Clearly unfinished first messages and creation requests without a purpose ask for clarification before file processing. Short contextual replies remain supported.
Submission locking and keyboard composition handling protect against repeated clicks and premature Enter. Failures restore instructions without overwriting a new draft.
Saved task receipts resume the same task. Expired or unconfirmed tasks offer instruction recovery instead of an automatic restart.
Connection and access errors have localized next steps. Reload recovery requires opt-in device storage; independent regression checks cover interrupted transport and saved sessions.
SUBTOTAL modes 1â11 and 101â111 support bounded A1 and structured table references, hidden rows, saved filters and nested subtotals.
Uploaded XLSX files receive a local edit preflight. Unsupported formulas and calculation errors show their sheet and cell before AI planning.
Uncertain hidden/filter distinctions remain blocked. Editing active filter criteria requires clearing the filter in Excel and uploading again.
Cell preview and approval remain required; version rollback stays available. Compatibility checks cover supported editing, not every Excel feature or business interpretation.
Bounded five-language routing keeps supported questions, how-to requests and negated commands out of controlled editing. Arbitrary natural-language intent is not certified.
Header candidates scan all imported rows within existing limits. Physical header and last-row selection require renewed scope approval before planning.
Excel table totals and recognized footer rows are shown outside suggested ranges. Gaps, repeated blocks and ambiguous total labels require an explicit last row.
Timeouts, interrupted connections, busy services and unsupported edits have localized next steps. Retry is manual and respects the returned waiting period.
Synthetic routing, layout and failure-path checks include late headers and independent saved-file recalculation. Desktop/mobile browser interaction remains unverified.
AI plans bounded edits from the instruction, selected headers and row boundaries; workbook records stay on the device in this workflow.
Cell-level before/after review includes dependent recalculations. Approval is bound to source, plan and output hashes; other parts and chart definitions are preserved.
Version comparisons follow the actual parent after a rollback or branch. Original downloads and the four retained versions remain available; device persistence is opt-in.
Up to 1,000 target cells and 5,000 formulas per edit. Unsupported formulas, complex structures, row/column shifts and table expansion are blocked or require clarification.
16 synthetic regression groups and 28 independent saved-file/native calculation assertions passed. Five-language controls; desktop/mobile browser interaction remains unverified.
Bounded Turkish/English request parsing checks literal cell values, types, coordinates and required live formulas against the saved file. Unrecognized prose is not certified.
Sums of products are verified from aligned raw input pairs, including zero and negative operands, with first/last row, blank and table-growth probes.
Added bounded CHOOSE, IFS, ROWS, COLUMNS and SIGN evaluation plus literal COUNTA arguments. Explicit empty operators over formula-produced empty text remain unverified.
Synthetic regression tests retain rejected originals, diagnostics and corrected outputs. Customer workbooks are not automatically archived for testing.
Round 1 totals distinguish unsold allocation from a verified burn. A successful receipt must contain matching finalization and ARXAI burn events; current supply is read separately.
Five-language round panel with confirmed-block timestamps, transaction links and explicit unavailable/stale states. Desktop/mobile browser interaction remains unverified.
Repeated invoice IDs and mixed currencies in recognized tables are flagged before AI analysis; unresolved business choices require clarification.
Approved file, sheet, period, column and record selections are checked by the server and browser; changed scopes with equal totals cannot authorize delivery.
COUNTBLANK and exact empty-string criteria recalculate supported formula dependencies. Explicit = and <> criteria on formula-produced empty text remain unverified.
Application regression checks run before the production build and stop it on failure. These checks do not create a new live AI accuracy score.
Header detection is bounded and does not certify arbitrary intent. Desktop and mobile browser interaction remains unverified for this release.
