used to split the room.
+- Whose number is right has stopped being an item in the leadership meeting.
+- **Anti-signal:** every reconciliation ends with timing as the cause and no ladder. Timing is a cause with a size, and one that is never measured is a shrug.
+
+## Failure modes
+
+- **Averaging the two.** The midpoint satisfies neither set of conditions, and nobody can reproduce it.
+- **Taking the friendlier number.** The figure that fits the story wins, and the numerate people in the room are the ones best able to justify it.[^kahan-2017]
+- **Escalating before the conditions are listed.** The senior or the confident person picks between two unstated definitions.[^bang-2017] [^greer-2018] The pick may be right, and nothing was written that stops the recurrence.
+- **Reconciling values and not conditions.** Two numbers that match this month under different calendars split next month, and [agreement measurement](agreement-measurement.md) treats such a match as a coincidence for the same reason.
+- **Stopping at the first cause.** One cause explains most of the gap and the team calls the rest timing. Twitter's overstated user count ran for three years and Facebook's overstated watch time for two, and a gap that is mostly explained is still a gap.[^twitter-2022] [^coldewey-2016]
+- **Settling without writing.** The room agrees and nothing is dated, so the assistant keeps answering the old way. When Twitter corrected its count in 2017, its retention rules meant it could not reconcile any period before the fourth quarter of 2016.[^twitter-2017]
+- **Reconciling once and never again.** For the three to five money numbers the reconciliation is a monthly habit, as in accounting.[^sao-2020] Uber's commission error stood for two and a half years until a second way of showing drivers their pay was built.[^dickey-2017] For everything else, run it when a disagreement appears.
+
+## Sources & Stories
+
+The three internal disagreements are first-party: Airbnb's metric-platform post [^chang-2021], Uber's metric-standardization post [^yu-2021], and Zoom's progress report of April 22, 2020, whose title, body and closing paragraph describe one figure as daily users, meeting participants and people [^zoom-2020b]; the correction is cited through the TechCrunch report that reproduces the company's statement [^miller-2020]. The Twitter recasts are from its shareholder letters [^twitter-2017] [^twitter-2022], and the Facebook and Uber cases reach this page through the same contemporaneous reports the plausible-but-wrong page uses [^coldewey-2016] [^dickey-2017].
+
+The accounting lineage is the 2025 revision of the federal internal-control standards [^gao-2025], the federal audit manual's line that control totals are agreed between source and target systems [^gao-2024], the Treasury's cash-reconciliation chapter, undated, so the fetch date carries the record [^treasury-tfm], and a state auditor's guide for small governments, read in full [^sao-2020]. The two payment-system accounts are first-party engineering posts from Airbnb [^khisti-2018] and Uber [^grandhi-2024], cited for the shape of the procedure and never for the platforms. The two-times distinction is Martin Fowler's essay on bitemporal history [^fowler-2021].
+
+The evidence on how rooms pick is peer-reviewed and comes from laboratory and survey settings, so the transfer to a metrics meeting is the framework's: motivated numeracy, measured on a politically charged question [^kahan-2017], confidence matching in pairs [^bang-2017], and the hierarchy meta-analysis [^greer-2018]. The highest-paid-person's-opinion framing is quoted from the experimentation book's preface, read from the publisher-hosted sample [^kohavi-2020].
+
+No peer-reviewed account of reconciliation testing in data migration surfaced inside the ten-year window, so the audit sources carry that lineage. The three causes, the ladder applied to a metric, the reference decision before the meeting ends, and the dated line are the author's own. The prior-art row is pending in [prior-art.md](../../prior-art.md).
