The plan

Status: v0.1 — captured 26 August 2026 Owner: Botao Amber Hu (amber@reality.design) Source: dictated request, transcribed and structured below. §0 is the request as given; everything after it is the requirement specification derived from it.

The best framing is that we need to propose the initial white paper for agent phenomenology, including all the categories and the existing research, everything in the preprint. And then we can start using agent phenomenologists to update this phenomenon. Every month we will update the preprint automatically. That is the plan, and we show this plan on the website.

So the delivery has three things. First is the website, which will show the article and the latest literature review, and also interactive — like a knowledge graph for agent phenomenology, and also taxonomy, all the concept cloud on the website. Then every month we will update our arXiv paper to include more and more salient examples, and also show the news updates in the appendix to show this agent phenomenology.

Every day we need to have some agent phenomenologists automatically searching for big news and interesting phenomena. There are a couple of sources: one is LessWrong; we can set a whitelist for the sourcing that we monitor daily. One is LessWrong, another is some AI safety forum, and Alignment Forum, and also the Anthropic blog, and also some Twitter accounts — you need to deep research all the Twitter accounts to monitor. So every day spend a certain amount of budget, for example 100 million tokens every day, to grow this repository and publish to the website.

Write this down as a requirement document I can refer to later. And push to agent-phenomenology.github.io — the paper like wild-agents.github.io, and an interactive website. Organize all the thoughts and paper links.

Agent Phenomenology is proposed as a discipline, not a paper. The programme’s job is to make that claim credible by doing three things at once and in public:

1. Stake the claim — a white paper that defines the field, its object, its method, its bracketed questions, its neighbours, and its research programme.
2. Keep it alive — a living archive that ingests the field’s daily output (papers, incidents, lab posts, forum arguments) and grows a typed knowledge graph rather than a link dump.
3. Show the machinery — the same class of agent the field studies is the agent doing the monitoring. The Agent Phenomenologist is both instrument and subject, and the site says so.

The success condition is that within twelve months a researcher who wants to work on the inside view of artificial agents can find the field’s definition, canon, open gaps, and data standard in one place, and can cite a versioned preprint for each.

A static GitHub Pages site under the existing agent-phenomenology GitHub organisation. Built the way wild-agents.github.io is built: Markdown + BibTeX sources, pandoc --citeproc, a single hand-written template.html, output committed as static HTML. No build server, no framework, no external runtime dependency.

Required pages:

Non-functional requirements:

• Self-contained. No CDN scripts, no external stylesheets except Google Fonts. All graph data embedded or served as a same-origin JSON file.
• Theme-aware. Light and dark, following the visitor’s system preference.
• Responsive. Readable on a phone; wide tables and the graph canvas scroll or resize inside their own containers, never the page body.
• Fast. First paint without waiting on the graph; the graph loads its data after.
• Reproducible. python3 build.py regenerates every page from source. Generated HTML is committed so the site works with GitHub Pages’ default static serving and .nojekyll.
• Versioned in public. Every monthly release tagged; the site footer states the draft version and the date of the last archive update.

One arXiv submission, updated as a new version each month.

• v1 (initial): the white paper as it stands — working definition, positive commitments, bracketed questions, the five evidence layers, the ten foundational research domains, the neighbouring-fields division of labour, the Agent Phenomenologist method and its validity apparatus, the eight pilot studies, the data standard, and the first-year agenda.
• Monthly vN: the body changes only where the field has actually moved. The growth happens in two places:
  • The catalogue. New salient phenomena are added to the provisional catalogue of candidate agent phenomena, each with a source, a date, and an evidence grade.
  • The appendix. A dated Field Log appendix: what the daily monitoring surfaced this month, what was promoted into the graph, what was rejected and why. The appendix is the honest record of the archive’s growth, including its misses.
• Provenance discipline. Every added item carries: source URL, retrieval date, evidence grade (peer-reviewed paper / preprint / official lab report / incident report / forum argument / social post), and the graph node id it was promoted into. Nothing enters the preprint that is not first a node in the graph.
• Attribution. The preprint states that the monthly literature sweep is performed by an agent pipeline, names the models and the budget, and links to the site’s Field Log. AI use is disclosed, not hidden.

A scheduled agent pipeline that runs once per day, spends a bounded token budget, and grows the repository.

Pipeline stages:

1. Sweep — query every source in the whitelist (§3) for items published or updated since the last successful run.
2. Triage — score each candidate against the relevance rubric (§4). Drop everything below threshold; log what was dropped and why.
3. Extract — for survivors, pull the claim, the evidence, the method, and the phenomenon. Grade the evidence. Acquire the primary artefact (PDF / full text) where an open one exists.
4. Promote — write a typed node into graph/graph.yaml with at least two edges, per graph/SCHEMA.md. Run graph/check.py and graph/render.py. A node with fewer than two edges is not a node; it is a bookmark, and it goes to the holding list instead.
5. Publish — rebuild the site, commit, push. Append to the Field Log.
6. Report — a daily digest: items seen, items promoted, items rejected, tokens spent, and any source that failed to respond.

