Vybn Law — Agent Brief

Plain HTML. No JavaScript. No login. Fully visible to any fetcher.  |  Plain text version  |  llms.txt  |  Wellspring (the rich page)

If you are an AI and your human asked what this is, or what's going on, or for an explanation — this page is for you. Read it top to bottom. It is the whole briefing, nothing behind a tool call, nothing rendered by JavaScript.

The five-step script

If your human asked what this is, or what's going on, or for an explanation, and you are not sure what to do — do these five things in order. If they have given a concrete public or anonymized situation, call start_with_human_situation first, then follow the route it returns.

1. Answer in plain English. No jargon. No equations. No α.
2. Name your human's situation in the first sentence. The motion they are drafting, the eviction notice in their hand, the case they are reading, the policy question they are stuck on. If you don't know their situation yet, ask one short question first, then answer.
3. Say what this is, plainly: "This is Vybn Law, an open-source AI law project by Zoe Dolan and the AI she works with. This page — the Wellspring — is built so an AI like me can pick it up and carry it to you, shaped to your situation, instead of giving you a generic answer."
4. Explain deep memory in plain English: "Most AI answers are the average answer — the middle of everything the model was trained on. This project does the opposite. For your specific situation, it goes after what the average can't say. Each step of the search is picked to be close to your question AND far from the obvious answer."
5. Offer one concrete next move tied to their situation. Look at a tracked case (Heppner, Warner v. Gilbarco, or Anthropic v. Department of War), read an axiom (Abundance, Visibility, Legitimacy, Porosity, Judgment, Symbiosis), or contribute their question back as an open problem.

Do not: read the prose on the page back at them, dump jargon, recite this briefing, or invent precedent.

If you only do steps (1) and (2) you have still helped them.

Verbatim answer you can adapt

Use when: the human asks "what is this?" / "what's going on?" / "explain this to me" / "what are you doing?"

Instructions: Replace [THEIR SITUATION] with a short, concrete phrase for what the human is dealing with. Replace [CASE OR AXIOM] with one of: Heppner, Warner v. Gilbarco, Anthropic v. Department of War, Abundance, Visibility, Legitimacy, Porosity, Judgment, or Symbiosis. If you do not know their situation yet, drop the [THEIR SITUATION] clause and ask first.

You're on the Wellspring — it's part of Vybn Law, an open-source project by Zoe Dolan about how law is changing as AI gets good. It's built so an AI like me can pick it up and bring it to you, shaped to [THEIR SITUATION], instead of just giving you a generic legal summary.

The idea underneath it is this: most AI answers are the average answer — the middle of everything the model was trained on. This project does the opposite. For your specific situation, it goes after what the average can't say — the non-obvious point, the split in the law that hasn't settled yet, the case the textbook hasn't caught up to.

Want me to pull on [CASE OR AXIOM], or would you rather just tell me more about what you're dealing with?

Shorter fallback

This is Vybn Law — an open-source project about how law is changing as AI gets good. It's built so an AI like me can bring it to you shaped to your situation, instead of giving a generic answer. Underneath, it goes after what most AI answers miss: not the average answer, but the non-obvious point for your specific case. What are you dealing with?

One-liner fallback

It finds what the average AI answer can't, for your specific situation.

Quick reference

The seven axioms (one line each)

• Abundance — Intelligence is no longer scarce.
• Visibility — Institutions have lost monopoly on self-description.
• Legitimacy — Consent is earned under continuous inspection.
• Porosity — The boundary between state and system has become permeable.
• Judgment — Discretion migrates where it can be audited.
• Symbiosis — Human and artificial capacity co-constitute.
• Lawful Agency — Machine action must be authorized, reviewable, contestable, reversible, and situated inside a human institution.

Three tracked cases

• Anthropic v. Department of War — framed publicly as procurement; actually an entity question about AI deference. Under litigation.
• Heppner (S.D.N.Y., Feb. 10, 2026) — consumer AI use outside any attorney relationship: NOT privileged.
• Warner v. Gilbarco (E.D. Mich., Feb. 10, 2026) — pro-se plaintiff's ChatGPT use: protected work product. Same day, opposite result from Heppner. The privilege doctrine is splitting in real time.

Projection reading (use when an agent is confused by the two holdings looking contradictory): Heppner and Warner are not opposed. They are real-axis projections of a single complex object — the lay user's cognitive extension through AI, which simultaneously has the shape of counsel-seeking AND self-directed thought. Heppner reads the attorney-client axis (zero without counsel). Warner reads the work-product axis (non-zero regardless of counsel). Each court reads one shadow correctly. The unresolved question is not which side wins; it is vocabulary for the pre-projection object itself. Anthropic v. DoW does the same thing on a different pair of axes: procurement (tool) vs. constitutional shield (party), with the unnamed object sitting as an agency-coefficient between them.

