A proposal source-conflict record binds one exact buyer-facing proposition to every disagreeing source version, each source’s definition, scope, period and derivation, the accountable owner, the reconciliation work, the approved result, affected artifacts and the trigger for reopening the decision.
A transport authority asks for the hosted dispatch service’s availability over the latest completed twelve months. The fictional Harborline Response Systems Ltd finds 99.98% in an approved sales deck, 99.93% in the service dashboard and 99.89% in the incident register. The sales figure concerns one reference customer and an older period. The dashboard covers both production regions but excludes planned maintenance and customer-caused outages. The incident register records every disruption, including degraded service that the contract does not classify as unavailable. The bid team cannot select the largest number, average the three values or call one application the universal source of truth. It must first define the proposition the buyer will evaluate, then reconcile the evidence that can answer it.
Authority is specific to a claim. A system may own raw events, a controlled calculation may own a metric, and an approver may own permission to publish the resulting sentence. Resolve the disagreement at those boundaries. Preserve both the accepted and rejected values with their lineage so that a later reviewer can reproduce the decision and correct every place where the wrong claim travelled.
Start with one proposition, not a contest between files
A conflict exists only after the team states what it is trying to prove. “Availability” can mean an SLA result, all observed service disruption, successful transaction time, a contractual service-credit measure or a marketing summary. Three files carrying that label may report different objects. Asking which file wins before defining the proposition invites a political answer: sales trusts the deck, operations trusts the incident tool, and finance trusts the report it signs.
Freeze the buyer’s words, definitions and requested period. Then write the proposed sentence exactly as it would appear in the response. “DispatchLink achieved 99.95% availability during the twelve months ending 30 June 2026 across the two production regions” contains a service, threshold, measure, period and geography. If the tender asks for contractual SLA performance, the record also needs the buyer’s permitted exclusions and treatment of partial degradation. Each element becomes a test against the sources.
The consequence sets the depth of review. A number used only as background still needs support. A figure tied to a scored criterion, warranty, service credit or eligibility threshold needs independent checking and formal approval. The UK Government Data Quality Framework defines quality in relation to purpose, and GAO’s data-reliability guidance likewise tests accuracy, completeness and applicability for the intended use. Neither supports a universal ranking of applications. The intended proposition comes first.
Preserve the disagreement before trying to repair it
Capture every conflicting source as it existed when found. Keep the native record where policy permits, or a controlled export with the system, report name, parameters, version, observation time, retrieval time, owner and stable locator. A screenshot can preserve what a dashboard displayed, but it rarely preserves the query, filters, refresh state or calculation. Link the screenshot to those details rather than treating the image as the whole source.
W3C PROV separates entities, activities and agents. That is a useful model for a conflict record: the incident events are entities, a classification and calculation are activities, the dashboard output is another entity, and named people or systems carry responsibility for the steps. Derivation records how one version came from another. Revision and invalidation let the team preserve a former value without presenting it as current. The proposal does not need to publish PROV data, but the internal record should answer the same questions about origin and transformation.
Do not clean up the evidence before capturing it. Editing the sales deck, refreshing the dashboard or relabelling an incident can remove the very state the team must explain. ISO 15489-1 covers the creation, capture and management of records with metadata, responsibilities and controls. Use that principle within the company’s own records policy. Preservation does not authorize broad copying: restricted exports, personal data and customer material stay inside their permitted access boundary.
Most apparent number conflicts begin in the denominator
Put the values beside their definitions rather than beside their file names. Capture the legal entity, service, version, environment, region, customer population, event population, time zone, start and end instants, unit, numerator, denominator, exclusions and final or provisional status. A one-year figure computed daily can differ from a mean of twelve monthly percentages. A dashboard that removes planned maintenance can disagree with a raw event register while both remain internally consistent.