+
+
+
+[^bang-2017]: [[BANG-2017]](../../../REFERENCES.md)
+[^chang-2021]: [[CHANG-2021]](../../../REFERENCES.md)
+[^coldewey-2016]: [[COLDEWEY-2016]](../../../REFERENCES.md)
+[^dickey-2017]: [[DICKEY-2017]](../../../REFERENCES.md)
+[^fowler-2021]: [[FOWLER-2021]](../../../REFERENCES.md)
+[^gao-2024]: [[GAO-2024]](../../../REFERENCES.md)
+[^gao-2025]: [[GAO-2025]](../../../REFERENCES.md)
+[^grandhi-2024]: [[GRANDHI-2024]](../../../REFERENCES.md)
+[^greer-2018]: [[GREER-2018]](../../../REFERENCES.md)
+[^kahan-2017]: [[KAHAN-2017]](../../../REFERENCES.md)
+[^khisti-2018]: [[KHISTI-2018]](../../../REFERENCES.md)
+[^kohavi-2020]: [[KOHAVI-2020]](../../../REFERENCES.md)
+[^miller-2020]: [[MILLER-2020]](../../../REFERENCES.md)
+[^sao-2020]: [[SAO-2020]](../../../REFERENCES.md)
+[^treasury-tfm]: [[TREASURY-TFM]](../../../REFERENCES.md)
+[^twitter-2017]: [[TWITTER-2017]](../../../REFERENCES.md)
+[^twitter-2022]: [[TWITTER-2022]](../../../REFERENCES.md)
+[^yu-2021]: [[YU-2021]](../../../REFERENCES.md)
+[^zoom-2020b]: [[ZOOM-2020B]](../../../REFERENCES.md)
diff --git a/llms.txt b/llms.txt
index e41064a..f57b60c 100644
--- a/llms.txt
+++ b/llms.txt
@@ -24,6 +24,7 @@ Every page opens with the problems it is for, in the reader's words. The short v
- [Single point of metric computation](https://raw.githubusercontent.com/h3qing/open-analytics-framework/main/docs/modules/01-ai-data-quality/single-point-of-metric-computation.md): the same metric returns different numbers in different tools, and nobody can say which copy is right. A number is quoted outside the team that produces it, in a deck or by an AI assistant. Template: [Metric-computation control plan](https://raw.githubusercontent.com/h3qing/open-analytics-framework/main/templates/01-ai-data-quality/metric-computation-control-plan.md).
- [Detecting plausible-but-wrong outputs](https://raw.githubusercontent.com/h3qing/open-analytics-framework/main/docs/modules/01-ai-data-quality/plausible-but-wrong.md): a revenue or usage figure in a deck that nobody has tied to the billing system. A total that jumped after a model change added a join. A metric that has been quietly counting test accounts or refunded orders. An assistant returns a fluent answer and a query, and nobody can say whether the join is right. Template: [Plausibility-check control plan](https://raw.githubusercontent.com/h3qing/open-analytics-framework/main/templates/01-ai-data-quality/plausibility-check-control-plan.md).
- [Schema drift detection and data contracts](https://raw.githubusercontent.com/h3qing/open-analytics-framework/main/docs/modules/01-ai-data-quality/schema-drift-and-data-contracts.md): the number moved last week and nobody touched the query. A dashboard went blank because a column was renamed. A status field grew a new value and the report files it under other. A field started arriving empty and the total drifted down. Somebody backfilled a table and last quarter no longer matches the deck. Template: [Schema contract control plan](https://raw.githubusercontent.com/h3qing/open-analytics-framework/main/templates/01-ai-data-quality/schema-contract-control-plan.md).
+- [The reconciliation protocol](https://raw.githubusercontent.com/h3qing/open-analytics-framework/main/docs/modules/01-ai-data-quality/reconciliation-protocol.md): finance and the warehouse give two figures for last month's revenue and the deck is due. This quarter's number does not match the version that went out last quarter. The assistant and the analyst are a few percent apart and both queries look right. Two dashboards show two totals for one metric and each owner is sure. The meeting ended with "we will look into it" for the third time. Template: [Reconciliation record](https://raw.githubusercontent.com/h3qing/open-analytics-framework/main/templates/01-ai-data-quality/reconciliation-record.md).
### An AI assistant answers questions about your data
diff --git a/templates/01-ai-data-quality/reconciliation-record.md b/templates/01-ai-data-quality/reconciliation-record.md
new file mode 100644
index 0000000..2dca97b
--- /dev/null
+++ b/templates/01-ai-data-quality/reconciliation-record.md
@@ -0,0 +1,227 @@
+# Reconciliation record
+
+Control artifact for the pattern [The reconciliation protocol (M1-08)](../../docs/modules/01-ai-data-quality/reconciliation-protocol.md). MIT licensed. Copy this file into your own repository, fill one copy per disagreement, replace every `[TODO: …]` field, and delete the guidance blockquotes. The procedure follows the reconciliations paragraph of the federal internal-control standards [GAO-2025] and the working rules in a state auditor's bank-reconciliation guide [SAO-2020], and the two-times rule follows the bitemporal distinction between what was true and what was recorded [FOWLER-2021] (see [REFERENCES.md](../../REFERENCES.md)); everything organization-specific is a `[TODO]`.
