Legally defensible AI tax answers, cited to the full chain of authority.

For any tax question, TaxCanon assembles the full chain back to controlling authority — statute, regulation, ruling, case — across a typed citation graph of 176,455 primary-source documents and 1.91 million typed relationships. Every claim resolves one click from the document, the snippet, and the authority that supports it. The practitioner reads the same chain the model walks.

W

Inside every authority · snippets · summary · relationships Tap any node to open.

  • Phase A

    Live

    Corpus, citation graph, lifecycle resolution. Queried from the database today.

  • Phase B

    Coded

    AI processing pipeline above the graph. Five iterations deep; current run awaits production-scale compute.

  • Wave 1

    Planned

    Public benchmark, throughput report, MCP reasoning chain. Ships when Wave 1 compute lands.

Demos throughout this page show Phase B output on a 30-authority anchor set, produced by running the same prompts the production pass will run across the full 150,028-document active corpus.

The landscape

No tool currently shipping returns an answer the IRS or a court can recognize.

Every existing AI tax product surfaces single-source citations — documents the model thought relevant, attached as footnotes. Tracing the chain back to controlling authority — statute, regulation, ruling, case — is the professional’s job to do by hand.

What's at stake

A taxpayer who follows IRS guidance, gets audited, and ends up in tax court still has to defend from primary authority — statute, regulation, case law. Sub-regulatory guidance is persuasive, not controlling. A careful reading of the IRS’s own words is no defense without the chain back to authority that does.

Existing AI tax solutions

The five products tax practitioners actually pay for. Each surfaces documents and single-source citations; the chain back to controlling authority is left to the user.

  • Search across primary authority

  • Editorial annotations and treatises

  • Per-document treatment indicators

  • AI document summaries (rolled out 2023+)

  • Single-source AI citations

What only the graph enables

Incumbents
TaxCanon

These products were built before a typed citation graph at this scale was tractable. The AI features layered on top in the past two years summarize documents and produce single-source citations — there’s no chain of authority underneath to rely on for accuracy or display for verification. A correct tax answer that can’t be cited back to primary authority isn’t correct. It’s only confident.

What TaxCanon does

Three things only a typed citation graph makes possible.

Showing both positions walked simultaneously. Walking 7 of 7 authorities.CONFLICTS WITH§170(h)§1.170A-14PL 96-541HewittGlade CreekOakbrook 6thOakbrook T.C.

Hover or focus a node to read its authority.

walking 7 of 7

Every defensible position, walked end-to-end.

Most tax questions admit multiple defensible positions, and which one binds depends on the taxpayer’s circuit, the lifecycle state of each authority, and how the chain assembles back to controlling primary law. The graph traces every position — every authority, every snippet, every lifecycle state, every binding-jurisdiction filter — so the practitioner compares them on the merits and advances the one that holds.

Refresh cadenceweekly
  1. Distinguished by 5th Cir. — narrowed to virtual-currency-only facts.

Diff addressable as chain@wk-18 → chain@wk-19

Weekly refresh, full version history.

New rulings publish, regulations finalize, cases come down — every week. The graph absorbs each change and versions the corpus, so a chain assembled today is addressable as a diff against the same chain a month ago, a year ago, or at the moment a position was originally taken.

  1. chain @ 2017-08

    When position was taken (pre-TCJA)

    chain hash

    4b7e2a91f3c5

  2. chain @ 2018-12

    TCJA enacted; statutory shift in progress

    chain hash

    8c1d9e44b7a2

  3. chain @ today

    Post-Treasury Regulation; statute + reg controlling

    chain hash

    e5a3f6c2d8b1

Each version is addressable, exportable, and citable back to. The position taken in 2017 is defensible from the chain that supported it in 2017 — even if the law has moved since.

Every chain you trace is an artifact you can cite back to.

Every chain TaxCanon assembles is captured as a versioned artifact — the question, the authorities, the snippets, the lifecycle state of each node, the binding-jurisdiction filter — addressable by URL and exportable to memo, audit response, or court filing. Six months later, when the IRS asks how the position was reached, the answer is the exact chain that supported it on the day it was taken.

Worked example · §6751(b) penalty approval

Every defensible chain assembled. The professional decides which holds in court.