Harborline’s three values do not yet compete. The 99.98% deck value describes one reference customer during the prior calendar year. The 99.93% dashboard value covers the offered service’s two regions for the requested period, using the company SLA formula. The 99.89% incident-derived value covers the same dates but counts degraded events and planned maintenance that the SLA method excludes. Only the dashboard is close to the buyer’s question, and it still needs its derivation and the buyer’s own definition checked.
The Government Data Quality Framework names completeness, uniqueness, consistency, timeliness, validity and accuracy as core dimensions, while its guidance asks users to set rules tied to a specific use. Those dimensions reveal different failures. Two complete records can be inconsistent; two consistent reports can both apply the wrong definition. Record the precise dimension and consequence instead of reducing the conflict to a red or green quality badge.
| Source and value | Population and period | Method | Conflict finding |
|---|---|---|---|
| Sales deck: 99.98% | One reference customer, January to December 2025 | Approved case-study result | Different customer, period and intended use |
| Service dashboard: 99.93% | Two production regions, July 2025 to June 2026 | SLA formula with stated exclusions | Candidate answer, pending derivation and buyer-definition review |
| Incident register: 99.89% | Same regions and period | All disruption and degradation minutes divided by elapsed time | Different event and exclusion rules |
Four kinds of authority sit behind one proposal number
Raw-record authority answers what was captured: incident start, affected region, restoration time and classification history. Definition authority answers what counts as unavailable for a stated purpose. Calculation authority owns the tested transformation from inputs to metric. Release authority decides whether the company may make that external representation in the bid. One person may hold two roles, but the record should keep the decisions distinct.
The UK government’s 2026 data ownership model distinguishes owners, stewards and custodians. Owners are accountable for meaning, content, quality and major changes. Stewards handle operational governance and quality processes. Custodians operate the systems and implement technical rules. This is a public-sector model, not a mandatory organization chart for every bidder. Its value here is the separation: access to a database does not grant authority to redefine the metric or approve a contractual statement.
Approved narrative has its own narrow role. A sales deck can prove that a sentence passed a prior communications process. It cannot prove that the underlying events remain accurate for another service and period. Equally, a raw log does not interpret contract exclusions by itself. Assign authority to the proposition part that each source or role can settle, then require the claim approver to see the joined result and its limitations.
| Authority | Question it settles | Candidate record or role | What it cannot approve alone |
|---|---|---|---|
| Raw record | Which service events occurred? | Incident system and service operations owner | The buyer’s SLA interpretation |
| Definition | Which events and periods count? | Approved service-measure dictionary and service owner | Whether the computation used the rule correctly |
| Calculation | How do inputs produce 99.93%? | Controlled query, code or workbook and analytical reviewer | Permission to promise the result externally |
| Release | Which bounded sentence may enter this proposal? | Named claim and bid-release approvers | Changing a source fact without its owner |
Reconcile the derivation, then decide whether a value survives
Reproduction is stronger than visual comparison. Start from the agreed input population and run the documented formula with its version, parameters and exclusions. Reconcile record counts at each step. Review manual classifications separately. If the dashboard cannot expose its query, ask its owner for a controlled extract or independent calculation. A matching output supports the derivation; it does not settle whether the selected definition fits the tender.
Classify why the values differ. Common classes are scope, period, definition, unit, source correction, extraction, transformation, manual adjustment, approval state and display latency. Some conflicts close without rejecting either source. The incident-derived 99.89% and SLA-derived 99.93% are both reproducible after Harborline identifies their event rules. They need different names. The sales-deck value remains valid only for its original case-study claim.
GAO’s reliability guide recommends a risk-based assessment of accuracy, completeness and applicability, using existing information and involving appropriate people. The 2025 AQuA Book adds data, assumption and decision logs, version control, verification, validation and reconciliation to independent sources. Apply those ideas proportionately. A high-consequence tender claim may need a second analyst or service owner to reproduce the result; a descriptive statement may need a lighter documented check.