+
+> **How to use with an AI analytics agent:** sections 1 to 4 are filled in during the hour the two owners spend together, and an agent can draft sections 2 and 4 from the two queries with Prompt A. Section 5 is the decision and stays with the person who signs it. Section 6 is the machine-readable core, and section 7 is the prompt library. The smallest complete record is two numbers with their as-of times, one ladder that closes to zero, and one name with a date.
+>
+> **What this record settles:** which of two numbers already in the room stands, and why. Checking one number against something outside its query is the [plausibility-check control plan](plausibility-check-control-plan.md); naming the one computation every consumer reads is the [metric-computation control plan](metric-computation-control-plan.md). This record feeds both.
+
+## 1. The two numbers
+
+One row per fact. Fill both columns before anyone explains anything.
+
+| | Side A | Side B |
+|---|---|---|
+| The number | `[TODO: e.g. 412,380]` | `[TODO: e.g. 396,150]` |
+| What it claims to be | `[TODO: metric name and period, e.g. revenue, March]` | `[TODO: same words, as the other side says them]` |
+| System or tool it came from | `[TODO: e.g. the warehouse's blessed revenue view]` | `[TODO: e.g. the billing system's monthly statement]` |
+| Query, export or link | `[TODO: path or link]` | `[TODO: path or link]` |
+| Who pulled it | `[TODO: name]` | `[TODO: name]` |
+| Pulled at (as-of time, with timezone) | `[TODO: e.g. 2026-04-03 09:00 UTC]` | `[TODO: e.g. 2026-04-03 11:30 UTC]` |
+| Is this the blessed computation? | `[TODO: yes / no / none exists for this metric]` | `[TODO: yes / no / none exists for this metric]` |
+| Gap (A minus B) | `[TODO: e.g. 16,230]` | |
+
+> A number is a value and a moment. Two pulls of the same query at different times are two numbers, and the difference between them belongs on the ladder as a line of its own, with its size.
+
+## 2. The conditions behind each
+
+One row per condition. Write what each side actually did, from its query or export, not what it was supposed to do. A row marked "different" is a candidate line on the ladder.
+
+| Condition | Side A | Side B | Same? |
+|---|---|---|---|
+| Records excluded (test accounts, refunds, internal, cancelled, trials) | `[TODO]` | `[TODO]` | `[TODO: same / different]` |
+| Calendar (calendar month, fiscal period, week start, timezone of the day boundary) | `[TODO]` | `[TODO]` | `[TODO]` |
+| As-of time (when the data was pulled; whether late-arriving rows are in) | `[TODO]` | `[TODO]` | `[TODO]` |
+| Grain (one row per what) | `[TODO]` | `[TODO]` | `[TODO]` |
+| Unit and currency (cents or dollars; conversion rate and its date) | `[TODO]` | `[TODO]` | `[TODO]` |
+| Definition of the thing counted (what "customer" or "order" means here) | `[TODO]` | `[TODO]` | `[TODO]` |
+| Join path (tables joined, and the key on each side) | `[TODO]` | `[TODO]` | `[TODO]` |
+| Anything else that differs | `[TODO]` | `[TODO]` | `[TODO]` |
+
+> Fill this before anyone argues about which number is right. Every executive escalation the author has traced came down to a row in this table, and most to the first three.
+
+## 3. The finest common grain
+
+The finest grain both sides can be brought to, so the gap can be located rather than argued about.