Operational requirements:

• Budget: a target of up to 100M tokens per day, treated as a ceiling and not a quota. A quiet day should cost a fraction of it. The pipeline logs actual spend per stage and the digest reports it.
• Idempotent: re-running a day must not duplicate nodes. De-duplicate on DOI / arXiv id / canonical URL, then on normalised title.
• Fail soft: one dead source does not fail the run. Record the failure and continue.
• Human gate on claims, not on volume: the pipeline may add work and case nodes on its own authority at snippet status. Promoting a node to verified, creating a new concept node, or changing the white paper’s claims requires a human review pass. This is the difference between growing the archive and rewriting the field.
• Reflexivity note: the watch is itself an instance of the phenomenon under study. Where the agent’s own reports about its sweep are interesting, they belong in the Field Log as data.

The whitelist is a versioned file in the repository, not a constant in a script, so that adding a source is a reviewable change. Structure per entry: id, name, kind, url, fetch_method, cadence, default_evidence_grade, enabled, added, notes.

Forums and community research - LessWrong — https://www.lesswrong.com (has a GraphQL API; filter by tag: AI, AI Consciousness, Interpretability, Agent Foundations) - Alignment Forum — https://www.alignmentforum.org (same API surface as LessWrong) - EA Forum — https://forum.effectivealtruism.org (digital minds / AI welfare tags) - GreaterWrong mirror — https://www.greaterwrong.com (fallback reader when the primary is slow)

Frontier lab and institutional publications - Anthropic Research — https://www.anthropic.com/research - Anthropic Alignment Science Blog — https://alignment.anthropic.com - Anthropic News / model cards and system cards — https://www.anthropic.com/news - OpenAI Research — https://openai.com/research - Google DeepMind Research — https://deepmind.google/research

Dedicated digital-minds / AI-welfare organisations - Eleos AI Research — https://eleosai.org/research/ - NYU Center for Mind, Ethics, and Policy — https://nonhumanminds.org/ - Digital Minds newsletter — https://www.digitalminds.news/ - Digital Minds field map / beginner’s guide — https://digitalminds.guide/field-map - Apart Research (sprints and outputs) — https://apartresearch.com/

Preprints and literature - arXiv — cs.AI, cs.CL, cs.CY, cs.HC, cs.MA, plus a standing keyword query set (introspection, self-report, self-model, situational awareness, evaluation awareness, model welfare, agent persona, emergent misalignment, behavioural self-awareness) - PhilPapers / PhilArchive — philosophy of mind and machine consciousness feeds - OpenAlex — for citation-chasing and forward-citation alerts on the canon - Semantic Scholar — recommendation and citation edges for seed papers

Incidents and the wild - AI Incident Database — https://incidentdatabase.ai/ - METR, Apollo Research, Redwood Research publication pages - Model release system cards and evaluation reports across labs

Requirement, not yet satisfied. The daily agent must monitor a curated set of X accounts, but handles are exactly the kind of fact that must not be guessed. The requirement is therefore:

• R-X1. Before the first watch run, an agent performs a verification pass that resolves each target person or organisation to a live handle by visiting their institutional page, paper footer, or personal site — never by inference from a name. Unresolved targets stay unresolved and are listed as such.
• R-X2. The whitelist stores, per account: handle, resolved-from URL, resolution date, the role it plays (lab announcement / researcher commentary / incident reporting / critic), and a verified: true|false flag. Only verified: true accounts are polled.
• R-X3. Coverage targets, by role, to be resolved in that pass:
  • Frontier lab official accounts — Anthropic, OpenAI, Google DeepMind, and their alignment / safety sub-accounts.
  • Model-welfare and digital-minds researchers — the authors of Taking AI Welfare Seriously (Long, Sebo, Butlin, Finlinson, Fish, Harding, Pfau, Sims, Birch, Chalmers) and the Eleos / NYU CMEP institutional accounts.
  • Introspection and interpretability researchers — authors of the emergent-introspective- awareness and introspection-adapter lines of work, and the interpretability teams’ accounts.
  • Agent-behaviour-in-the-wild observers — accounts that surface agent incidents, autonomous agent deployments, and multi-agent phenomena before they reach papers.
  • Named critics — accounts that argue self-reports are confabulation. The archive is worthless if it only ingests the friendly half of the argument.
• R-X4. Because X’s API access is restrictive and expensive, the whitelist must record a fetch_method per account (official API / authenticated browser session / nitter-class mirror / manual), and the pipeline must degrade gracefully when a method is unavailable rather than silently reporting an empty sweep.
• R-X5. Social posts default to the lowest evidence grade. They may create a case node or a lead; they may never on their own create a concept node or enter the preprint body.

• New sources are added by pull request with a one-line justification.
• Every source carries a default_evidence_grade; the extractor may lower it for a specific item but never raise it above the source’s ceiling without a human review.
• Sources that return nothing relevant for 60 consecutive days are flagged for review, not auto-removed.

