Site Financial
Health Engine.
From fragmented reporting to accountable variance investigation.
A six-version record: an SAP and planning reporting foundation that became the Financial Variance Command Center — a transparent six-agent investigation room where every variance carries a theory, a trace, and a human disposition.
Can a variance be made accountable rather than merely reported?
Financial reporting reliably tells an organization what moved. It rarely tells it why the movement happened, who investigated it, what was ruled out, and what remains unexplained. The number arrives; the theory behind it does not.
System 008 began as a reporting foundation — the Site Financial Health Engine — and has since become a research question: if calculation is separated from interpretation, and interpretation is assigned to bounded, inspectable roles, does the variance conversation become faster, more complete, and more auditable?
Every variance deserves a theory.
What is public, and what it is not.
The Financial Variance Command Center is published as a synthetic, browser-only research release. The reporting foundation it grew out of remains an internal, in-development enterprise system; only its written record is published here.
Six versions, each answering the failure of the one before it.
- V12026 · Q1Fragmented reportingFive sources, reconciled by hand every close.
KSB1 actuals, FBL3N general ledger detail, Dodeca Plan and LBE extracts, cost center and GL master data, and operational KPI exports arrived in different shapes on different cadences. Each close was rebuilt manually: combine, enrich, validate, standardize, then finally analyze.
LessonThe constraint was never data availability. Effort spent preparing data had quietly overtaken the time available to interpret it.
- V22026 · Q1Automated pipelineExtraction, validation, normalization, mapping, refresh, delivery.
A deterministic PowerShell pipeline replaced the spreadsheet choreography: extract from source, validate completeness, normalize schema, apply GL and cost center mapping, refresh the semantic model, and publish to Power BI in a fixed sequence.
LessonAutomating steps is not the same as removing ambiguity. The pipeline was fast and still produced numbers people argued about, because the business logic had not yet been unified.
- V32026 · Q2Unified semantic modelOne schema: Actual, Plan, and LBE finally comparable.
Every record was forced into a shared structure — Cost Center, Business Lead, Function, GL Account, Category, Scenario, Period, Month, Amount. Scenario became a dimension rather than a separate report, which is what made Actual, Plan, and LBE directly comparable.
LessonComparability is a modeling decision, not a dashboard feature. Once scenario was a dimension, whole classes of reconciliation disputes disappeared.
- V42026 · Q2Financial health surfaceThree depths of focus, one decomposition.
Executive, Finance Analyst, and Operations views were built over the same model, with variance decomposed into volume, rate, mix, and timing. A rate variance and a timing variance start different conversations, so the surface refuses to present a delta without its decomposition.
LessonDecomposition changed the meeting. People stopped asking what the number was and started asking which component moved.
- V52026 · Q2–Q3Driver intelligenceConnect financial movement to the operations that caused it — and keep the residual.
Labor, throughput, utilization, cutoff, and mapping signals were joined to variance components so movement could be attributed to a driver. Crucially, whatever remained unattributed was preserved as an explicit unexplained residual rather than absorbed into the nearest plausible bucket.
LessonForcing full attribution manufactures false confidence. The residual is the most honest field on the page and the best pointer to the next investigation.
- V62026 · Q3Public research prototype · SyntheticFinancial Variance Command Center — a transparent investigation room.
Six bounded agents work a ranked portfolio of synthetic variance cases. Cases are editable, every claim is traceable end to end, a human disposition is required, findings export as JSON, and a counterfactual lab lets a reviewer test alternate assumptions before committing to a theory.
LessonMaking the investigation inspectable matters more than making it autonomous. The value is a defensible brief, not an unattended conclusion.
Six roles, each with a remit narrow enough to be checked.
No role is allowed to be a general assistant. Each holds one responsibility, emits one kind of claim, and can be inspected independently of the others. Disagreement between roles is preserved rather than averaged away.
Validates completeness, mapping integrity, and whether the record is fit to reason over at all.
Decomposes movement into volume, rate, mix, and timing using deterministic calculation only.
Connects financial movement to labor, throughput, and utilization signals from the operating floor.
Contests the outlook, proposes alternate baselines, and argues the case against the emerging theory.
Checks cutoff, accrual treatment, anomalies, and control exposure implied by the variance.
Assembles an accountable management brief: theory, evidence, residual, dissent, and open questions.
Standardized record → deterministic calculation → bounded interpretation → human authority.
