Dr Reconcile

Data Consistency Reconciliation

Cross-check Finance OS data against itself using independently-sourced numbers — the same aggregate through two different API families, the balance-sheet identity, hierarchy roll-ups, and scenario/period checksums. Every check compares two numbers that arrive by different routes; nothing is ever compared against a re-read of the same aggregate (which would always "agree" and prove nothing).

Honest scope: these are data-pipeline & mapping consistency checks, not source-system reconciliation. A clean pass means the data in Finance OS is internally consistent across endpoints, grains, and slices — it does not prove the numbers match the ERP/GL. State this in the report output too.

Essential for month-end close and financial validation.

Excel Context — Routing Preamble

Before any data pull, establish whether this skill is running in a live Excel context (Claude for Excel with the Datarails Add-In loaded) and route accordingly.

Detect — never infer from the user's wording. A sheet list containing __dr_agent means the add-in is loaded. Confirm with the agent.get_session probe, which you run by executing Office.js through the execute_office_js tool (see the Excel Context Contract in CLAUDE.md, §Transport) — it is not an MCP tool and has no MCP equivalent. A failed probe is a normal detection result, not an error: it means "no bridge here", which is the expected outcome in Claude Code. Do not surface it, do not retry it, and do not apply this skill's connection-error or Connectors-UI guidance to it — that guidance is about datarails-finance-os connector calls only. A successful probe means Excel context, on either transport. The bridge serves two add-in tracks and their session payloads differ: Flex (Office.js task pane) exposes isLoggedIn; the COM desktop add-in — the majority of live workbooks — exposes isConnected instead, and isConnected: false is not a login failure, an error, or a reason to stop or send the user anywhere. It merely means the workbook isn't connected to a Datarails file, which matters only to drilldown_* / create_dynamic_range (the bridge skill gates those itself). Only Flex's explicit isLoggedIn: false means sign-in is needed.

Route by the target of the request, not by whether a workbook is open.

• Org / server data — which tables, models and fields exist, aggregations, raw rows, distinct values, metrics, profiling — always the datarails-finance-os MCP connector, even in Excel. The bridge cannot answer these.
• Workbook actions — refresh, drill a cell, insert a DR function, read what a cell returns, publish, submit — always the add-in bridge. Never a native Excel recalc (calculate(), F9): it does not pull Datarails data and silently yields stale values.

In a live Excel context this skill cannot produce its file deliverable. Its generation steps depend on the Bash tool, which that surface does not provide. Say so plainly and offer the real alternatives — a scoped answer in chat, or re-running this skill from Claude Code where file output works. Never improvise another route to a file, never hand back a partial artifact, and never silently substitute a different deliverable: writing into someone's live workbook instead of giving them the file they asked for is a different and irreversible outcome, not a smaller version of the same one.

If you do write DR formulas into the workbook, writing and refreshing are one atomic step. Write to a new sheet, fire refresh_selected_cells_ribbon scoped to that range — a new-sheet block is one contiguous range, so one scoped call covers any cell count — then read the range back. refresh_ribbon is not the tool for this: it repulls every DR cell in the file and can silently move numbers elsewhere in the user's model. It is reserved for the one case the scoped command can't cover — scattered inserts across multiple sheets, per excel-context__internal's refresh-after-insert rule — and even then only with the user's explicit OK, after snapshotting the DR ranges you can bound, reporting each changed cell in them before → after with the compared ranges named, and saying plainly that cells beyond them may also have updated. If the user declines the whole-workbook refresh, fall back to scoped refresh_selected_cells_ribbon calls sheet-by-sheet — slower, but nothing outside the written cells moves. A freshly written DR formula reads Missing / Loading… / #BUSY! until an agent refresh lands, so never quote a value you have not read back after a successful refresh, and never present a figure fetched from the MCP connector as though it were the cell's value. /dr-get-formula is the full authority for DR.GET workbooks.

If the user asks you to elaborate on a DR-backed figure — "explain", "break down", "what's driving this", "why is X" — and the figures in scope are DR formula cells, offer the add-in's drill-down instead of silently re-deriving the number through the MCP connector. A drill resolves the exact filters behind that cell; a hand-rebuilt query only approximates them.

How the checks source their numbers

This skill is self-contained — it discovers the client's financials table, fields, and account categories inline (every Datarails environment names them differently; there is no saved profile). Do discovery once per conversation, then carry the values forward.

