AI Research Pipeline — How This Module Stays Current

How the HigherSelf Network keeps the tax-architecture module up to date via the ML-Intern integration. Cross-references the canonical strategy in the SecondBrain vault.

NOT LEGAL OR TAX ADVICE. This page describes infrastructure, not tax positions. Circular 230 notice applies to the rest of the module.

Why this page exists#

Tax law moves. IRS notices, Tax Court decisions, Treasury regulations, and statutory sunsets reshape the architecture every quarter. A static reference goes stale fast. The HigherSelf Network is integrating Hugging Face's open-source ML-Intern agent (https://github.com/huggingface/ml-intern) to keep this module current with a documented provenance chain.

Canonical strategy: ~/Documents/SecondBrain/2 - Areas/Grace/ML-Intern-Strategy.md — the full per-entity mission map, governance, and 90-day rollout plan lives in the vault.

What ML-Intern does in this module#

On a weekly schedule, an ML-Intern mission scoped to tax-architecture/ runs against:

IRS newsroom updates (last 14 days)
Tax Notes free articles
arXiv econ.GN, cs.CY papers tagged with tax / regulatory keywords
Hugging Face Papers ranked by relevance to module keywords
GitHub code search for new public Solo 401(k) / S-corp / QBI calculators

The agent assembles a candidate set of updates, drafts new or revised content with frontmatter and provenance metadata, and opens a pull request against the knowledge-base repo. The PR body lists every source URL with retrieval timestamps.

What ML-Intern does NOT do here#

Never auto-merges. Every PR is reviewed by Lumina (UPL / tone), the outside CPA (where tax substance is touched), and Grace Fields (authorization gate).
Never modifies files outside tax-architecture/.
Never publishes specific dollar amounts or filing recommendations without an explicit human approval label.
Never trains on cohort PII or facts without written consent.
Never invokes Hugging Face Jobs (paid GPU) without Grace authorization. Phase 1 is curation-only at $0 compute spend.

Governance summary#

| Gate | Owner | Trigger | |---|---|---| | needs-lumina | Lumina | Every PR | | needs-cpa | Outside CPA | Any tax-substance change | | needs-erisa | ERISA counsel | Any 401(k) / ROBS / plan-loan change | | needs-grace-auth | Grace Fields | Merge gate | | Auto-close at 14 days idle | Bot | Stale PRs |

Why a cohort member should care#

Two reasons:

Provenance. Every research artifact in this module has a source_provenance frontmatter field listing the arXiv / IRS / case-law sources that contributed. You can audit the chain back to primary sources.
Drift detection. If § 199A sunsets, IRS Notice 2025-67 is superseded, or a Tax Court decision changes how Watson-reasonable comp is benchmarked, the ML-Intern weekly run will surface it within days, not quarters.

What this is NOT#

ML-Intern is research infrastructure, not a substitute for licensed counsel. Every cohort-facing implementation still goes through:

Engaged outside CPA for tax positions on returns.
Engaged ERISA attorney for retirement-plan structure.
Engaged state tax counsel for state-level conformity.
Lumina + Grace Fields authorization before publication.

The ML-Intern integration shortens the cycle between "the law changed" and "the research library reflects it." It does not change the requirement that real implementation needs real professionals.