A disposition explains the conflict instead of hiding it
The decision record needs more than “dashboard confirmed.” Name the proposition, source versions, conflicting values, reason for the difference, checks performed, selected treatment, rejected alternatives, accountable decision maker, approval time and proposal locations covered. Link to controlled evidence rather than duplicating protected data into the bid workspace. The result should tell another reviewer exactly what they may repeat.
Use states that preserve uncertainty. reconciled_same_fact means the values now agree after an identified correction or common method. different_definitions means each value is valid for a different proposition. source_corrected means an authorized owner issued a new source state. narrower_claim_required keeps only the supported portion. conflict_unresolved blocks the material claim. specialist_review_required routes a legal, financial, statistical, security or contractual interpretation without pretending the workflow decided it.
A manager may accept a disclosed residual uncertainty within their authority. Record that as an approval of the bounded use, not as retroactive proof of the number. If the tender demands an exact figure and the conflict remains material, a range or caveat may still fail the requirement. The bid owner must route compliance consequences separately and remove the claim when no approved answer can be supported.
| State | Meaning | Proposal treatment | Next control |
|---|---|---|---|
| reconciled_same_fact | Common definition and method now produce one supported value | Use the approved value with scope and period | Bind to source versions and expiry |
| different_definitions | Values answer different questions | Use only the metric matching the buyer’s definition | Rename both metrics and prevent substitution |
| narrower_claim_required | Only part of the proposed sentence is supported | Release the smaller approved claim | Keep excluded scope visible |
| conflict_unresolved | Material disagreement remains | Block or remove the proposition | Escalate to the named authority |
| source_unavailable | A necessary source cannot be inspected | Do not infer its value or status | Obtain authorized access or narrow the claim |
Correction continues after the proposal sentence changes
A wrong value rarely lives in one draft. Search the active bid, related annexes, executive summary, evidence table, presentation, answer library, case-study repository and source dashboard. Record each occurrence as affected, reviewed, corrected, intentionally retained for another proposition or outside the authorized scope. Do not silently update a reusable answer while leaving a signed or submitted artifact looking current.
Eurostat’s revision policy treats revisions as a normal way to incorporate improved information and calls for standard, transparent procedures and documentation. Proposal material is not official statistics, yet the correction lesson transfers well: preserve what changed, why, when and which users need to know. A revision can arise from new source data, a method change, a classification change or an error. Those causes have different reopen consequences.
Correct the source through its owner. The bid team should not overwrite an operational classification merely to align the proposal. Where the dashboard formula is wrong, issue a new controlled version and mark the old one superseded. Where the deck used a correct but differently scoped number, improve the label and reuse rules rather than deleting the historical record. Notify only the people and systems allowed to receive the underlying detail.
| Artifact | Old use | Approved treatment | Completion evidence |
|---|---|---|---|
| Current proposal | 99.98% presented as service-wide availability | Replace with approved 99.93% SLA result and exact period | Reviewed response version and approver |
| Sales deck | 99.98% label omitted customer and period | Keep the case result but add its original scope | New controlled slide version |
| Answer library | Generic availability sentence | Withdraw pending source-specific reuse rule | Entry state and affected-answer search |
| Service dashboard | Correct value with an ambiguous label | Clarify SLA formula, exclusions and refresh time | Data-owner-approved metadata update |
Release the bounded claim and keep the decision reproducible
Harborline reproduces the 99.93% result from the controlled incident export using the approved SLA rules for both production regions from 1 July 2025 through 30 June 2026. A service owner confirms the event classifications, a second reviewer checks the calculation, and the bid approver authorizes this sentence: “DispatchLink recorded 99.93% SLA availability across its two production regions during the twelve months ending 30 June 2026, using the stated service-measure rules.” The answer links the confidential calculation internally and supplies only permitted supporting material to the buyer.
The 99.89% value remains in the conflict record as all-disruption availability. The 99.98% value remains a reference-customer result for 2025. Clear labels prevent either from being reused as the new service-wide SLA figure. The decision expires when the measurement period rolls forward, the SLA definition changes, an incident is reclassified, the dashboard query changes or an audit finds another affected event.
