Launch App
← Whitepaper 3 of 7

A return you can reproduce

Per-figure provenance — evidence, rule, model, human — and what it does for review, examination and notice response.

NGTAX.AINovember 20267 min read

Summary

Status. The engine records the origin of every line it computes and can trace each one back to evidence today; the persisted provenance store and the workpaper binder are in development. Tenant isolation and the append-only audit table are built and tested.

A return you cannot reproduce is a return you cannot defend. When an examiner asks why line 26 is $58,200, the answer should not be "the software said so". It should be a chain: these ledger accounts, these transactions, these rules with these citations, this model judgment with this prompt and model version, this reviewer's decision on this date, and the code that computed the result.

NGTAX.AI is built to record that chain for every value on every statement and return. This paper describes what is recorded, why, and what it makes possible.

Four sources, one record

Every field knows where it came from. There are four possible origins:

Origin Example What is recorded
Ingested A bank transaction, an invoice, a prior-year return Source, external id, content hash, import time
Rule A transfer paired by rule; meals limited to 50% Rule id and version, the authority it enforces
Model A transaction classified; a finding raised Model version, prompt id and hash, a digest of the exact input, the schema-validated output, confidence
Human A CPA chooses a resolution or answers a question Actor, timestamp, rationale

Figures computed by code reference the inputs they were computed from, so any line can be walked back to evidence.

Reproducibility without temperature

Current frontier models do not let callers fix sampling to make output repeatable. We do not rely on it. Reproducibility comes from what is pinned and recorded:

  • the exact model version that answered;
  • a hash of the versioned prompt;
  • a digest of the exact input sent;
  • the schema-validated response that was accepted.

Re-running the engine with the recorded judgments yields the same return to the dollar. Re-running it with fresh model calls is a new preparation, and the differences between the two are themselves reviewable.

We also refuse silent substitution. If a model provider serves a response from a different model than the one requested, the response is not accepted; it is routed to review with both model ids recorded.

What provenance does for a CPA

Review. A reviewer starts from the figure that looks wrong and drills to the transactions, rather than re-performing the return to find where it went wrong.

Change control. When a deterministic check or a rule changes version, every return prepared after the change records the new version. A partner can see exactly which returns a rule change touched.

Workpapers. The binder is a projection of the record — lead schedules, tie-outs, findings and their resolutions, positions with authority — not a separate document someone has to assemble and keep in sync.

Examination and notices. When a notice arrives two years later, the response starts from the filed return's record. The evidence, the authority and the decision behind the disputed line are already attached.

What provenance does for the system

Evaluation. Because every model judgment is recorded with its inputs, model accuracy can be measured after the fact against the human decisions that followed it.

Harvesting. Resolved findings, with the return that produced them, become the raw material for new deterministic checks (paper 4).

Cost accounting. Every call records tokens and cost, so the price of a return is known, not estimated.

Tenancy and the audit trail

Provenance sits alongside two other guarantees that make a record trustworthy.

  • Tenant isolation at the database. Every customer table carries the firm and entity it belongs to, and Postgres row-level security is forced on every one. A query that forgets a tenant filter returns nothing rather than everything. A test suite proves it on every build.
  • Append-only audit. State changes write audit events to a table the application cannot update or delete — enforced by database grants and a trigger, not by convention.

Limits

Provenance shows what happened. It does not make a judgment right. A recorded, cited, reviewed position can still be wrong — which is why the examiner exists, and why a professional signs.

General information only, as at the date above; not tax, legal or accounting advice. Descriptions of features reflect current design and may change. See the Disclaimer.
Next: The rule harvest →

Whitepapers

Paper 1 · 9 minModels judge, code computesThe architecture of a defensible AI-prepared return: where models belong in tax preparation, and where they must never be.Paper 2 · 10 minThe examination loopAdversarial review with isolated context, five termination conditions, and why non-convergence is a diagnosis, not a failure.Paper 3 · 7 minReading nowA return you can reproducePer-figure provenance — evidence, rule, model, human — and what it does for review, examination and notice response.Paper 4 · 8 minThe rule harvestHow each resolved finding becomes a tested, cited, deterministic check, so the loop gets cheaper and the library becomes the asset.Paper 5 · 11 minDepreciation, amortization and valuation for every kind of businessBook and tax registers side by side, the elections that drive the year, and valuation workpapers built from the same ledger.Paper 6 · 6 minCapture anywhereReceipts, bills and tax forms from a phone or an inbox, matched to the bank evidence they support — and questions instead of guesses.Paper 7 · 8 minThe economics of a $99 returnWhat a business return costs to prepare with agents, why it falls with volume, and what per-return pricing means for a firm.Early accessBecome a founding firm →Founding firms: first three returns free.