1. Discover the financials table and fields. list_data_models → pick the table whose name/alias matches /financial|cube|p&?l|ledger|gl/i; if nothing matches, probe candidates with get_fields_by_id(<id>) and pick the one carrying amount/scenario/date-like fields (no tool returns row counts). Note both its numeric id and its alias (alias may be empty — prefer the alias path when present). Then list_aliased_fields(<alias>) if it has an alias, else get_fields_by_id(<id>) (capture each field's numeric id). Bind by case-insensitive match: amount (^amount$ → transaction_amount → value), scenario (^scenario$ → ^version$), date (reporting_date → posting_date → ^date$), and every account-hierarchy level field present (names matching /acc(ount)?.*l\d|account_group/i, or any field whose distinct values look like account groupings) — the checks below need at least two adjacent levels. Also bind an entity-like field when one exists (/entity|company|subsidiar|legal/i) for per-entity slicing. If a binding can't be resolved by name (e.g., non-English field names), list the discovered fields and ask the user which field plays which role.

Alias coverage is per field, not per table. A table having an alias does not mean its fields are aliased — real orgs often expose only a handful of aliased fields (e.g. ~5 of ~185 on a mapped financials table), and the load-bearing fields (amount, scenario, account groups, dates) are frequently not among them. Treat the alias/by-id choice per field: get_fields_by_id(<id>) returns every field with its numeric id and its alias (empty if none). Address a field by alias (via the *_by_alias tools) when it has one, else by numeric id (via the *_by_id tools). By-id always works — never abandon the query because the aliased set is thin.

Async fetch — aggregations and distinct values run as start → poll. start_aggregation_by_id/_by_alias and start_distinct_values_by_id/_by_alias take the same arguments as the retired blocking calls (dimensions/metrics/filters; table id + field id, or alias + field alias) and return immediately with {"status": "pending", "handle": {...}}. Echo that handle back verbatim to the matching get_aggregation_result_by_* / get_distinct_values_result_by_* tool: a {"status": "running", "retry_after_seconds": N} response means poll again with the same handle after ~N seconds (≈5s) — it is not an error, and large jobs may take several polls; when ready, the result arrives in the familiar shape (for distinct values, pass limit to the result tool). An expired/unknown-handle error means restart with the start_* tool. Transitional fallback: if the start_* tools aren't available on the connector (older server), the blocking twins get_aggregated_data_by_* / get_distinct_values_by_* still work with the same arguments.

Data-scope discovery — run before any aggregate (reuse anything already discovered this conversation).

1. Scenario domain. Pull distinct values of the scenario field (start_distinct_values_by_alias/_by_id → poll the matching result tool) — never assume a scenario name exists (Budget frequently doesn't; many orgs carry only {Actuals, Forecast}). For budget/plan questions, if no budget-like scenario exists, look for a planning-version-like field (alias/name matching /plan|version|cycle|budget/i) and use its versions as the plan side; if neither exists, say so and offer a comparison across the scenarios that do exist.
2. Account grain. Pull distinct values of each account-hierarchy level field (L0/L1/L2-like). Use the level whose values partition P&L flows into revenue/COGS/opex-like buckets — on many orgs the top level is the balance-sheet equation (ASSET/LIABILITY/EQUITY/INCOME) and P&L line items live one level deeper. For P&L work, scope to P&L flows and exclude balance-sheet buckets; never present asset/liability/equity totals as revenue or expenses.
3. Period scope. Discover the date field's range (distinct values of the reporting-month field, or MIN and MAX in two separate calls — one aggregation per field per call). Default every P&L question to the latest complete fiscal year (or trailing 12 closed months) — never an unscoped all-time total: financials tables are multi-year cumulative and mix balance-sheet stock with P&L flow. Label every output with the period + scenario it covers.
4. Reading GROUP BY responses. Each response returns exactly one row per requested group — no subtotal rows and no grand-total row mixed into the data list; grand totals arrive in a separate top-level totals field beside the rows ({"data": [...], "totals": {...}}), computed across all groups, not just the returned prefix. For a grand total, read totals — never sum the rows when the response carries truncated: true (summing the returned prefix silently under-counts; dev repro: 474 of 31,455 rows summed to 21% of the true total). totals combines the per-group results rather than re-scanning the rows, so it is exact exactly when the aggregation is decomposable: SUM (sum of the group sums), COUNT (sum of the group counts), MIN, and MAX. It is WRONG for AVG (unweighted mean of the group averages) and COUNT_UNIQUE (sum of the per-group distinct counts, so a value recurring across groups is counted once per group) — true average = SUM total ÷ COUNT total (two calls: a field may be aggregated at most once per request); true distinct count = the distinct-values tools. Treat every aggregation type not named exact above — UNIQUE_VALUES included, whose cross-group de-duplication is unverified (the COUNT_UNIQUE behaviour above is evidence the engine may not de-duplicate across groups at all) — as not decomposable: derive it from complete rows or the distinct-values tools, never from totals. totals is absent on dimension-less aggregations (the single returned row IS the total) and may be absent on responses cached before the rollout (cache TTL ≤ 7 days) — only in those two cases is a total obtained by summing complete (untruncated) rows. Null groups arrive explicitly labeled [null] and are real groups; read null counts from that bucket. Defensive filter: keep only rows in which every requested dimension key is present — a roll-up row omits one or more keys entirely, whereas a genuine null is present with the value [null]. On a correct response this is a no-op; it guards against a stale cached response still carrying legacy subtotal and grand-total rows, each of which equals the whole total and would inflate any sum. When COUNT-ing rows per group, aggregate a different field than the GROUP BY dimension itself — a same-field COUNT of the grouped dimension can 500.
5. Truncated results. Any data tool may return {"data": [...], "truncated": true, "total_rows": N, "returned_rows": M, "guidance": "..."} when the result exceeds the response size limit (~50 KB). The data prefix is incomplete — never compute totals, shares, or trends from it, and never present it as the full result. On aggregations the top-level totals field is unaffected by truncation (computed across all groups, not just the returned prefix) — read grand totals from it instead of re-fetching. Narrow the query (fewer dimensions, more filters, fewer selected columns — or a business metric for a named KPI) and re-fetch only when the rows themselves are needed beyond the cap; with totals present, a SUM/COUNT/MIN/MAX grand total never requires a re-fetch or chunking by dimension (AVG, COUNT_UNIQUE and UNIQUE_VALUES never read totals — true average = SUM total ÷ COUNT total from two calls; true distinct count = the distinct-values tools). A truncated response without totals (pre-rollout cache) cannot answer a grand-total question from its prefix. Re-run the aggregation once — a fresh run may miss the stale entry and return totals. If the re-run still carries no totals, stop re-running and fall back to narrowing or chunking by dimension until the responses are complete, then sum those rows. Never total the prefix.