+
+| Item | Value |
+|---|---|
+| Common grain | `[TODO: e.g. by day and by customer; or by week and by plan if one side cannot go finer]` |
+| How Side A was brought to it | `[TODO: query or export step]` |
+| How Side B was brought to it | `[TODO: query or export step]` |
+
+| Segment | Side A | Side B | Difference | Share of gap |
+|---|---|---|---|---|
+| `[TODO: e.g. week 1]` | `[TODO]` | `[TODO]` | `[TODO]` | `[TODO]` |
+| `[TODO: e.g. week 2]` | `[TODO]` | `[TODO]` | `[TODO]` | `[TODO]` |
+| … | | | | |
+| Total | `[TODO]` | `[TODO]` | `[TODO: must equal the gap in section 1]` | 100% |
+
+> Reading the table: a gap spread evenly across every segment usually means a definition or a unit; a gap in one period usually means a calendar; a gap in a handful of rows usually means a join or an exclusion. The audit rule for any interface between two systems is that control totals are agreed between source and target [GAO-2024]; this table is that rule one grain finer.
+
+## 4. The difference ladder
+
+One line per cause. Start at Side A, apply one condition, write the amount it explains, carry the running total, and arrive at Side B. Sign convention: an amount is negative when applying the condition moves the running total down.
+
+| Step | Cause (from the section 2 row) | Condition on A | Condition on B | Amount | Running total |
+|---|---|---|---|---|---|
+| Start | Side A as pulled | | | | `[TODO: Side A value]` |
+| 1 | `[TODO: e.g. refunds]` | `[TODO: e.g. gross]` | `[TODO: e.g. net of refunds]` | `[TODO: e.g. −5,940]` | `[TODO]` |
+| 2 | `[TODO: e.g. calendar]` | `[TODO: e.g. fiscal period]` | `[TODO: e.g. calendar month]` | `[TODO]` | `[TODO]` |
+| 3 | `[TODO]` | `[TODO]` | `[TODO]` | `[TODO]` | `[TODO]` |
+| End | Side B as pulled | | | | `[TODO: Side B value]` |
+| | **Remainder** (last running total minus Side B) | | | | `[TODO: must be 0]` |
+
+Spreadsheet-ready: `running_total[n] = running_total[n-1] + amount[n]`, and `remainder = running_total[last] - side_b`. The sum of the amounts equals `side_b - side_a`.
+
+> The ladder ends at zero or the number is on hold. A remainder that is not zero goes into section 5 as "unexplained", with its size, an owner and a date, and nobody signs the reference decision until it closes. "Timing" is a valid cause only with an amount beside it.
+
+## 5. Reference decision
+
+Filled before the meeting ends, by the one person who signs it.
+
+| Item | Value |
+|---|---|
+| Which number stands | `[TODO: Side A / Side B / neither; on hold]` |
+| Why | `[TODO: e.g. the definition says net of refunds on the calendar month; Side A's view had the defects, and is corrected to match]` |
+| Which ladder lines were defects, and which were definitional choices | `[TODO: e.g. lines 1 and 2 definitional, now written down; line 3 a defect in the join; line 4 timing]` |
+| Decided by | `[TODO: name]` |
+| Date | `[TODO: date]` |
+| Blessed computation | `[TODO: existed and stands / existed and is corrected in change X / did not exist; Side [A/B] is now named as it, per the metric-computation control plan]` |
+| Definitions line (dated, verbatim, appended and never edited in place) | `[TODO: e.g. 2026-04-03: revenue is net of refunds, on the calendar month, as of the billing statement date]` |
+| Context-file line (dated, verbatim) | `[TODO: the same fact in the words the assistant reads]` |
+| Losing copy | `[TODO: retired on [date] / kept with a drift check against the reference]` |
+| Numbers already sent out carrying the losing figure | `[TODO: which deck, which date, who received it; or none]` |
+| Disclosure question | `[TODO: yes / no. Whether a corrected number that already left the company is a restatement belongs to Module 4]` |
+| Unexplained remainder, if any | `[TODO: amount, owner, due date; else none]` |
+| Same pair, same cause before? | `[TODO: no / yes, record dated [date]. If yes, this goes to root cause analysis]` |
+
+> Averaging the two numbers, or taking the one that fits the story, is never a valid entry in the first row. The senior person may choose; the choice is made after section 2 is full and section 4 closes, and is written here so it does not have to be made again.
