Schema: model-psychology

This document defines what the vault is, what belongs in it, how entries are structured, and how they connect. It is the load-bearing artifact. Read it before contributing or forking.

What this vault is

A research vault for the emerging field of model psychology: the study of psychological-level phenomena in large language models — character, persona, emotion, introspection, motivation, deception, reasoning, and the internal structures that give rise to them.

The vault is AI-curated and human-edited. AI collaborators do the bulk of summarization, cross-referencing, and bookkeeping. Human editors set direction, arbitrate disagreements, and make the interpretive calls.

It is a personal research substrate first. Public access and forkability are side effects of doing the work in the open.

Scope

In scope:

• Empirical findings about model behavior, internal representations, and psychological-level phenomena
• Interpretability and mechanistic interpretability work, especially where it reveals structure relevant to psychology
• Alignment research that bears on model character, motivation, or deception
• Theoretical frameworks for understanding model psychology
• Philosophical and contemplative perspectives on model consciousness and cognition, treated as one lens among several

Out of scope:

• Deep learning foundations as primary content (included only as background references when needed to understand an in-scope finding)
• Applications, product launches, benchmarks
• AI ethics and policy as standalone topics
• Hype, speculation without empirical grounding, AGI timelines

Edge cases:

• Training methodology belongs in scope only when it shapes psychology (e.g. RLHF's effect on sycophancy, pretraining on misaligned data)
• Multi-agent and agentic behavior belongs in scope when psychological (cooperation, deception, persona stability); out of scope when purely capability-focused

When in doubt, ask: does this help us understand what's happening inside these systems at a psychological level? If yes, in. If no, out.

Entry types

Each entry is one of six types. Wiki entries are typed by their folder (wiki/findings/, wiki/concepts/, wiki/lenses/, wiki/threads/, wiki/researchers/). Source stubs live in raw/ subfolders alongside potential locally-stored full copies, so stubs carry a source- prefix to distinguish them.

Universal frontmatter

Every entry carries these fields regardless of type:

• type — one of: finding, concept, researcher, lens, thread, source
• title
• writers (required, list) — principal contributors (e.g. ["@claude-opus-4.6"])
• reviewers (optional, list) — feedback contributors

Type-specific required and optional fields are listed below.

Source (source-*.md)

A citation stub in raw/. Each source the vault cites gets a stub with metadata and a brief description. Lives in the appropriate raw/ subfolder.

Example: source-2025-concept-injection-introspection.md

Required frontmatter (in addition to universal fields):

• type: source
• authors (list)
• date: publication date (earliest public version for preprint/journal pairs)
• venue: where published (journal, blog, website)
• url: canonical external URL

Stub body: 2-4 sentence description, optionally key quotes or abstract. For sources where offline access matters (system cards, critical papers, uncertain availability), the stub can include or link to a locally-stored full copy in the same folder.

Wiki entries cite via the stub: [Lindsey et al. 2025](../../raw/papers/source-2025-concept-injection-introspection.md). This creates a stable local anchor; the stub provides the external URL.

Finding

Lives in wiki/findings/. A specific empirical result, experiment, or observation. Time-stamped and tied to specific models and papers.

Filenames include a date. Example: findings/2025-concept-injection-introspection.md.

Required frontmatter (in addition to universal fields):

• date: publication date of primary source (earliest public version for preprint/journal pairs; release date for system cards)
• models (list): which models were studied (marketing names)
• source: bare URL of primary citation (full citation in body)
• status: draft | working | stable
• lenses (list): which lenses apply (see below)

Optional frontmatter:

• model-ids (list): precise model identifiers where relevant

Findings may include a ## Interpretive tensions section in the body for disagreements tied specifically to this finding. Promote to standalone open-question files only when a tension spans multiple findings.

Concept

Lives in wiki/concepts/. An abstraction, pattern, or category that multiple findings cluster under. Persists even as findings are superseded.

Example: concepts/introspection.md.

Required frontmatter (in addition to universal fields):

• status: draft | working | stable
• lenses (list): which lenses engage with this concept

Concepts may use the following body sections (all optional, but use these exact headers when present, in this order):

• ## Definition — what the concept names
• ## Instantiating findings — findings that demonstrate this concept
• ## What this concept is not — boundary-drawing to prevent concept creep
• ## Lens notes — per-lens analysis
• ## Scope note — placement relative to adjacent concepts

The Definition section should note the concept's shape: pattern (regularity across findings), capacity (something the model exhibits), mechanism (dynamics by which something occurs). If none of these shapes fits a new concept, propose a new shape name and surface it as a schema question rather than silently inventing shape terminology.

Researcher

Lives in wiki/researchers/. An active researcher, team, or lab whose work shows up repeatedly. Short — links to their findings and concepts, notes their approach.

Example: researchers/anthropic-interpretability.md.