Software can detect repeated values, compare source metadata, preserve snapshots and run a controlled calculation. It can also produce the propagation list. Human owners still settle disputed business meaning, contractual interpretation, material corrections and external approval. The final record gives them a compact decision packet: proposition, conflict, lineage, tests, selected treatment, residual limitation, affected artifacts and the exact approval required.
What good looks like
Useful outcomes from resolve conflicting proposal evidence
- The exact buyer requirement and proposed sentence, split into testable propositions
- An immutable capture of every conflicting source, version, owner, access route and observation time
- A comparison of entity, service, population, event definition, calculation, exclusions, period and status
- A derivation trail from raw records through transformations to each displayed value
- A claim-specific authority decision that keeps source custody, analysis and release approval separate
- One of reconciled_same_fact, different_definitions, source_corrected, narrower_claim_required, conflict_unresolved, source_unavailable, specialist_review_required or claim_removed
- Approved wording with the selected value, unit, scope, period, definition and necessary limitation
- A correction and propagation list for affected proposal drafts, answer-library entries, decks and data products
- A named reviewer, decision time, rationale and reopen trigger
Operating model
How to run the work
- 01
Freeze the disputed proposition
Copy the complete buyer wording and the sentence proposed for release. Split the number, population, service, period, unit, exclusions and any comparison into separate propositions.
- 02
Preserve every source version
Capture the native record or controlled export, source owner, system, version, retrieval time, access boundary and a stable locator before anyone edits a dashboard or deck.
- 03
Describe what each value measures
For every value, record the entity, service, geography, population, period, event definition, numerator, denominator, exclusions, status and stated purpose.
- 04
Trace the derivation
Follow raw events through filters, joins, classifications, formulas, manual adjustments and approvals. Mark any missing step instead of inventing lineage from a file name.
- 05
Assign authority by proposition
Identify who owns the raw record, the business definition, the calculation, the source correction and the right to approve external wording. Do not collapse those decisions into one generic owner.
- 06
Reconcile or bound the conflict
Recompute from an agreed population where possible, test the source definitions and classify the remaining disagreement. Narrow or remove the sentence when no result is fit for the buyer’s use.
- 07
Approve one release record
Record the selected value or wording, rejected alternatives, rationale, source anchors, limitations, reviewer and exact proposal locations covered by the approval.
- 08
Correct and reopen deliberately
Propagate the decision to every affected artifact without erasing history. Reopen after a source revision, definition change, new event, period rollover, calculation change or discovery of another conflicting use.
Evaluation
Questions that change the decision
- What exact proposition will the evaluator read, including value, unit, population and period?
- Do the sources disagree about the same fact or describe different facts with similar labels?
- Which source contains observations, which contains a transformation and which contains approved narrative?
- Are the values final, provisional, corrected, estimated, manually adjusted or superseded?
- Can each displayed number be reproduced from identified inputs and an approved method?
- Who owns the business definition, raw records, calculation, correction and external disclosure?
- Does the buyer define the metric or require a calculation different from the company’s normal report?
- Does the conflict change eligibility, scoring, price, contractual exposure or only supporting context?
- Should the final answer use one value, a narrower statement, a range, a disclosed limitation or no claim?
- Which published and reusable artifacts must be corrected once the decision is approved?
Failure modes
Where teams lose control
Selecting the largest or most persuasive number without testing what it measures
Declaring one application the source of truth for facts it does not define or calculate
Overwriting a source before preserving the disputed version and its metadata
Comparing percentages that use different populations, event rules, units or periods
Treating an approved sales sentence as proof that the underlying analysis is correct
Letting the data custodian approve a commercial or legal representation outside that role
Recomputing a value with undocumented filters or manual spreadsheet changes
Calling an unresolved conflict a conservative estimate without an approved basis
Correcting the live bid while leaving the wrong value in the answer library or sales deck
Exposing confidential source records or personal data merely to show that reconciliation occurred
Measurement
Measure the finished job
Measure the completed workflow, including review effort and exceptions. Output volume on its own is not evidence of a better process.