Review one original XLSX/CSV sheet, numeric metric, optional category/date/currency fields, period and exact category filter before running the package
Nine-sheet report includes all selected source rows, numeric summaries, groups, months, highest/lowest records, IQR review flags, source hashes and native Excel charts
Blank values stay distinct from zero; exact duplicate rows are retained and flagged. Invalid dates with active period filters, selected mixed currencies and unsupported source formulas block delivery
Original source bytes and reviewed settings are embedded for deterministic replay; every saved workbook part, including chart ranges and caches, must match before delivery
Optional AI explanation must return all 13 metrics exactly; one bounded correction is allowed, and persistent mismatches or changed report hashes withhold delivery
Reusable settings require identical sheet headers and renewed scope approval; controls support English, Turkish, German, Spanish and French
Four live application client/API scenarios passed: quality and outliers, recalculated formulas, empty numeric values and approved month/category filters; all 52 metric claims matched
22 synthetic package groups and 23,124 independent Python comparisons passed, including complete export/replay of 10,000 records; existing settlement, reviewed explanation and connection tests also passed
One original source up to 2 MiB, 10,000 selected rows, 50 columns and 100,000 source cells; output limits also apply. Floating-point statistics use 15 significant digits and are not exact-decimal accounting
The report embeds the entire original source. AI explanation sends both to the external AI service. Desktop/mobile browser interaction and native Office rendering remain unverified
Uploaded XLSX source formulas and filter cells now use the bounded dependency calculator, with stale-cache counts and dependency audits; unsupported formulas and errors withhold source answers
Additional IFNA, error/type predicates and conservative text functions; original COUNT/COUNTA/record semantics stay distinct
A reviewed primary analysis scope pins the original source hash, metric, ranges, exclusions, operation and up to six filters; a correct total from the wrong scope is withheld
Advanced settlement supports up to three exact typed key columns, explicit positive payment/refund/credit records and user-approved dated conversion rates
Per-record integer conversion with half-away-from-zero rounding; credits beyond invoices, refunds beyond selected receipts and unknown rates require review
Original amounts, currencies, rates, transaction types and source addresses remain in the replayable report; net invoices/receipts and gross adjustments are separate
Five-language controls. These checks do not certify arbitrary natural-language intent, unsupported Excel formulas, live FX rates or accounting completeness. Browser/mobile interaction remains unverified
User-reviewed invoice and payment tables from XLSX/CSV sources or separate sheets; up to 10,000 selected records with exact typed, case-sensitive IDs
Multiple unique payment IDs may reference one invoice; duplicate invoices or payments, missing amounts, source formulas, negative amounts and currency differences stay under review with all affected records
Explicit single currency and independent input number formats; integer minor-unit arithmetic with 0â6 decimal places, no source rounding or offsetting open balances against overpayments
Ten-sheet Excel report separates closed, partial, unpaid, overpaid, unmatched and review records; the payment ledger, selected source snapshots, file hashes and original cell references remain available
Exports re-read original files and reopen the saved report; every row, amount, classification and source reference must agree before delivery
Optional AI explanation returns 11 required metrics checked against a replay of the exact report; one bounded repair, with unresolved or mismatched responses withheld
English, Turkish, German, Spanish and French controls; source or setting changes invalidate the previous result
27 new synthetic test groups and 142 independent Python Decimal, SQLite and saved-XLSX comparisons passed
The initial scope does not allocate refunds, convert currencies, evaluate source formulas, infer due dates or certify accounting completeness; desktop/mobile browser interaction remains unverified
v3.25.0IA AREXAI15 oct. 2026
PUBLIĂ
Comptages explicites et contrĂŽles conditionnels des sources
Record counts and distinct-customer counts preserve repeated transactions, real zero, leading-zero IDs and exact text matching
Up to six conditions are reapplied to original source rows, including numeric comparisons, status/customer matches and calendar-day boundaries
ISO dates and XLSX 1900/1904 calendars are supported; ambiguous dates and formula-based conditions remain unverified
Source cards show included, filtered, excluded and ignored addresses; incomplete scope inspection is clearly marked
