Skip to content
GitHub
US Tax Prep

US Individual Tax Preparation

US Individual Federal Tax Preparation (Form 1040 and its schedules) is Spectral’s first world-model domain. The selection rationale lives in ADR-029; this subtree is the operator-facing and dev-facing companion — a problem-space breakdown, a source-material inventory, and a rule-coverage overview that evolves as the Operations-app dogfooding proceeds.


Three things have to be true of the first domain: it has to bootstrap the three-source corpus and the scan-pipeline integration tests, it has to be interesting enough to stay in the apps/test-agents harness indefinitely, and it has to be something a careful generalist can validate without an SME on call. US individual federal tax prep clears all three bars — publication-grade Authoritative sources (IRS publications + Form 1040 instructions) are freely available; the rule surface is genuinely diverse (filing status, deductions, credits, schedules); and a careful generalist can validate rules against published IRS text without a CPA’s sign-off per rule.

See ADR-029 for the full selection analysis and the rejected alternatives (Healthcare PA, Legal Contract Analysis, Building Codes, Financial Compliance).


Tax year, jurisdiction, form surface:

  • Jurisdiction: United States federal individual income tax only. No state, no local, no foreign-tax-credit-as-primary. (State and local come later; foreign-tax-credit appears only as a line item on Form 1040 / Schedule 3 within the federal return.)
  • Form surface: Form 1040 and the schedules attached by individuals — Schedules A, B, C, D, E, SE, 1, 2, 3, plus the 1040-instruction content that governs them.
  • Tax year discipline: The world model is versioned per tax year. Rule corrections across years flow through the release-notes restatement mechanism (per ADR-026) — prior years’ WorldModelCards remain authoritative for those years.
  • Taxpayer type: Individuals. Businesses (Form 1120 / 1120-S / 1065) are out of scope, even for pass-through income reported on the individual return.

Out of scope (permanently, for this domain):

  • State income tax, local income tax, sales tax, property tax, payroll-tax withholding beyond what appears on the 1040
  • Corporate, partnership, and trust returns
  • Estate and gift tax (Forms 706 / 709)
  • International filing complexity beyond what Form 1040 itself accommodates

Initial problem spaces selected to reach the 50-enshrined-rule minimum-viable floor while exercising the full provenance / mutation / holdout machinery:

Problem spaceIn in-scope?Why
Filing status determination✅ YesNarrow, decidable, high edge-case density (married-filing-separately qualifying conditions, head-of-household dependency rules, qualifying widow(er) continuation). Exercises rule-relationship surface (depends-on, conflicts-with, qualifies).
Standard deduction✅ YesTightly coupled to filing status — natural pair. Covers age / blindness / dependent modifiers; exercises provenance-tier behavior (amounts change annually from a single primary source).

These two together give roughly 50–70 enshrinable rules — enough to cross the corpus-bootstrap floor and enough variety (status classification + quantitative modifier lookup + edge-case handling) to stress-test the scan pipeline end-to-end.

Problem spaces enumerates the full list, including problem spaces outside the alpha corpus and their prerequisites.


The tax domain is rich; the alpha corpus deliberately starts narrow. The following problem spaces are not in the alpha corpus; each is listed with why it waits and what would lead the operator to pull it in.

  • Dependents (qualifying child / qualifying relative tests)
  • Itemized deductions (Schedule A — medical, SALT cap, charitable, home-mortgage interest)
  • Self-employment income + SE tax (Schedule C + Schedule SE)
  • Capital gains and losses (Schedule D, short-term vs long-term, wash-sale rules)
  • Tax credits (Child Tax Credit, Earned Income Credit, education credits)
  • Retirement contributions + distributions (IRA, 401(k), RMDs)
  • AMT (Alternative Minimum Tax triggers and computation)

See problem-spaces.mdx for the why-outside-alpha and pull-in-trigger for each.


Why this domain double-duties as the test-agent backbone

Section titled “Why this domain double-duties as the test-agent backbone”

apps/test-agents is the alpha home for a single tax_prep agent that exercises the full scan pipeline in CI. A pluggable OTEL emitter parameterized over two axes (instrumentation framework × LLM-vendor span shape) provides ingestion diversity at the trace layer while keeping the agent itself a single, deterministic tax-prep workflow. The agent + emitter described here land under the test-substrate epic; see Test Agents — Pluggable OTEL emitter for the current-state framing, the coverage matrix, and the CI tier policy.

The world-model domain and the test-agent agent are the same thing: the rules the tax-prep test agent exercises come from the published tax world model. This is deliberate — it means ingesting a CI trace and ingesting a real design-partner trace go through the same pipeline against the same world model.


Spectral is domain-agnostic infrastructure by construction — no customer domain-specific structure, features, or knowledge is designed into the system. Tax-prep is the first world model, not the only one the system supports. A second world-model domain ships when three conditions are met: an operator commits to the curation work for the second domain (the operator-walkthrough effort is the gating cost — bootstrap of a new world model end-to-end); the source-material inventory clears the same Authoritative-grounding bar tax-prep cleared (domain authority publishes specifications, regulatory text, or standards-body material that can ground the rule corpus); and the test-agent harness expands to cover the second domain so CI exercises both. The criteria for picking domain #2 are commercial — driven by design-partner demand and operator-team capacity — and live outside the Codex.

  • Problem spaces — full enumeration, alpha boundaries, prerequisites for pulling each space into the alpha corpus
  • Source materials — IRS publication inventory (links, revision years, relevance per problem space)
  • Rules — rule-coverage overview; populated as bootstrap proceeds
  • ADR-029 — the selection decision and rejected alternatives
  • ADR-022 — the three-source corpus this domain bootstraps
  • Operations App — the surface that authors the world model