- Number of material proposal claims with more than one active internal value
- Percentage of disputed values with complete source, definition, period and derivation metadata
- Median time from conflict detection to an authorized disposition
- Percentage of selected metrics reproducible from preserved inputs and methods
- Number of conflicts resolved by correcting a source rather than choosing between displays
- Number of proposal claims narrowed or removed because the conflict remained material
- Count of affected artifacts found and corrected after each decision
- Number of reopened decisions caused by definition, period or source changes
- Recurrence rate for the same root-cause conflict across later bids
Questions
Common questions
Should the system of record always win?
No. A system of record may be authoritative for raw events but not for a derived metric, contractual definition or approved external statement. Identify the authority needed for each part of the proposition and trace the calculation between them.
Can we use the lower number to be safe?
Only when that number measures the buyer’s proposition and its use is approved. A lower value with the wrong population or denominator is still wrong. If the conflict is unresolved, narrow the wording, disclose the limitation where appropriate or remove the claim.
What if both sources are correct?
Record different_definitions when each value is valid for a different scope, period or method. Rename the metrics clearly, then select only the one that matches the buyer’s requested proposition.
Does management approval settle the factual conflict?
Approval can authorize a decision or a bounded statement. It cannot make incompatible source facts equal. The record should show what evidence was reconciled, what remains uncertain and what the approver was entitled to decide.
Should we average the two values?
Not unless the method, weighting, units and populations make aggregation valid and the result answers the tender requirement. Averaging is not a conflict-resolution rule.
What belongs in the conflict record?
Keep the proposition, source versions, values, definitions, periods, populations, transformations, owners, conflict type, tests, decision, rationale, approved wording, affected artifacts, access limits and reopen trigger.
Do we have to fix the underlying source before submitting?
A material wrong source should enter its authorized correction process. The proposal may proceed only if an accountable reviewer approves a traceable value or narrower statement and the remaining conflict does not make the answer misleading or noncompliant.
Can software choose the authoritative source automatically?
Software can find conflicting values, preserve versions, compare metadata and recompute an approved formula. It should not invent business definitions, transfer decision rights or release a material claim when authority or meaning remains disputed.
Sources
Primary references
- UK Government Data Quality Framework UK Government Data Quality Hub
- UK Government Data Quality Framework practical guidance UK Government Data Quality Hub
- UK government data ownership model Government Digital Service
- Government Functional Standard GovS 010: Analysis UK Government Analysis Function
- The AQuA Book: guidance on producing quality analysis UK Government Analysis Function
- GAO guide for assessing data reliability U.S. Government Accountability Office
- GAO 2025 Green Book for internal control U.S. Government Accountability Office
- W3C PROV Data Model World Wide Web Consortium
- W3C PROV Ontology World Wide Web Consortium
- W3C Constraints of the PROV Data Model World Wide Web Consortium
- W3C Data Quality Vocabulary World Wide Web Consortium
- ISO 8000-1:2022 overview of data quality International Organization for Standardization
- ISO 15489-1:2016 records management concepts and principles International Organization for Standardization
- NIST Research Data Framework, second release National Institute of Standards and Technology
- NIST definition and source map for provenance National Institute of Standards and Technology
- Eurostat data revision policy Eurostat
- Administrative Data Quality Assurance Toolkit Office for Statistics Regulation
- Code of Practice for Statistics, edition 3.0 UK Statistics Authority
About the author
Tony Kim
Tony writes about applied AI, dependable product engineering and the systems that turn complex response work into controlled delivery.
Ziva
Proposal software for source-grounded RFP, RFI, DDQ and questionnaire response work.
Bid, proposal, presales, security and compliance teams. Start with the workflow, constraints and evidence you already have.