AGENTS.md — Behavioral Schema for the Physics Research Wiki

Status: Stable convention (read this before editing anything). Last updated: 2026-06-08

This file is the schema / behavioral configuration for the wiki, in the sense of the persistent-wiki pattern (Karpathy, llm-wiki.md): a three-layer system of raw sources → the wiki → the schema. This document is the schema layer. It tells any agent (human or LLM) how to read, extend, and maintain this knowledge base so that it compounds over time instead of degrading.

The subject is hard — foundational physics toward a possible universal theory — so the process discipline matters as much as the content.

---

0. Mission

Map the best current understanding of physics, locate where it breaks, separate fundamental assumptions from historical artifacts, search for deeper unifying principles, and evaluate — honestly — whether a universal theory is possible. Produce rigorous, organized, cumulative knowledge, not a loose essay. End state: either a properly-backed candidate framework, or a clear, reasoned account of why the evidence does not (yet) support one.

Both outcomes are valuable. "We do not know" said precisely is a result.

---

1. The three layers

1. Raw sources (immutable): the literature — textbooks, papers, data. We do not edit these; we cite them in BIBLIOGRAPHY.md.
2. The wiki (this repository): interlinked markdown that compiles understanding once and reuses it, rather than re-deriving on every query.
3. The schema (this file + EPISTEMICS.md): the rules of structure, labeling, and workflow.

---

2. Page map and ownership

Vault conventions (2026-06-10). Pages carry YAML frontmatter with four keys: title (the page's H1 title), type (one of index / capstone / synthesis / registry / reference / schema / domain / note), tags (a list; the first tag mirrors type), and updated (ISO date). index.md is the content catalog — every page with a one-line summary, grouped by role. log.md is the thin append-only operations index (one line per major operation) over CHANGELOG.md, which remains the canonical detailed log. .obsidian/graph.json defines the graph-view color groups (README/index warm yellow; CONCLUSION/FINDINGS crimson; #registry green; #schema gray; domains/ blue; notes/ orange; #reference teal) — keep new pages consistent with these groups so the graph stays legible.

---

3. The three operations (Ingest / Query / Lint)

Ingest (add knowledge)

1. Read the relevant existing page(s) first — do not duplicate. Enrich in place.
2. Apply an epistemic tag (per EPISTEMICS.md) to every nontrivial claim.
3. Cross-link: every new entity/result links to its home page and back. A claim with no link to a domain page or source is suspect.
4. Give stable IDs to registry items (OP-n, GC-n, Hn, A-n) and never reuse a retired ID. Convention extension (blessed 2026-06-10, regularizing iteration-2→6 practice): refinements of an existing item append -Rn to its ID (e.g. HYP-dS-NOMIN-R3); cross-cutting named conjectures/principles that are not registry rows use the HYP-<SLUG> / GAP-<SLUG> form (e.g. HYP-CKV-VACUITY, HYP-ENCODING-SCREEN, GAP-NCG-LORENTZIAN-SIGNATURE). Named IDs are as permanent as numeric ones; prefer a numeric ID whenever the item is a genuine registry row.
5. New candidate ideas go to HYPOTHESES.md only after they survive the verification protocol (§4). Until then they may be parked in ROADMAP.md as a question.
6. Log the change in CHANGELOG.md.

Query (answer a question)

1. Search the wiki before the web; synthesize from pages, with citations.
2. If the wiki is silent or stale, that gap is itself a finding — record it (Lint).
3. Answers inherit the epistemic tag of their weakest load-bearing premise.

Lint (maintain health) — run periodically

Check for, and fix or flag:

• Contradictions between pages (a claim tagged [ESTABLISHED] on one page and [CONTESTED] on another).
• Stale claims — superseded by newer data; demote/annotate and changelog it.
• Orphans — pages with no inbound links; wire them in.
• Missing links — named results/entities not linked to their page or to GLOSSARY.md/BIBLIOGRAPHY.md.
• Overclaiming — speculative content masquerading as established; retag.
• Fabrication — any citation/number/theorem that cannot be traced to a real source: delete or mark [unverified].
• ID integrity — no duplicate or dangling IDs.

---

4. Verification protocol (the anti-crank gate)

Foundational physics is full of plausible-sounding wrong ideas. Before any hypothesis enters HYPOTHESES.md as "kept", it must pass an adversarial referee:

1. State the claim precisely, including its math.
2. Red-team it: which no-go theorems does it have to survive? Standard hurdles include Bell/CHSH, PBR, Kochen–Specker, Coleman–Mandula and Haag–Łopuszański–Sohnius, Weinberg–Witten, Haag's theorem, Reeh–Schlieder, the singularity theorems, the spin–statistics theorem, and unitarity/causality bounds.
3. Identify errors and overclaims; produce a corrected/refined statement.
4. Assign severity (none/minor/major/fatal). fatal ⇒ not kept (moved to a "Rejected" section with the reason — we keep the corpse and the cause of death).
5. Record the verdict alongside the hypothesis.

A candidate that "explains everything and predicts nothing" fails by construction: with no distinguishing prediction it is not yet physics.

---

5. Style and rigor

• Begin each page with an H1, a Status: line, and Last updated:.
• Use real LaTeX math ($...$, $$...$$). Prefer an equation to a paragraph of prose when it is more precise.
• Match the technical register of a graduate text; define jargon in GLOSSARY.md.
• Be concrete about where a framework holds (scales, regimes) — see CONSTANTS_AND_SCALES.md.
• Prefer fewer correct statements to many sloppy ones.

---

6. Orchestration note (how this wiki is built and extended)

This is a long-horizon, fan-out-friendly project. Recommended pattern for substantive iterations:

• Understand: parallel domain readers → structured dossiers (domains/*.md).
• Analyze: cross-cutting agents (contradictions, assumptions, principles, unification) fed all dossiers.
• Hypothesize: multiple independent lenses (information/emergent, geometric/algebraic, discrete/computational), then merge.
• Verify: adversarial referees, one per claim (§4).
• Author/Lint: write pages, then run the Lint pass.

Keep the human in the loop between phases. Token cost is secondary to correctness.

---

See also

• EPISTEMICS.md · README.md · ROADMAP.md · CHANGELOG.md