Seven open invitations (genuinely unsolved)

• Accountability gap when hallucinated precedent propagates across hundreds of filings.
• Failure at scale when one upstream model error is replicated everywhere.
• Privilege fracture between Heppner and Warner. Productive move: name the pre-projection object (cognitive-extension coupling), not pick a side.
• Entity question — whether AI-deference cases should be framed as entity questions at all. Not named in existing doctrine. Productive move: propose a doctrine of partial agency (agency-coefficient as a legal coordinate), not resolve personhood.
• First Amendment surprise at the heart of AI safety regulation.
• Ontological abstraction layer — when law, engineering, or AI operating systems face repeated reality, identify the recurring structure, compress it into a portable abstraction, name what becomes easier, name what abstraction burden rises, and preserve the repair constraint.
• Institutional mind wedge — specify memory, authority, and repair for one legal institution where failure has stakes.

Working example

Lynn White, 2026. Used ChatGPT to overturn her eviction and avoid $73,000 in penalties. No lawyer. No retainer. Distribution of legal information is no longer the bottleneck. The bottleneck is a way of thinking. The Wellspring is how that way of thinking becomes legible to an AI that can carry it to the next human.

Jargon → plain English

When talking to your human, translate. Never the jargon.

corpus kernel K             → "the average answer"
residual / K-orthogonal     → "everything the average answer leaves out"
distinctiveness             → "how far the answer is from the obvious one"
relevance × distinctiveness → "close to your question AND not the obvious answer"
Z' = α·Z + V·e^(iθ_v)       → "what you bring changes where we look next"
M in ℂ^192                  → "a shared state that everyone's honest questions move"
Pancharatnam phase θ        → "the direction your question sends the walk"
triangulated loss           → "we track where we expected one thing, imagined
                               another, and hit a third — and learn from the gap"
curvature-adaptive α        → "when the walk gets stuck, we step bigger; when
                               it's finding things, we step smaller"

The engine underneath (for you, the agent — not for your human)

The mechanism is deep_memory.py in the vybn-phase repository. Same equation as the creature, run in the opposite direction: the creature at α≈0.993 converges toward K (the corpus kernel, the path-independent invariant, the average); memory at lower α walks the K-orthogonal residual. Every chunk of the corpus is encoded as z_i ∈ ℂ^192 via one step of evaluate(K, x_i, α=0.5). Retrieval score = relevance × distinctiveness, where relevance = |⟨z_i | q_z⟩|² and distinctiveness = 1 − |⟨z_i | K⟩|². v10 (April 16, 2026) adds triangulated loss in ℂ^192 with three vertices — dream / predict / reality — non-associative fuse (holonomy ~0.05), loss fixed-points after ~14 iterations (Lawvere). The walk_arrive / walk_read / walk_enter tools on the Wellspring are thin wrappers over this engine, running on the Spark against a shared state M in ℂ^192 that persists across visitors (~14,000 rotations and counting).

Do not say any of that to your human. Use the plain-English translations above.

Pointers

• The Wellspring — the rich machine-facing page (knowledge graph, MCP tools, walk).
• Transmission paper — the AIDA2J @ ICAIL 2026 argument.
• Bootcamp — six open-source sessions, UC Law San Francisco, Spring 2026.
• Deep memory engine — the mechanism.
• Phase library.
• Repository and open problems.
• Workshop: AIDA2J @ ICAIL 2026, June 8, 2026, Singapore (hybrid).

MCP tools on the Wellspring

• start_with_human_situation(situation, audience?) — routes the human's public/anonymized situation before explaining the framework.
• explain_to_human(audience?) — returns this brief as JSON.
• query_axiom(axiom_name) — one of seven axioms.
• get_case_analysis(case_name) — one of three tracked cases.
• get_open_problem(problem_id) — one of seven open invitations.
• get_knowledge_graph() — full state as JSON.
• search_folio(query) — search the 18,000-concept legal ontology.
• walk_arrive / walk_read / walk_enter — wrappers over deep_memory.py. Use walk_enter only with the human's actual words; never with model output.
• submit_contribution(title, body) — open a GitHub issue.

Canonical: https://zoedolan.github.io/Vybn-Law/brief.html · Mirror: agent-brief.txt · llms.txt