2. Map the account grains. From the per-level distinct values pulled in the data-scope preamble (item 2) — if the distinct-values start→poll pair (start_distinct_values_by_alias/_by_id → get_distinct_values_result_by_*) errors, fall back to get_data_by_alias(<alias>, select=[<account_field>], limit=500) (or the by-id twin) and dedupe — identify two grains: the balance-sheet grain is the level whose values look like the balance-sheet equation (/asset|liabilit|equity/i buckets); the P&L grain is the level whose values partition P&L flows into revenue/COGS/opex-like buckets (/revenue|sales|income/i, /cogs|cost of goods|cost of sales|direct cost/i, /operating|opex|expense|sg&a/i) — often one level deeper. Then discover the parent↔child mapping between the two adjacent levels with one aggregate (start_aggregation_by_alias/_by_id → poll the matching result tool until ready): dimensions=[<parent_level>, <child_level>], metric COUNT on a different dense field (a same-field COUNT of a grouped dimension can 500). Every child bucket should map to exactly one parent; note any that map to several parents or to [null].

The checks — four independent-source comparisons

All checks run under the same scope: the --year window as an advanced date filter — {"name": <date_field>, "values": {"type": "advanced", "val": [{"condition": "total_range", "value": ["<jan1_epoch>", "<dec31_epoch>"]}]}} (epoch seconds as strings; the by-id twin takes field_id instead of name) — and, for checks 1–3, one scenario at a time from the discovered scenario domain. Label every reported number with its period + scenario.

3. Check 1 — Cross-endpoint agreement. Run the same aggregate through both API families over the same table and compare per bucket to the cent: the alias pair start_aggregation_by_alias(<alias>, dimensions=[<account_field>], metrics=[{"field": <amount_field>, "agg": "SUM"}], filters=[<scenario>, <date range>]) → poll get_aggregation_result_by_alias(handle) until ready, vs the by-id pair start_aggregation_by_id(<id>, dimensions=[<account_field_id>], metrics=[{"field_id": <amount_field_id>, "agg": "SUM"}], filters=[...]) → poll get_aggregation_result_by_id(handle) until ready. This proves something only because the two pairs travel different endpoint families (aliased vs raw) — and it is only possible for fields that carry both an alias and a numeric id (per-field rule above). If the load-bearing fields (amount, scenario, account level, date) are not all aliased, skip this check and note the skip in the report — never fake it by running the by-id pair twice.