Linked saved output cells must match independently recomputed original-source results before delivery; unsupported or mismatched bindings withhold the file
One unresolved business choice is asked before calculation; bounded repairs cannot discard declared conditions or output bindings
31 new synthetic verification groups and 40 independent Python Decimal/date/scope comparisons passed; the retained browser fixture now passes all three source counts/totals and all three output bindings
Checks apply to declared scopes and supported operations, not arbitrary business rules, joins or untested mobile behavior
v3.24.1IA AREXAI15 oct. 2026
PUBLIĂ
Corrections de lâaperçu Excel et de lâexport source
Separate numerator and denominator scopes support sums, numeric counts, record counts, equality filters and multiple worksheets
Weighted means include only numeric value/weight pairs, preserving real zero while excluding the weight of a missing value
Blank or explicitly requested zero fallback is checked when the denominator is empty or zero
Bounded table probes append and delete rows and reorder columns; fixed summaries that miss new records are blocked
Automatic repair must preserve the original metric definitions and their zero-denominator policy
Verification cards show checked ratio and weighted-average counts in five languages
22 new metric test groups and 80 independent Python Decimal checks passed; 42 existing table groups and 2,264 retained office-oracle comparisons passed
Live fixed-range output produced correct rate and weighted mean values; automatic metadata verified only the rate. Explicit raw-source replay checked both metrics across 20 probes. Table-based live acceptance remains incomplete
Negative weights, unsupported formula syntax and unchecked business assumptions remain unverified; browser and mobile interaction are not covered by these checks
Counts, comparison-entry percentages, individual differences and per-comparison sums are recalculated from the exact reviewed workbook before display
A strict model response selects bounded findings; the app supplies the displayed values and labels instead of treating free-form prose as verified evidence
âHow was this calculated?â shows the calculation, included entry count, report cells and original source addresses; downloadable JSON retains every reference
Decimal calculations preserve blanks versus real zero, selected numeric formats and leading-zero record IDs; percentages disclose their denominator and six-decimal rounding
Sums stay within one status group and one source pair; duplicate or ambiguous review records are never silently aggregated
The original reviewed file hash is bound to the signed task; changed source bytes and unsupported or incorrect claims block delivery
One bounded repair must preserve all original claim IDs, including when waiting is paused and resumed
English, Turkish, German, Spanish and French finding cards; causal comments and suggestions are explicitly marked as unverified
18 new synthetic verification groups and 57 independent Python Decimal checks across inventory, invoice-payment and personnel-call fixtures passed
Real browser/mobile interaction, actual IndexedDB reloads and an anonymized customer pilot remain pending; these checks do not certify arbitrary AI prose, business causes or tax compliance
Queued, working, checking and reconnecting states show elapsed time and the last server update; delayed work is disclosed without a made-up progress percentage
Transient poll and download failures retry the same task twice; pause and resume retain the existing token within its 20-minute session
Report source addresses are checked against the original uploaded files and exact worksheet names; a repair cannot remove a rejected citation to bypass validation
Historical references copied from an earlier report are retained and labelled separately, without claiming the original sources were reverified
Reviewed comparisons produce a deterministic six-sheet workbook with summary, four record groups and complete source scope; every saved row and value is checked before download
AI explanations read the complete reviewed workbook; four summary totals must match independent checks, and delivery preserves the exact reviewed file bytes
13 new delivery checks, 28 workspace checks, 13 recovery checks, 16 evidence checks and 24 workbook-agent checks passed; an independent Python oracle verified all 55 synthetic source records
Desktop/mobile browser interaction and real IndexedDB reload tests remain unavailable; arbitrary AI-generated reports and qualitative recommendations are not certified by reviewed matching checks
v3.21.1IA AREXAI14 oct. 2026
PUBLIĂ
Reprise des sessions IA et contrĂŽles des processus
Pending tasks block source replacement, rule changes and version restoration until they are resumed or discarded, including uploads from local mode
