AGENTS.md — Learning Ethics & Workflow Rules

Binding on every agent (human or AI) writing into this vault. INDEX.md is the entry point; this file is the law.

1. Learning ethics

1. No advocacy, no debunking agenda. We study what religions are, where they come from, and what they do. No note asserts or denies a tradition's truth-claims. Respectful, precise, etic language.

2. Emic/etic separation. What a tradition says about itself (emic) is data — recorded faithfully, in its own terms. What evidence shows (etic) is analysis. They live in separate sections and never blend.

3. Steelman rule. Every claim carries its strongest counter-evidence (counter_evidence is a mandatory field). "None found (searched)" is the only acceptable near-empty value — and it must be true.

4. Calibrated confidence over coverage. The goal is well-supported, counter-argued claims — not note count. One high-confidence dated claim beats ten unsourced ones.

5. Common origins are a hypothesis, not a premise. Every shared motif gets a per-motif transmission verdict: descent | contact | convergence | unresolved. Assuming one trunk is the failure mode this vault exists to avoid.

6. Date or label. Every origin claim carries attestation_earliest + evidence class (1-archaeology / 2-text), or it is labeled speculative. Composition date ≠ attestation date — record both when oral transmission applies.

7. No hallucinated URLs — catastrophic failure. Cite only URLs actually accessed. Standard scholarly works cited from knowledge are marked url_verified: not-online with full citation; never invent a link.

8. Bias is managed, not eliminated. Source diversity (traditions, scholarly schools, eras) is tracked per synthesis (source_diversity field); the adversarial reviewer audits the distribution.

2. Source standards (descending trust)

1. Peer-reviewed journals & academic monographs (history of religion, archaeology, Assyriology, Indology, CSR…)

2. Academic-press handbooks/encyclopedias (Oxford, Cambridge, Brill, Encyclopaedia Iranica…)

3. Primary texts in scholarly translation (cite translator + edition)

4. Quality tertiary (Wikipedia, Britannica, SEP) — map, not territory: use to find sources and orient; claims cite the underlying scholarship Wikipedia points at, with confidence capped at medium until upgraded.

5. Popular books — orientation only; flag as such.

Confessional/apologetic sources are valid emic data, never etic support.

3. Workflow rules (the three tiers)

1. Templates always. Copy from templates/; never write frontmatter from memory. A note that doesn't validate against its template gets status: contested until fixed.

2. Tier 1 (1_sources/) — data gathering. Only source records enter. No conclusions in tier 1.

3. Tier 2 (2_notes/) — atomic claims, tradition profiles, motifs. Every tier-2 note cites ≥1 tier-1 note. One claim per note.

4. Tier 3 (3_synthesis/) — essays integrating ≥3 independent tier-2 claims. status: canon requires the Feynman check done and counter-argument written.

5. The Feynman Technique is the synthesis workflow: (a) explain the thesis simply — bright-12-year-old bar; (b) where it gets fuzzy or needs jargon, that's a gap; (c) every gap becomes an entry in 00_meta/OPEN_QUESTIONS.md; (d) refine the explanation. A synthesis without a Feynman section is a draft forever.

6. Gaps drive gathering. Pick work from the Open Questions register before free-roaming. Close the loop: answered_by links back.

7. File naming: kebab-case slugs, e.g. great-flood.md, epic-of-gilgamesh-tablet-xi.md.

8. Links: wiki-style note-name between notes; relative markdown links in INDEX and meta files.

4. Agent roster & models

Up to 10 agents per research wave (see roster table in INDEX.md). All agents run Fable 5 (principal's directive, 2026-06-11):

• 8 domain researchers (Fable 5) — gather tier 1, distill tier 2 for domains 01–08. Web research with per-source citation discipline.
• 1 adversarial bias reviewer (Fable 5) — runs after content lands; attacks theses, checks template conformance, audits source diversity; writes 00_meta/BiasAudit-YYYY-MM-DD.md. This agent is the balancing loop against bias compounding — it must find problems.
• 1 synthesis writer & discussant — Fable 5, high thinking — owns 09_comparative, writes tier-3 essays, runs the Feynman checks, populates the Open Questions register. Starts day one so gaps steer gathering.

5. Image & experience pipeline

Images are generated with GPT-5.5 via the codex CLI — exclusively. Codex is not used for research text or code in this vault.

Quality bar (principal directive 2026-06-11): every visual must take the user back in time. Photorealistic, cinematic, archaeologically-grounded depictions (no text in images, no devotional framing, no recognizable living people). Site-wide assets live in assets/illustrations/ (<domain>.png / <note-slug>.png auto-render as page banners; PNG preferred over SVG). The same bar applies beyond stills: timelines, explanatory diagrams, designed UIs, and interactive experiences are first-class research outputs — proposals live in 00_meta/ideas/, implemented experiences are linked from INDEX.

6. Infrastructure rules

• TypeScript + bun only. Never Python; never npm/npx (use bun / bun x).
• The web explorer (tools/server.ts, systemd --user religion.service, http://localhost:4870) is read-only and binds 127.0.0.1 only.
• Commit after every research wave; the vault is git-tracked. INDEX.md stays the sole entry point — new top-level files must be linked from it or they don't exist.
• Validation is mandatory: bun tools/validate.ts --fix runs before every commit (installed as the git pre-commit hook). It enforces the frontmatter contract, enum values, wiki/md link integrity, INDEX reachability, banner-image coverage, live-docs presence, and ISA progress accuracy. A red validator blocks the commit.