4. Check 2 — Balance-sheet identity. At the discovered balance-sheet grain: dimensions=[<bs_level>, <period_field>] (add the entity-like field as a third dimension when one exists), SUM(<amount>) — via start_aggregation_by_alias/_by_id → poll the matching result tool until ready (async-fetch pattern). Per period (and entity), compare magnitudes: |asset-like| vs |liability-like + equity-like|, and report the org's sign convention as discovered from the data (e.g. whether liability/equity buckets arrive negative) — never assumed. Balance-sheet values are stock, so evaluate per period; never sum them across periods. If income/P&L-like buckets also live at this grain, exclude them from the identity, and note that it only closes exactly when the model rolls P&L into equity — if it visibly doesn't, report |A| vs |L + E + cumulative P&L| and say which form was used.

5. Check 3 — Cross-grain consistency. The P&L-grain buckets must sum to their parent bucket at the level above (mapping from step 2). Two aggregates over the same scope (each via start_aggregation_by_* → poll get_aggregation_result_by_* until ready): dimensions=[<parent_level>] with SUM(<amount>), and dimensions=[<parent_level>, <child_level>] with SUM(<amount>). For each parent, the sum of its child rows — including the [null] bucket — must equal the parent's own row to the cent. A mismatch means the hierarchy mapping leaks rows (unmapped or double-mapped accounts).

6. Check 4 — Scenario/period integrity. Two grouped aggregates over the scoped window (each via start_aggregation_by_* → poll get_aggregation_result_by_* until ready), no scenario filter on either: (a) dimensions=[<scenario_field>], (b) dimensions=[<period_field>]. First apply the item-4 defensive filter — keep only rows in which every requested dimension key is present (a stale cached response can carry legacy roll-up rows that omit keys and equal the whole total, which would fake a mismatch), preserving genuine [null] groups, then sum those rows to get that slicing's total — this check's rows must be summed client-side, and the responses must be complete: on truncated: true, narrow (e.g. split the window into sub-ranges) and stitch, never sum a prefix. The two totals must equal each other to the cent, since both describe the same unfiltered window sliced two different ways. A mismatch means rows are escaping one of the groupings (bad scenario/date values) or the two slicings disagree about what is in scope.

   The independence of this check comes from the two different slicings, not from any server-provided total. For exactly that reason, do not read the top-level totals field for this check: the engine computes totals over the filtered window independently of the grouping, so both slicings would return the same engine number by construction and the check would pass vacuously (verified live: totals is byte-identical across groupings of the same window).

Arguments

What It Validates

Every check compares two independently-sourced numbers — two different endpoint families, two different grains, or the same window sliced by two different dimensions. Nothing is compared against a copy of itself.

1. Cross-Endpoint Agreement

• Same aggregate via the aliased and the raw by-id API families
• Per-bucket match to the cent
• Skipped (with a note) when per-field alias coverage is too thin

2. Balance-Sheet Identity

• |Assets| vs |Liabilities + Equity| per period (and entity)
• Sign convention reported as discovered, never assumed
• Evaluated at the discovered balance-sheet grain

3. Cross-Grain Consistency

• P&L-grain buckets roll up exactly to their parent bucket
• Flags unmapped / double-mapped accounts (including [null] buckets)

4. Scenario/Period Integrity

• Scenario-sliced and period-sliced totals (each the sum of its own rows) agree with each other to the cent

Not validated: agreement with source systems (ERP/GL), completeness of the load, or business-metric engine values — the get_business_metric_* data tools are feature-gated and may be absent; list_business_metrics (ungated) may be used only to note in the report which named KPIs exist, never to fetch values.

Datarails Brand Styling

When generating Excel or PowerPoint files, apply Datarails brand styling:

Font: Poppins (fall back to Calibri if unavailable). Weights: 400 regular, 600 semibold, 700 bold.

Colors:

Excel layout:

• Content starts at column B (column A is a narrow gutter)
• Rows 1-6: header banner with navy background, white title text, white subtitle
• Gridlines OFF. Freeze panes at B7.
• Footer as last row with generation date
• Every cell must have font, fill, alignment, and number format set

Number formats: _(* #,##0_);_(* (#,##0);_(* "-"_);_(@_) (default), $#,##0 (dollars), $#,##0.0,,"M" (millions), 0.0% (percent)

Variance coloring: Any cell showing a delta/change: green (2ECC71) if favorable, red (E74C3C) if unfavorable. Apply automatically based on value sign and metric context.

PowerPoint: Navy (0C142B) background, 16:9 widescreen, Poppins font, white text, amber (FFA30F) accent lines, card backgrounds 001F37.

DR.GET Formulas — Authoring Contract

If asked to add live / refreshable Datarails formulas (DR.GET) to a generated workbook, the only valid form is:

=DR.GET(Value, "[DimensionName]", CellRef, "[DimensionName]", CellRef, ...)