The architecture exists to enforce one separation: arithmetic is never a matter of opinion, and interpretation is never mistaken for arithmetic. Every figure in the command center is produced by deterministic code over a standardized financial record. Agents may argue about meaning; they may not change the numbers.
- Layer 01Standardized financial recordOne schema across Actual, Plan, and LBE. Nothing enters the room without conforming to it.
- Layer 02Deterministic calculationsVolume, rate, mix, timing, and residual are computed in code. No agent can alter a figure.
- Layer 03Bounded agent rolesSix narrow remits interpret the computed record; each emits claims attributable to its role.
- Layer 04Shared variance bridgeA common bridge reconciles every claim back to the same decomposition, so disagreement is visible.
- Layer 05Human-reviewed briefA reviewer disposes of the case — accept, revise, escalate — and the disposition is part of the record.
Calculation is deterministic, interpretation is bounded, and authority stays human. The architecture is the argument.
What the build has actually taught so far.
- ObservationTraces get read when they are shortFull traceability is only useful if a reviewer will actually open it. Traces that opened at claim level with an expandable path were read; traces that opened as a full log were skipped.
- DecisionAgents may argue, never recomputeEvery figure comes from the deterministic layer. Agents annotate, contest, and interpret, but the arithmetic is a single source. This removes an entire class of disagreement about whether the numbers themselves are trustworthy.
- FailureAn unbounded 'analyst' agent produced fluent nonsenseAn early single general-purpose reviewer wrote confident narratives that were difficult to check because no part of the claim mapped to a role or a computation. Splitting the work into six bounded remits was a response to that failure, not a design flourish.
- ConstraintPublic release cannot touch production dataThe prototype is deliberately synthetic and browser-only. That constrains realism, and it is the correct trade: the research question is about the structure of investigation, not about any specific site's financials.
- ObservationDisposition is the missing artifact in most reportingDashboards record what happened; they rarely record what a human concluded and committed to. Requiring an explicit disposition turned the tool from a report into a record of accountable judgment.
Bounded challenge should raise explanation coverage without hiding residual.
The working hypothesis is that a structured investigation room — deterministic decomposition, an explicit challenger, an operations channel, and a mandatory trace — produces briefs that explain more of the variance, surface unexplained residual honestly, and reach a first human review faster than an unassisted analyst working from a dashboard.
The hypothesis is falsifiable in both directions. If the challenge role only adds noise, or if the trace slows review without changing conclusions, the design should lose those components rather than defend them.
Measures proposed, none yet empirical.
The following measures are proposed. No study has been run and no result is claimed. They are recorded here so the design can be judged against a standard it committed to in advance.
Share of total variance attributed to a stated, evidenced driver.
Magnitude preserved as explicitly unattributed after review.
Proportion of claims that resolve to an inspectable computation and source.
Elapsed time from case surfacing to a human opening and reading the brief.
Frequency with which reviewers revise or reject the synthesized theory.
How often the Forecast Challenger triggers a reconsidered outlook.
- AblationRemove challenge
Drop the Forecast Challenger and test whether coverage rises while override rate rises with it.
- AblationRemove operations context
Withhold labor, throughput, and utilization signals and measure residual growth.
- AblationRemove trace
Hide the inspectable path and observe the effect on review time and override behavior.
Boundaries stated plainly.
- The public release uses synthetic data and deterministic browser calculations.
- It has no live SAP or Power BI connection.
- It makes no external LLM calls.
- It posts no journal entries.
- It claims no measured organizational improvement.
Finance and operations retain decision authority. The system prepares an accountable brief; people dispose of it.
What comes after the synthetic release.
- StudyRun the proposed measures
Instrument the prototype so coverage, residual, and review time are recorded rather than asserted.
- ProtocolTyped claim schema
Formalize claims — statement, computation reference, evidence, confidence, dissent — across all six roles.
- CalibrationResidual honesty checks
Test whether reviewers trust systems that admit residual more than systems that force full attribution.
- BridgeFoundation-to-prototype path
Define what a governed, read-only connection between the reporting foundation and the investigation room would require.
One variance, walked from detection to disposition.
Direct link: Watch the narrated demo (MP4)
Everything published alongside System 008.
- Live prototype · ExternalFinancial Variance Command Center↗
- V6 technical report · PDFDownload technical report — Financial Variance Command Center↓
- Public repository · Externalgithub.com/fa366193/financial-variance-command-center↗
- Narrated demo · MP4Watch narrated demo (Entry 11)↗
- Reporting foundation report · PDFReporting foundation report — Site Financial Health Engine↓