Temporary request failures retain the existing task checkpoint; recovery polls the same task without a new upload or task start
Late checkpoints and answers from discarded tasks cannot overwrite a newer workspace task
Device-saving transitions share one ordered write queue; blocked consent storage is reported without claiming saving is enabled
AI file limits are enforced before updating saved workspace sources, including files brought in from local analysis
13 session and recovery regression groups plus 28 existing workspace groups passed; three synthetic business scenarios account for all 55 source records and 63 pairwise source occurrences
Independent Python Decimal and XLSX checks compare every reviewed report row with original source records, including leading-zero IDs, duplicates, blanks, credit amounts and exact tolerance boundaries
Live staff and clarified invoice reports passed independent record-level checks; the original invoice source-label failure and a timed-out three-source report remain recorded
A separate three-source recovery replay used one task start and independently matched totals 66 / 63 / 59; an earlier source-evidence assertion failed, so consistent live reliability is not claimed
Desktop/mobile browser interactions and actual IndexedDB reloads remain unverified because the supported browser test tool was unavailable; synthetic fixtures are not customer pilot data
Travaillez avec trois sources Excel/CSV au maximum, examinez les correspondances et rĂšgles, et comparez les versions. Le stockage facultatif sur lâappareil permet de reprendre une tĂąche existante.
Up to three complete XLSX/CSV inputs per AI task, 6 MiB each and 12 MiB combined; source checks identify the exact original file as well as the worksheet and cells
Reviewed one-to-one comparisons separate matched, differing, unmatched, duplicate, blank-key and formula-dependent records; no fuzzy joins, silent aggregation or currency conversion
Up to 10,000 selected source records per reviewed comparison, with explicit sheet/header/end-row/key/value choices and downloadable Excel plus source evidence
User-approved business rules accompany file-agent requests; imported rules remain disabled until reviewed
Latest four output files can be selected as the next working copy; cell values, formulas and cell styles can be compared with disclosed visual limits
Optional browser-device storage retains sources, rules, the last answer and recent outputs for recovery; active jobs use their existing signed token within a 20-minute session
Portable import profiles reuse field mappings and settings for business-system Excel/CSV exports; they contain no source rows or credentials and require a new comparison approval
Independent formula checks add bounded rounding, weighted sums, exact lookups and numeric date/time functions, including 1900 and 1904 workbook date systems
28 new workspace regression groups and 127 expanded calculation comparisons/boundary checks passed; existing source, formula and table suites remain green
Two synthetic live AI tasks passed: three-source rule totals with recovery of the same provider job, and a five-sheet reconciliation file independently checked against all 15 source rows; browser reload and visual behavior were not tested
Live ERP connections, tax-law certification, chart/pivot comparison and automatic fuzzy matching are outside this release
Supported Excel structured references now resolve saved table and column names, including current-row calculated columns and explicit data/header/total selectors
Declared raw input checks can follow named table columns instead of fixed row ranges
Temporary scenarios append one or three rows beyond the saved table, remove a seeded first row and swap input columns; first/last, zero, blank and category probes remain active
Incorrect supported calculations enter the existing single repair attempt; the original table and column check scope remains anchored across repair
Explicit automatic-growth requests check for saved Excel table metadata; existing fixed-layout requests retain their scope
Five-language result cards distinguish table formulas, passed growth scenarios and checks that could not be verified
2,264 table-formula results across 21 fixtures match independently constructed A1 references recalculated by LibreOffice; this does not certify native Excel editing behavior
Live sales, inventory, accounting and call-center outputs pass 12 key calculations and 100 temporary scenarios; 60 summary calculations also match raw inputs across 20 separately saved edited copies
Four blocked synthetic attempts remain recorded. Invalid control JSON led to provider-constrained output schemas; the final inventory replay passed without file repair; strict validation and independent calculation gates remain enforced
Checks are bounded to supported formulas and declared raw inputs. Table totals-row relocation, general Excel edit rewriting, chart/pivot expansion and validation-range expansion remain unverified