Most nuanced tax questions admit multiple defensible positions, and which one binds the taxpayer depends on the circuit, the lifecycle state of each authority, and how the chain assembles back to controlling primary law. TaxCanon assembles every chain, ranks them by the criteria a court would apply, and surfaces both for the practitioner’s verification. The judgment call — which position to advance for a specific client and jurisdiction — stays with the professional.

The question

When must the IRS supervisor approve a penalty assertion under §6751(b)(1)?

Both rules hold. Each node opens for the source — the multi-layer summary, the verbatim snippets, the connected authorities. Which rule binds depends on the taxpayer’s appellate circuit.

TaxCanon returns

Two defensible positions exist in the active corpus, both grounded in §6751(b)(1) — the requirement that initial penalty determinations be “personally approved (in writing) by the immediate supervisor” before the IRS assesses a tax penalty. The positions diverge on what “initial determination” means, and therefore on when, procedurally, that approval must occur.

Position A — Chai rule.Approval must precede the IRS’s first formal communication of the penalty determination to the taxpayer — the notice of deficiency, or FPAA in partnership cases. Established in Chai v. Commissioner, 2d Cir. 2017; the Tax Court adopted the rule court-wide in Graev v. Commissioner III, T.C. 2017 and continues to apply it for cases not appealable to a contrary circuit.

Position B — Laidlaw / Kroner rule. Approval is timely if obtained at any point before the formal assessment under §6203, even after the taxpayer has received a notice of deficiency. Established in Laidlaw v. Commissioner, 9th Cir. 2022; reached independently in Kroner v. Commissioner, 11th Cir. 2022.

Ranking is jurisdiction-contingent.Both positions sit lifecycle-active at the controlling appellate level. Neither has been resolved by the Supreme Court, by Treasury Regulation, or by amending statute. Under the Tax Court’s Golsenrule, the controlling holding is the one of the Circuit to which the taxpayer’s case is appealable. TaxCanon returns the rule binding on the taxpayer’s litigation forum and surfaces both for verification.

Position A · Chai rule

Source
Chai (2d Cir. 2017), adopted in Graev III
Authority tier
4 (Circuit)
Lifecycle
Active
Posture
Appellate precedent
Binding in
2d Cir.; T.C. cases otherwise unappealable to a contrary circuit
Conflicting auth
Laidlaw; Kroner

Position B · Laidlaw / Kroner rule

Sources
Laidlaw (9th 2022), Kroner (11th 2022)
Authority tier
4 (Circuit × 2)
Lifecycle
Active
Posture
Appellate precedent
Binding in
9th Cir.; 11th Cir.; T.C. cases in those circuits under Golsen
Conflicting auth
Chai; Graev II

Confidence rank for both positions is jurisdiction-dependent: the controlling holding is determined by the Circuit to which the taxpayer’s case is appealable.

TaxCanon ranks per jurisdiction. The professional decides. Both chains are below — every authority, every snippet, every relationship — and the binding-jurisdiction filter resolves to the taxpayer’s actual forum.

The conflict

Both positions begin at §6751(b)(1), enacted by PL 105-206 §3306 (RRA 1998) to constrain IRS penalty assertion procedures. They diverge on what “initial determination” means: the IRS’s first formal communication to the taxpayer (Chai) or the internal assessment step under §6203 (Laidlaw, Kroner). Both rules sit lifecycle-active, joined by a typed CONFLICTS_WITH edge — neither resolved, both binding within their geographic scope.

The engineering

A typed citation graph of U.S. federal tax law. The first at this scale.

176,455 primary-source documents across 24 authority types — every IRC section, every Treasury Regulation, every Revenue Ruling since 1954, 50,911 federal court tax decisions, the Internal Revenue Manual, IRS Notices and Announcements, Revenue Procedures, Private Letter Rulings, Chief Counsel Advice, Treasury Decisions, JCT Bluebooks, and the rest. Acquired from 11 public-source providers. Year coverage: 1866 to present.

The shape of the work

A two-pane reading of the project today: on the left, what the loaded corpus actually contains, broken down four different ways. On the right, the extraction work that is already done, the AI processing that is coded and waiting on compute, and the claim-grounding work scoped for after.