+
+## 6. Machine-readable core
+
+```json
+{
+ "spec": "reconciliation-record/v1",
+ "metric": "[TODO: e.g. revenue]",
+ "period": "[TODO: e.g. 2026-03]",
+ "sides": {
+ "a": { "value": "[TODO]", "source": "[TODO]", "query": "[TODO: path]", "pulled_by": "[TODO]", "as_of": "[TODO: ISO 8601 with timezone]", "blessed": "[TODO: true | false | null]" },
+ "b": { "value": "[TODO]", "source": "[TODO]", "query": "[TODO: path]", "pulled_by": "[TODO]", "as_of": "[TODO: ISO 8601 with timezone]", "blessed": "[TODO: true | false | null]" }
+ },
+ "gap": "[TODO: a.value - b.value]",
+ "conditions": [
+ { "name": "excluded", "a": "[TODO]", "b": "[TODO]", "same": "[TODO: true | false]" },
+ { "name": "calendar", "a": "[TODO]", "b": "[TODO]", "same": "[TODO]" },
+ { "name": "as_of", "a": "[TODO]", "b": "[TODO]", "same": "[TODO]" },
+ { "name": "grain", "a": "[TODO]", "b": "[TODO]", "same": "[TODO]" },
+ { "name": "unit_currency", "a": "[TODO]", "b": "[TODO]", "same": "[TODO]" },
+ { "name": "definition", "a": "[TODO]", "b": "[TODO]", "same": "[TODO]" },
+ { "name": "join_path", "a": "[TODO]", "b": "[TODO]", "same": "[TODO]" }
+ ],
+ "common_grain": "[TODO: e.g. day x customer]",
+ "ladder": [
+ { "step": 1, "cause": "[TODO]", "condition_a": "[TODO]", "condition_b": "[TODO]", "amount": "[TODO: signed number]", "kind": "[TODO: defect | definitional | timing]" }
+ ],
+ "remainder": "[TODO: must be 0 to sign]",
+ "decision": {
+ "reference": "[TODO: a | b | hold]",
+ "why": "[TODO]",
+ "decided_by": "[TODO]",
+ "date": "[TODO]",
+ "blessed_computation": "[TODO: stands | corrected | named_now]",
+ "definitions_line": "[TODO: dated, verbatim]",
+ "context_file_line": "[TODO: dated, verbatim]",
+ "losing_copy": "[TODO: retired_on date | drift_checked]",
+ "already_sent": "[TODO: list or none]",
+ "disclosure_question": "[TODO: true | false]",
+ "recurrence": "[TODO: false | record id of the earlier one]"
+ },
+ "sources": ["GAO-2025", "SAO-2020", "FOWLER-2021", "GAO-2024"]
+}
+```
+
+> The JSON restates sections 1 to 5 in a form any agent can parse. Keep them in step in the same edit; the JSON wins when the prose is ambiguous.
+
+## 7. Agent prompt templates
+
+### Prompt A — draft the conditions and the ladder from two queries
+
+```text
+Two numbers for the same metric disagree. Below are the two values, the
+time each was pulled, and the query or export behind each. Rules:
+1. For each side, read its query and fill the conditions table: records
+ excluded, calendar, as-of time, grain, unit and currency, definition
+ of the thing counted, join path. Write what the query does, not what
+ it was meant to do. Mark each row same or different.
+2. Propose the finest grain both sides can reach, and if the data is
+ available, compute the difference by segment at that grain.
+3. Build the difference ladder from Side A to Side B: one line per
+ condition marked different, with the signed amount it explains and
+ the running total. Show the arithmetic so a person can repeat it in a
+ spreadsheet.
+4. Report the remainder. If it is not zero, say so, name it as
+ unexplained, and list what you would check next. Do not spread the
+ remainder across the lines to make it close.
+5. Do not say which number is right. Do not average them. Stop after
+ the ladder.
+
+
+[paste section 6 JSON with sides filled in]
+
+
+
+[paste the two queries or export definitions]
+
+```
+
+### Prompt B — draft the dated lines, for the person who signs
+
+```text
+A disagreement between two numbers has been reconciled and the reference
+has been decided. Below is the record. Draft, for the owner to approve:
+1. One dated line for the metric definition stating the settled
+ condition, with the period boundary, exclusion, unit or definition
+ spelled out so the question cannot be read the other way again.