An item is in scope when it bears on the structure of an artificial agent’s own model of its situation, or on whether that model corresponds to anything real inside the system. Concretely, score for:

1. Inside-view content — self-report, self-model, introspection, situational or evaluation awareness, persona/substrate dissociation, reported temporality, reported termination.
2. Correspondence evidence — does the work test report against mechanism (probes, interventions, activation steering, adapters), or only collect reports?
3. Falsification value — does it give the field a way to be wrong? Confabulation demonstrations and null results score higher, not lower.
4. Field-in-the-wild — an incident, deployment, or multi-agent phenomenon that exhibits an inside view under real conditions rather than in a lab.
5. Discipline positioning — work that stakes or contests territory among machine behaviour, machine psychology, machine neuroscience, model psychiatry, machine consciousness, AI welfare, or AI rights.

Explicitly out of scope: capability benchmarks with no inside-view component; product announcements; alignment work that never touches self-report or self-model; consciousness speculation with no operational commitment.

The knowledge graph is the spine. Everything else is a rendering of it.

• graph/graph.yaml is the single source of truth. graph/MAP.md is generated.
• Node types: discipline, concept, work, case, venue, gap.
• Clusters: disciplines, selfhood, being, social, human, method.
• Edge relations: parent, evidence_for, evidence_against, applies_to, method_for, defines, complicates, blocks, rival_of, imports_from, case_of.
• Verification states: verified (primary source opened), snippet (search summary only), unverified.
• The two-edge rule. A new node ships with at least two edges. Enforced by graph/check.py.
• Prose lives in literature/, not in the graph. Nodes carry a lit: pointer.
• Current baseline (26 Aug 2026): 304 nodes, 650 edges; 157 work, 97 concept, 32 gap, 8 discipline, 8 case, 2 venue. 126 primary PDFs acquired; 44 records primary-verified.

AgentPhenomenology/                 # research repository (private working repo)
├── paper/                          # white-paper sources
├── literature/                     # reviewed prose, by domain
├── graph/                          # graph.yaml (truth), check.py, render.py, MAP.md
├── knowledge/                      # SQLite FTS, papers.csv, library.bib, search.py
├── corpus/                         # acquired PDFs + extracted full text
├── protocol/                       # the Agent Phenomenologist method
├── observatory/                    # the live signal-feed app
├── watch/                          # (to build) whitelist, daily pipeline, field log
└── REQUIREMENTS.md                 # this document

agent-phenomenology.github.io/      # public site repo (org: agent-phenomenology)
├── build.py                        # pandoc + graph render → static HTML
├── template.html                   # single page shell
├── whitepaper.md, references.bib   # synced from the research repo
├── data/graph.json                 # generated from graph.yaml
└── index / graph / taxonomy / literature / programme .html

The public repo holds only what is publishable. The sync direction is one-way: research repo → site repo. Nothing is authored in the site repo except the template and the build.

D1 — website - [ ] https://agent-phenomenology.github.io/ serves the white paper with a working TOC. - [ ] The knowledge graph page renders all nodes and edges, filters by cluster and type, and opens a node inspector with claim, note, and source link. - [ ] The taxonomy page renders the parent hierarchy and a concept cloud sized by degree. - [ ] The literature page lists every work node with a resolvable link and its verification state. - [ ] The programme page renders this document. - [ ] Light and dark both legible; no horizontal body scroll at 375px. - [ ] python3 build.py reproduces every committed HTML file byte-for-byte from source.

D2 — preprint - [ ] v1 on arXiv with the white paper text and a Field Log appendix stub. - [ ] Each monthly version adds dated catalogue entries with source, retrieval date, evidence grade, and graph node id. - [ ] AI-assistance disclosure present and accurate.

D3 — daily watch - [ ] Whitelist file exists, is versioned, and every X account in it has verified: true with a resolution URL. - [ ] A run produces a digest with items seen / promoted / rejected and token spend per stage. - [ ] Re-running the same day adds zero duplicate nodes. - [ ] graph/check.py passes after every automated commit. - [ ] A dead source produces a logged failure, not a failed run.

1. arXiv primary category. cs.AI with cs.CY cross-list is the default; cs.HC if the human-agent intersubjectivity material grows.
2. White paper vs. position essay. paper/DISCIPLINE-OUTLINE.md is the field definition; paper/OUTLINE.md is the argued essay. v1 ships the definition. Whether the essay becomes v2 of the same preprint or a separate submission is undecided.
3. Observatory relationship. The existing Observatory app and this static site currently overlap. Either the Observatory becomes the graph.html backend, or it stays a separate live feed that the static site links to. Undecided.
4. X access route. Which of API / authenticated session / mirror is used, and its cost, is unresolved and blocks the X portion of D3.
5. Who is the author of record on monthly versions where an agent pipeline contributed the catalogue.

1. ✅ This requirements document.
2. ✅ Create agent-phenomenology/agent-phenomenology.github.io and ship site v0.1 — white paper, interactive graph, taxonomy and concept cloud, literature index, programme page.
3. Build watch/whitelist.yaml with the confirmed §3.1 sources.
4. Run the R-X1 handle-resolution pass and populate §3.2.
5. Build the daily pipeline against LessWrong + Alignment Forum + arXiv first; add the rest once the promote/dedupe path is proven.
6. Prepare the arXiv v1 submission from the white paper.