AI that participates in the model.
Build in the grid. Branch the work. Run the model. Explain what changed. Replay the decision.
AI built in. Audit built in. No black box.
One approved model can be versioned (“Desk” vs. “House”), run across many companies and/or with many scenarios, tracking versioned lineage, approved data, approved AI, replay, and Audit behind every result. AI is native, diffing is native, versioning is native.
Build in the grid. Mark Assumptions and Outputs directly in the spreadsheet your team already knows.
Keep work organized by lineage. Branch analysis, promote reviewed work, and return working models to a known-good state without filename sprawl.
Run one model across the coverage universe. Each company keeps its own assumptions, data, Runs, Results, Audit, and replay context.
Diff runs, not just spreadsheets. Compare Scenarios, Sensitivity Cases, and Results side by side.
Connect approved data directly in the grid.
=DATA("bank.net_interest_income") maps business concepts to approved
internal and external data through one formula interface.
AI workflows run inside the model.
=AI() formula can execute reviewed Methods: chained analysis, approved
context, validation, typed output, Model Memory, YAP chronology, and Audit. Inside
the model.
Replay historical decisions. Reconstruct models with the model state, data, provider evidence, AI method, memory, chronology, and time rules that applied then.
Audit every result. See what ran, what changed, which evidence was used, who approved it, and why the result was valid at that point in time.
Sheet Warnings. Every Run checks for errors before they become problems. See what formulas may have drifted, hardcodes replaced formulas, and other changes that need review. Spreadsheet linting for finance.
Why fin123? Spreadsheets used to be publishing artifacts: investment banking models, sell-side models, printed reports, static assumptions. Then the buyside started building live internal models for capital allocation, scenario analysis, and execution. Models forked. Assumptions drifted. AI entered the workflow. The spreadsheet became a living institutional decision system, but most software still treats it like a file. fin123 was built for what spreadsheets actually became.
fin123 is not another AI spreadsheet. It turns spreadsheet work into a replayable, explainable decision system.
Analysts keep working in the grid: assumptions, outputs, formulas, approved data, and approved AI stay where the model work already happens.
=DATA() and =AI() as model steps.Model work moves by lineage, not filenames. Teams can branch analysis, promote reviewed work, and return working models to a known-good state.
One approved model can run across a whole coverage universe. Each company keeps its own assumptions, data, Runs, Results, Audit, and replay context.
Build -> Branch -> Run -> Explain -> Replay
When a number moves, fin123 shows what changed, which evidence was used, who approved it, and why the result was valid at that point in time.
Most systems answer: What is the latest number?
fin123 answers: What did we know, what did we run, who approved it, and why was the result valid then?
The spreadsheet became a governed decision system.
The root surface is still the spreadsheet. The model state around it is versioned, replayable, and explainable.
There is no File menu and no Print menu because fin123 is not a document editor. It is a system for building, branching, running, explaining, and replaying model decisions.
Mark cells as Assumptions or Outputs, save model versions, select Scenarios, choose company context, Run, and inspect Results without leaving the spreadsheet surface.
The abstraction is not open, print, and close. It is build, branch, run, explain, replay.
Type =AI() into a cell. The
approved definition lives behind the cell, and it may reference an approved Method:
versioned institutional methodology for how the formula should reason, validate output,
and preserve evidence during a Run. Approved Model Memory and YAP chronology can supply
decision context when they are part of the reviewed model record.
Results are based on marked Outputs from the saved model state. Run All can execute one model across selected companies and Scenarios while preserving the context behind each number.
Charts and visualizations can help, but they are secondary to the explainable Run record.
Audit opens from Results. It shows what ran, what changed, which model state, Scenario, data evidence, provider evidence, AI method, approved memory, chronology, approval, and time rules supported the result.
Keep model work organized by lineage, not filenames. Branch analysis, promote reviewed work, and return working models to a known-good state.
Run the same approved model across a coverage universe while preserving each company’s context.
=DATA() gives analysts one formula shape for approved internal and external
data.
Every execution is reviewable. Replay explains the decision; Audit shows the evidence behind it.
The hard part is not: “run 10 models.” The hard part is: “prove what the model knew, what it ran, and why the decision was valid then.”
Not a side chat. Not copy/paste. =AI() lets reviewed AI workflows
execute as model steps.
=AI() sits in a normal worksheet cell. The formula
definition lives behind it, and the definition can reference an approved Method: versioned
institutional methodology that defines how the formula should reason, validate output, and
preserve evidence during a Run.
A Method can be more than a prompt. It can define a chained analytical workflow: retrieve evidence, apply approved context, compare history, validate assumptions, reconcile intermediate outputs, and return a typed result to the model.