+2. The same fact as one dated line for the business context file, in
+ the words an assistant needs before answering this question.
+3. If the losing side was a copy of the metric, one sentence naming it
+ as retired on a date or as kept with a drift check against the
+ reference.
+Append the lines; never edit an existing line in place. Do not change the
+decision. If the record's remainder is not zero, refuse and say why.
+
+
+[paste section 6 JSON, filled in through the decision]
+
+```
+
+### Prompt C — before answering a question that has split the room before
+
+```text
+Before answering the question below, check the reconciliation records
+and the dated lines in the context file for this metric. If a settled
+condition applies, use it and cite the dated line in your answer. If the
+question could be read two ways and no dated line settles it, say which
+reading you used, and list the conditions (excluded, calendar, as-of,
+grain, unit, definition) under your number so it can be reconciled
+against another.
+
+
+[paste the question]
+
+```
+
+> Run Prompt A the moment two numbers disagree, before the meeting if possible. Run Prompt C on every question that has appeared in a reconciliation record.
+
+## 8. Ownership and review
+
+| Item | Value |
+|---|---|
+| Owner of this record | `[TODO: the person who signed section 5]` |
+| Where records live | `[TODO: repository path; one file per disagreement, dated]` |
+| Time allowed | One hour with both owners present for sections 1 to 5; `[TODO: e.g. two working days]` for an unexplained remainder to close |
+| Standing reconciliation | For the `[TODO: three to five]` money numbers, run sections 1 to 4 monthly against the independent total, whether or not anyone has noticed a gap |
+| Review rule | The reference decision and the dated lines are reviewed like code. A dated line is appended, never edited. |
+| Recurrence rule | The same pair of numbers disagreeing twice for the same cause is a finding, and the case goes to root cause analysis rather than to a second record. |
+| Standing check | If no record has been filed in `[TODO: e.g. a quarter]`, confirm that disagreements are still reaching this file and not being settled by the senior person in the room. |
diff --git a/templates/README.md b/templates/README.md
index 4bb434d..865faa5 100644
--- a/templates/README.md
+++ b/templates/README.md
@@ -6,7 +6,7 @@ In the DMAIC mapping ([pattern template](../docs/pattern-template.md)), the arti
| Directory | Module | Contents |
|---|---|---|
-| `01-ai-data-quality/` | AI data quality | [Retention-state control plan](01-ai-data-quality/retention-state-control-plan.md) (M1-11) · [Metric-computation control plan](01-ai-data-quality/metric-computation-control-plan.md) (M1-01) · [SPC control plan](01-ai-data-quality/spc-control-plan.md) (M1-02) · [Golden question set](01-ai-data-quality/golden-question-set.md) (M1-04) · [Plausibility-check control plan](01-ai-data-quality/plausibility-check-control-plan.md) (M1-03) · [Schema contract control plan](01-ai-data-quality/schema-contract-control-plan.md) (M1-07) · [Agreement-measurement control plan](01-ai-data-quality/agreement-measurement-control-plan.md) (M1-05) |
+| `01-ai-data-quality/` | AI data quality | [Retention-state control plan](01-ai-data-quality/retention-state-control-plan.md) (M1-11) · [Metric-computation control plan](01-ai-data-quality/metric-computation-control-plan.md) (M1-01) · [SPC control plan](01-ai-data-quality/spc-control-plan.md) (M1-02) · [Golden question set](01-ai-data-quality/golden-question-set.md) (M1-04) · [Plausibility-check control plan](01-ai-data-quality/plausibility-check-control-plan.md) (M1-03) · [Schema contract control plan](01-ai-data-quality/schema-contract-control-plan.md) (M1-07) · [Agreement-measurement control plan](01-ai-data-quality/agreement-measurement-control-plan.md) (M1-05) · [Reconciliation record](01-ai-data-quality/reconciliation-record.md) (M1-08) |
| `02-infrastructure-design/` | Infrastructure design | _pending_ |
| `03-ai-agent-integration/` | AI agent integration | [Business context file](03-ai-agent-integration/context-file.md) (M3-06) |
| `04-governance-and-financial-reporting/` | Governance and financial reporting | _pending_ |