Supported saved formulas are evaluated without trusting AI claims or cached values; unsupported syntax and exhausted budgets remain visibly unverified
Declared key totals are independently reduced from literal input ranges; this checks arithmetic within the declared scope, not the interpretation of every business rule
Temporary first-row, last-row, zero, blank and category changes exercise saved formula dependencies without changing the downloaded file
Explicit supported worksheet counts, quoted sheet names, required header lists and requested dropdown presence are checked against the saved workbook
Calculation, scenario and structural failures enter the existing single repair attempt; unresolved failures block download
Repair cannot evade a failed check by omitting or changing its original declared target and source scope
Five-language result cards distinguish evaluated formulas, matched calculations, passed scenarios, unavailable checks and stale saved results
48 basic formula results and 139 formulas in a populated 23-sheet workbook match independent LibreOffice recalculation; 3 key reductions and 28 mutation scenarios pass on that workbook
Two real AI runs delivered a new workbook and a preserved 23-sheet revision without repair: 142 saved formulas evaluated, 6 key calculations matched and 41 temporary scenarios passed; downloaded copies received independent file and LibreOffice checks
Advanced formulas, named and structured references, volatile functions, some empty-string criteria and ambiguous request requirements remain outside the supported verification scope
v3.18.2IA AREXAI14 oct. 2026
PUBLIĂ
Classeurs plus grands, navigation interne et erreurs claires
A shared 100-sheet AI limit covers input, generated output and flexible preview; dashboards and helper sheets count toward the total
20 representative worksheets plus dashboard, management and lookup sheets pass deterministic file and preview checks
Native internal links and literal HYPERLINK targets are checked against existing worksheet names and valid cell addresses
External and computed link destinations remain blocked; byte, cell, row and archive limits still apply
Specific output failure reasons and opaque error references survive the single automatic file-repair attempt
Previous-file availability is mentioned only when this conversation actually contains an earlier generated workbook
15 new regression groups cover sheet limits, navigation, blocked destinations, error reporting and chart placement near navigation links
A live replay produced 23 worksheets after one status clarification; 111 independent checks passed, while a visual chart/link overlap informed a new bounded repair check
Structural and label-width heuristics do not certify every formula, business rule or spreadsheet rendering
v3.18.1IA AREXAI14 oct. 2026
PUBLIĂ
Couverture complĂšte des enregistrements et placement des graphiques
For explicit full-record helper requests with one unambiguous ID table, original source cells and the complete ordered record-ID sequence are fingerprinted from the upload
Missing, duplicated or reordered helper records and changed source cells enter the bounded repair flow; unresolved faults block download
Supported explicit chart rectangles are checked alongside the print area, including both chart corners
Value-axis titles that incorrectly repeat the category source heading are detected in supported bar charts
Ten new regression groups cover historical failures, valid localized headers, unsupported formulas, source changes and signed server-side enforcement
Six unchanged synthetic sector requests delivered six workbooks, with 30 of 30 summary values independently recalculated correctly; three used automatic file repair
A separate accounting-chart replay passed all five summary calculations and visual review after the axis-title check was strengthened; retained failed outputs remain in the report
These bounded checks do not certify arbitrary Excel formulas, ambiguous layouts, all languages or visual rendering
Source mismatches receive one reinspection; corrected evidence is recalculated against the original uploaded file
Previously failing calculations cannot bypass checks by disappearing from the repaired answer; unresolved mismatches remain blocked
Generated workbooks are screened for direct blank-cell references counted as numeric data, charts outside a single print area and numeric limits on category axes
Detected output faults enter the existing bounded file-repair loop; originals are preserved
Incomplete runs and eligible provider server failures receive one bounded continuation in the same workspace; completed outputs still undergo source and file checks
Instructions now separate record-weighted means from group means, preserve blanks and real zero, and verify saved chart extents and value axes
16 regression groups cover historical failure files, repair paths, bounded continuation and blocked bypasses