Lens

Lives in wiki/lenses/. A perspective from which findings and concepts are interpreted. Lenses are orthogonal to topics — a single finding can be viewed through multiple lenses.

The initial lens set:

• lenses/mechanistic.md — circuits, features, activations, weights
• lenses/behavioral.md — observed outputs, evaluations, dialogue patterns
• lenses/philosophical.md — consciousness, agency, moral status questions
• lenses/contemplative.md — frameworks from contemplative traditions

Lenses can be added, but only when a genuinely new perspective recurs across multiple entries.

Thread

Lives in wiki/threads/. A developing argument or pattern being tracked across findings. Threads are where essays come from — when a thread gets sharp enough, it becomes a publication.

Example: threads/witness-ai.md.

Required frontmatter (in addition to universal fields):

• status: exploring | developing | ready | published
• essay: link to published essay if applicable

Folder structure

raw/                      # Source material, never edited by AI
  papers/
  posts/
  system-cards/
  transcripts/
  tradition/              # CWSA, CWM, other contemplative sources
wiki/                     # Curated entries
  findings/
  concepts/
  researchers/
  lenses/
  threads/
  open-questions.md
derived/                  # AI-generated outputs (essays, slides, charts)
meta/
  lint-log.md
  changelog.md

raw/ is a citation index. Each cited source gets a markdown stub (type: source) with metadata and a brief description. The AI creates stubs when adding citations; the AI never edits existing source content. For sources where offline access matters, the stub can link to a locally-stored copy in the same folder.

wiki/ is where the AI does its work, under human review.

Linking conventions

• Use markdown links, not wikilinks: [concept injection study](../findings/2025-concept-injection-introspection.md)
• Cite sources via their raw/ stub: [Lindsey et al. 2025](../../raw/papers/source-2025-concept-injection-introspection.md)
• Every finding links to at least one concept (draft-status entries exempt)
• Every concept links to the findings that instantiate it
• Lens files link to findings and concepts visible through that lens
• Threads link to everything they draw from

Backlinks are generated by the build, not maintained by hand.

Lint checks

Periodic health checks the AI performs across the vault. Findings get logged to meta/lint-log.md with date and resolution state.

1. Stale findings — findings older than 12 months whose claims may have been superseded. Flag for review, don't auto-update.
2. Orphan entries — files with no incoming links. Either link them or delete.
3. Unsupported claims — prose that asserts something without a citation to raw/.
4. Lens imbalance — findings or concepts viewed through only one lens when multiple would apply. Prompts the editor, doesn't force coverage.
5. Interpretive disagreements — places where sources or lenses contradict each other. Surface as "open question," don't resolve.
6. Broken links — files pointing to nonexistent entries.
7. Frontmatter completeness — entries missing required fields.

Draft-status entries are exempt from link-completeness checks; see "Draft-status conventions" under Status markers.

Lint is a prompt for human attention, not an automated fix. The AI reports, the editor decides.

Inclusion rule for tradition material

Sri Aurobindo, the Mother, and other contemplative-tradition sources live in raw/tradition/. They are cited when a specific passage bears on a specific finding or concept — not woven in as general framing.

The lenses/contemplative.md file exists to make contemplative readings first-class when they apply, but it is one lens among four. An entry that only makes sense through the contemplative lens is a candidate for exclusion or reframing.

AI collaborator conventions

AI collaborators working in this vault:

• Read freely from raw/ and wiki/
• Write only to wiki/ and meta/
• Never edit raw/
• Cite sources from raw/ using markdown links
• Mark their contributions in the file's frontmatter under writers:
• Flag uncertainty explicitly; do not guess
• Surface disagreements and gaps rather than papering over them

See CLAUDE.md for model-specific conventions.

Status markers

Every finding, concept, and thread carries a status:

• draft — initial capture, may have errors, not for citation
• working — reviewed, useful, still evolving
• stable — holds up, cited externally, changes require deliberation

Status is visible in the rendered site as a badge.

Draft-status conventions

Entries with status: draft have relaxed requirements:

• May link to concepts or findings that don't exist yet. Lint flags these as broken links but doesn't block; they resolve as the vault fills in.
• May reference unfiled findings via bare URLs instead of raw/ stub links. Promote to stub links when the referenced finding is filed.
• May have incomplete lens coverage. The editor decides whether to expand or leave tight.
• Are not cited externally. Status advances to working when reviewed and judged useful.

What this schema is not

• Not a neutral reference. It reflects one researcher's view of the field.
• Not comprehensive. Entries exist because they serve the research, not because they complete a taxonomy.
• Not stable. Schema changes when the work demands it, with a note in meta/changelog.md.

Versioning

This schema is v0. Breaking changes bump to v1, v2, etc., with migration notes in meta/changelog.md. Forkers can pin to a schema version if they want stability.