The corpus, dimensionally

Top eight document types in the active corpus; fifteen smaller types are rolled up.

  • Fifteen other types (IRC sections, IRM, AOD, treaties …)+15,169 · 10.1%

Top eight + rollup · 149,571

From the loaded corpus to the AI pass

Phase A

LIVE · queried from database

176,455 documents · 1.91M typed relationships

The corpus, the citation graph, the lifecycle states, and the authority tiers are loaded and queryable today. Every figure above is read from the live database.

Corpus figures

Total documents
176,455
Active corpus
150,028
Typed citation edges
1,909,091
Pinpoint resolution (overall)
87.4%
Pinpoint IRC (with subsection)
59.4%
Validity matrix triples
761
Phase A extraction runtime
7h 49min · single MI300X
Pipeline iterations
Run end-to-end across five corpus generations

Phase A is the substrate of every dimension at left — the corpus, the graph, and the lifecycle states are read directly from staging.db.

The AI processing pipeline

The AI pipeline that produces every demo on this page. Four capability layers:

  1. Classification

    Every authority placed in the legal taxonomy and tagged with the tax doctrines it implicates.

  2. Content extraction

    Multi-layer summaries and structured metadata for every authority, with every claim resolving to its source citation.

  3. Graph structure

    Authority tier, binding-jurisdiction filter, and lifecycle / supersession resolution.

  4. Retrieval & calibration

    Paragraph-level retrieval and per-field confidence scoring.

Of the 176,455 documents in the graph, 150,028 form the Phase B active corpus — the documents that pass the quality gate for multi-layer summarization and structured-metadata extraction. The remaining 26,427 are queried directly from the graph but are not subject to Phase B processing in Wave 1. Every demo on this page comes from this pipeline. Run end-to-end four times before this iteration; Wave 1 scales the current iteration to production depth on the full 176,455-document corpus.

The production pass

Wave 1 runs the pipeline in roughly 1,000 GPU-hours on a single AMD MI300X. Wall-clock from compute access to first deliverable: approximately three weeks. Full Wave 1 — the MCP-served reasoning chain product, a 100-question benchmark against Checkpoint and CCH, the throughput-and-cost report, and the LoRA recipe — lands in eight to ten weeks. The compute conversation is open with AMD, GCP, and Azure.

176,455 documents · 1.91M typed relationships · 24 authority types · 11 public-source providers · 100% lifecycle-resolved · 1866 — present

Phase A is live and queried from the database. Phase B has been run end-to-end four prior times on smaller corpora; the current iteration is the largest yet and adds modifications targeting full citation-chain assembly. Those modifications have not yet been tested at scale; Wave 1 runs this iteration at production depth.

One graph, three commercial shapes

Built for the practitioner. Exposed for the platforms that build on top.

The same typed citation graph drives a first-party application for tax attorneys and accounting firms, an API for tax software vendors, and an MCP server for legal-tech and AI platforms. One graph layer, three commercial shapes — chosen so TaxCanon sits as infrastructure, not as a single-vertical bet.

Graph layer

One typed citation graph. Three commercial shapes. The same source of truth.

Questions

The questions a careful reader asks next.

The page above answers the load-bearing questions; the ones below come up after a careful reader has read it.

  • Models hallucinate when they’re inventing content from training memory. TaxCanon’s model doesn’t invent — it reads from the typed citation graph during inference, and every authority, every snippet, every relationship the model relies on is one click from the source document on the same surface the practitioner reads. A claim either resolves to a real authority in the corpus, or it doesn’t appear.

    The architectural posture: the model isn’t trusted to remember tax law; it’s required to look it up. The practitioner verifies the same chain the model walked, on the same authority records, on the same page. If the model and the practitioner disagree about what an authority says, that disagreement is visible immediately rather than buried in a fluent-sounding answer.

On the waitlist

Deliberate access. Personal write-back.

Tax attorneys, accounting firms, tax software vendors, and legal-tech and AI platforms get priority. We write back personally before granting access — no auto-confirmation cascade, no upsell, no marketing list. Leave an email; we’ll be in touch.

We'll write back personally before granting access. No newsletter, no list-share.