Repeated live tests produced workbooks in six sectors; 29 of 30 latest summary values matched independent recalculation, with a retail row-coverage defect and chart-label/layout issues still open
These targeted checks do not certify all formulas, business rules or Excel rendering
Shared CSV parsing chooses stable column structure over separator frequency and preserves physical lines in multiline records, BOM and explicit separator files
Empty and cover worksheets remain available without blocking data sheets; the first sheet containing records opens automatically
Unsafe large-number arithmetic stays unverified; tighter comparison prevents decimal-sized errors being accepted in large totals
Analysis answers with missing or invalid evidence receive one bounded repair with a strict response schema, followed by independent source recalculation
Conflicting duplicate join keys require a user decision before grouping; original records must not be silently multiplied
16 added regression groups cover real source files, parser failures, blank versus zero, stale formula caches, evidence repair, ownership and cleanup
Synthetic live-model evaluation complements automated checks; source arithmetic checks do not certify business rules or full Excel recalculation
AI inspects title rows, merged and multi-level headings, independent table regions, repeated headers and subtotal candidates before calculating
One material clarification at a time, with selectable answers and follow-up context for scope, dates, numeric separators, units and missing-value rules
Independent source recalculation for declared sums, averages, minimums, maximums and numeric counts; mismatched answers are withheld
Review source sheets and ranges, exclusions, blank rules, numeric interpretation, recalculated values and a downloadable evidence record with the source file fingerprint
Formula caches, unsupported calculations, business rules and table-selection correctness remain explicitly outside the arithmetic guarantee
Local imports retain colliding column names and trailing unnamed fields, recognize headers after longer prefaces and correctly read grouped decimal-comma values
Source check and clarification controls support English, Turkish, German, Spanish and French
Opt-in full-file XLSX and CSV processing with clear disclosure before sending data to an external AI service provider
Python tool execution can inspect uploaded workbooks, calculate from actual cells and create new XLSX copies
Follow-up requests use the latest generated file; earlier downloadable versions can be selected again within the browser session
Independent output checks block malformed packages, formula error cells, macros and external connections; one repair attempt is available for rejected output
Preview reads actual downloaded cells, shared strings, basic formatting and saved formula results
Missing formula caches are disclosed; file-structure checks do not certify business rules or fully recalculate Excel formulas
Background progress, cancellation, bounded task duration and signed file-access tokens
Temporary provider workspace cleanup is requested after each task; uploaded files also have a 24-hour expiry
File processing controls and privacy explanations support English, Turkish, German, Spanish and French
Local mode retains its existing browser-only workbook tools
Creation runs independently of an uploaded analysis file; new requests open a workbook plan
Explicit row counts, charts, monetary exclusions, named sheets and appearance are checked before download
Worksheet previews show sample values and calculated cells; four appearance choices preserve the plan
Uploaded Excel files can explicitly guide a new AI template through sheet names, column types and calculated-field flags
Reference planning sends structural metadata; source cell values and original formulas stay local. Exact data, formula and style preservation uses the existing workbook editor
Bounded text-column additions and appearance changes preserve existing blueprint data and formulas
Local student-grade planning with shared exam weights complements the existing starting plans
AI plans receive one repair attempt for structural or explicit-requirement mismatches
File output uses content-aware widths and coordinated chart and table themes
Immutable rates and 115 million ARXAI reward cap; no automatic compounding
Flexible rewards calculated for complete 24-hour periods from each deposit
Voluntary locked early exit returns principal and cancels the entire position reward
Principal withdrawal and allocated flexible claims remain independent of batch settlement
Permanent emergency closure controlled by a 2-of-3 Safe; earned liabilities preserved and only unallocated rewards burned after settlement
At activation on 7 October, 115M of the original 575M reserve funded staking; the reserve was 460M then. See the 13 October contribution allocation update for the current 431.25M reserve.