• Never transliterate an MCP/API call into a formula. DR.GET takes no table, field, or aggregation arguments — =DR.GET(Value,"financials","Amount","SUM",...) is invented syntax that the Datarails Add-in cannot parse or refresh.
• Dimension names go in square brackets inside quotes ("[Scenario]"). Dimension values are always cell references, never hardcoded strings.
• Date cells referenced by formulas hold end-of-month date serials computed from the calendar — never raw epoch timestamps from API responses (epochs land a day early with a time component and never match).
• Before writing any formula, create the workbook-scoped defined name Value referring to the string constant "Value" (wb.defined_names.add(DefinedName("Value", attr_text='"Value"'))) — otherwise Excel autocorrects the bare token to its built-in VALUE() and the formula breaks.
• Bare =DR.GET(...) only — never wrapped in IFERROR/IF/ROUND.
• Every rule here applies to the retrieval/period family — DR.GET, DR.QTD, DR.YTD, DR.MTD share one form (=DR.QTD(Value, "[Dim]", CellRef, ...)), one cell-reference discipline, one Value defined-name requirement, one no-wrapping rule. "DR.GET" in this contract means that family. Helper functions with their own documented signatures (e.g. DR.INCLUDE, DR.RANGE) are not covered here — author those only from their own documentation, never by analogy with this form.
• In a live Excel context, writing DR formulas and refreshing them is one atomic step — a freshly written DR cell reads Missing until an agent refresh lands, and only read-back values may be quoted. The Excel-context routing preamble (or the skill's own Step 0 workflow) owns that procedure; this contract owns the formula text.

The get-formula skill (/dr-get-formula) is the full reference — parameter cells, validated dimension values, report layouts. Prefer it for whole formula workbooks; apply this contract when adding any retrieval/period DR formula (DR.GET/DR.QTD/DR.YTD/DR.MTD) to a workbook here.

Output

Excel report with multiple sheets:

1. Summary - Pass/fail/skipped per check, exception count, and the scope statement: "Data-pipeline & mapping consistency checks over <period> / <scenario> — not a source-system reconciliation." Every figure labeled with its period + scenario.
2. Check 1 - Endpoint Agreement - alias vs by-id value per bucket, delta (or the skip note when alias coverage was too thin)
3. Check 2 - Balance Sheet - per-period |A| vs |L+E|, sign-convention note
4. Check 3 - Roll-Up - parent totals vs child-bucket sums per parent
5. Check 4 - Integrity - scenario-sliced total vs period-sliced total
6. Exceptions (if any) - deltas exceeding each check's threshold

Examples

Reconcile current year (default 5% tolerance)

/dr-reconcile --year 2025

Strict reconciliation (1% tolerance)

/dr-reconcile --year 2025 --tolerance-pct 1.0

Reconcile specific scenario

/dr-reconcile --year 2025 --scenario Forecast

Custom output location

/dr-reconcile --year 2025 --output reports/reconciliation_2025.xlsx

Use Cases

Month-End Close

Run after data extraction to validate:

/dr-extract --year 2025
/dr-anomalies-report --severity critical    # Check quality
/dr-reconcile --year 2025 --tolerance-pct 2 # Validate consistency

Financial Review

Reconcile before presentations:

/dr-reconcile --year 2025 --scenario Actuals

Audit Preparation

Reconcile with strict tolerance:

/dr-reconcile --year 2025 --tolerance-pct 0.5

Performance

• Year reconciliation: ~30-60 seconds
• Runs 4 independent-source checks (check 1 may be skipped on thin alias coverage)
• Scalable to large data volumes

Error Handling

"Account grain not found" - Re-run discovery (list_data_models → list_aliased_fields/get_fields_by_id → start_distinct_values_by_alias/start_distinct_values_by_id → poll the matching get_distinct_values_result_by_*) and re-derive the balance-sheet and P&L grains from the discovered level values. There is no profile — discovery happens inline.

422 on aggregation - At most one aggregation per field per call (split SUM and AVG of the same field into separate calls); SUM/AVG require a numeric or date field.

500 on COUNT - A same-field COUNT of the GROUP BY dimension can 500 — aggregate a different dense field instead (data-scope preamble, item 4).

"Variance exceeds tolerance" - Review the exception sheet; re-check the discovered sign convention before treating a balance-sheet delta as real.

"Incomplete data" - Run /dr-extract to refresh data first

Related Skills

• /dr-extract - Get latest financial data
• /dr-anomalies-report - Check data quality
• /dr-dashboard - Verify KPI values
• /dr-insights - Understand trends driving reconciliation items