Methods stay behind the formula. Analysts still work in the grid. Results receive only validated typed output. Audit shows the approved Method, prompt, inputs, output, validation, memory, and chronology used.
Model Memory adds the missing layer: approved institutional context attached to the model. YAP preserves the decision chronology around the work. Together they let AI use approved context without becoming a side-chat that mutates the model.
FINRA has made the direction clear: as firms adopt GenAI, supervision and compliance expectations follow the technology. Its October 2025 AI update describes internal AI governance, human-in-the-loop controls, training, security review, and ongoing engagement with member firms on GenAI use cases and supervisory implications. fin123 gives firms the operating layer for that reality: approved AI methods inside the model, typed outputs, evidence, replay, and Audit for every AI-assisted result.
Cell B12: =AI()
Reviewed Method:
- retrieve management guidance and prior quarter commentary
- compare historical SSS seasonality and recent comp trends
- apply approved Model Memory
- validate assumptions against source evidence
- reconcile intermediate outputs
- return typed model output
Inputs:
- prior quarter SSS
- management guidance text
- recent comp commentary
- macro / consumer read-through
- historical seasonality
Typed output:
Forecast SSS: 0.035
Forecast revenue: 1035
Audit:
- Method version
- inputs and evidence
- validation checks
- approved memory
- YAP chronology
- Run identityScheduled execution is the same Run model, just delayed. It references a saved model state, selected Scenarios, and selected company contexts. It never runs the mutable grid.
The modal is a delayed Run confirmation: saved version, selected Scenario, and execution time. Unsaved changes are not included.
Run later means: run this saved model state with these Scenarios at this time. Use it for one name, a scenario pack, or a whole coverage universe. When it triggers, fin123 leaves the same Results and Audit trail as an interactive Run.
Headless running is not unattended Office automation. Microsoft’s official guidance says it does not recommend or support automating Microsoft Office applications from unattended, non-interactive client applications or components because Office can become unstable or deadlock in that environment. By contrast, fin123 model Runs are executed from the saved model record and approved formula/data/AI definitions, not by opening Excel on a server. Read Microsoft’s unattended Office automation guidance.
The product contract is Model -> Version -> Scenario -> Run -> Results -> Replay -> Audit. Every institutional decision becomes a replayable, explainable model record.
Model
-> Branch / Version
-> Company + Scenario
-> Run
-> Approved DATA + AI Workflow Execution
-> Results + Diff
-> Replay + Audit
-> Data + Provider Evidence
-> AI Method + Approved Memory
-> YAP Chronology + Time Rules
YAP remembers how the pod thinks. fin123 governs what the firm executes.
@yap why did we lower $AAPL gross margin in Q3?
Response:
Gross margin was reduced from 46.2% to 44.8% in Scenario Base on 2026-05-12.
Primary drivers:
- Management commentary suggesting elevated NAND costs
- China mix deterioration concerns
- Temporary services mix normalization
Approved memory:
- Source: earnings reaction thread
- Approved by: JG
- Confidence: medium
Affected formulas:
- gross_margin_q3
- product_margin_assumptionasof123 governs what time meant for the decision. It records market state, holidays, publication windows, effective dates, stale status, and replay-valid time context.
Replay uses those same time rules instead of judging the past with today’s information. A PM, analyst, risk team, or compliance reviewer can reconstruct what the desk could actually know at the time.
YAP preserves the thesis record. fin123 runs the model record. asof123 preserves the market clock. Together they make institutional decisions replayable and explainable.
Reckoning Machines is the runtime layer behind chained AI execution: methods, evidence, validation, typed outputs, replay, and audit.
ReviewedMethod
-> Evidence
-> Validation
-> TypedOutput
-> Replay
-> Auditprompt123 is a product in development and is not yet deployed to fin123.
prompt123 is optional prompt improvement before running. Reviewable, approved, audited, and nothing done silently.
It flags ambiguity, missing schema, hidden assumptions, nondeterministic wording, and unsafe external dependencies. If the prompt is unclear, prompt123 records findings instead of silently deciding what the user meant.
LLM-assisted proofing may propose clarifications or normalized draft language, but every suggestion remains advisory. Approval and execution stay with the analyst.
PromptIntent
-> PromptDraft
-> ApprovedPrompt
-> ExecutionArtifact
-> AuditOpen the hosted app: app.fin123.dev
Want the walkthrough? Email reckoningmachines@gmail.com
fin123
The spreadsheet became the operating system of business. fin123 makes it a replayable, explainable decision system.
Reckoning Machines is a financial-services-focused software company founded by a senior developer and former portfolio manager.
Most systems start with UI and execution workflows, then discover the hard problems later: replay, provenance, time-aware reconstruction, and review.
We build from replay, Audit, and time-aware execution forward, because institutional systems only become trustworthy when the decision record is native to the product.