Loading the atlas
Loading the atlas
Method · standards · limitations
The product standard is not maximum confidence. It is inspectable reasoning, honest provenance, and a willingness to publish “we do not know.” This page states the standards the atlas applies and the checks that are still missing.
Scope of this build
The autonomous worker is connected. It plans a new thesis, retrieves SEP and Wikipedia pages, checks quoted wording against retrieved text, builds both sides, and publishes into D1. It does not perform independent human review, and the relevance of a quotation remains AI-assessed. Fully verified counts therefore remain zero.
Operational statusFoundational rules
Generated text is never treated as proof of itself. Model output can propose an argument; it cannot certify one.
Sources must be retrieved, inspected, and connected to the exact claim they are said to support. A working URL proves only that a URL works.
Observational relationships are labelled as such. Causal conclusions require intervention, mechanism, or an explicitly defended inference.
A valid argument with unestablished premises does not establish its conclusion. Premise support is assessed separately.
A thesis does not become established because no contradiction has yet been found. Underdetermination is a real outcome.
The system searches for counterexamples, counterarguments, and competing explanations, and states what evidence would discriminate between them.
Some questions cannot currently be settled. The annotation is not a failure to be hidden by confident prose.
Published assessments are versioned. New evidence, retractions, or contradictions trigger reassessment and a recorded reason.
A conceptual argument, a mathematical proof, a historical claim, and a scientific hypothesis cannot be evaluated identically.
A reader should be able to follow the path from source to claim to inference to status, without trusting the system's summary.
Status vocabulary
A theorem can be proven within its axioms; a broad empirical theory generally remains open to revision. Status is always relative to a stated standard of proof and a stated scope. The atlas currently applies:
Evaluation
Reliability, evidential strength, validity, premise support, counterargument strength, independence, overall status, and uncertainty are stored separately. They are not averaged into a universal truth score.
If a number is shown, its meaning must be documented. A calibrated probability is appropriate for a well-defined, testable prediction with validation data — not automatically for a metaphysical or ethical claim.
Is the source authentic, credible, and methodologically appropriate?
Rated separately from how strongly the source bears on the claim.
How strongly does this evidence bear on this specific claim?
A reliable source can provide weak evidence for a claim outside its scope.
Does the conclusion follow from the stated premises?
Deductive validity, inductive strength, and abductive support are kept distinct.
How well established is each premise?
Established, disputed, assumed, or unknown — recorded per premise.
Pipeline
Not one enormous prompt. Each stage has a schema-validated intermediate result and a persisted state, so work can resume after a restart without duplicating a publication.
Choose a question from the pool, unresolved branches, or graph gaps; score for importance, researchability, information gain, and cost.
Convert the question into one or more precise propositions with scope, definitions, assumptions, and an evaluation standard.
Specify key premises, search queries, likely primary sources, competing hypotheses, verification procedures, and stop conditions.
Search trusted repositories and records. Preserve original metadata and retrieval time. Treat all retrieved content as untrusted input.
Record the specific claim, data point, or passage; attach a location where possible; separate the author's claim from the system's interpretation.
Graph integrity
The graph is the system's current understanding, not a declaration that every connection is objectively true. Every edge carries a type, rationale, confidence, origin, and verification status.
Verification
Existence
The source can be located and retrieved.
Authenticity
Known failure modes
AI research systems fail in patterned ways. These are design targets, not hypotheticals. The evidence gate catches some of them; a research-quality benchmark with expert-reviewed references does not yet exist.
The final standard: beautiful enough to invite exploration, rigorous enough to make its reasoning inspectable, honest enough to say “we do not know.”
Enter the atlasStatuses are assigned by criteria, not by the number of sources. No label is a numerical probability of truth.
| Status | Definition | Criteria |
|---|---|---|
| Proven | Established within a clearly defined formal system, or by an explicitly stated standard of proof appropriate to the domain. | A checkable derivation from stipulated axioms, or an equivalent domain-specific demonstration. Empirical theories are not labelled proven merely because evidence is abundant. |
| Strong | Multiple high-quality, independent lines of evidence support the claim, while material uncertainty remains. | Convergent independent evidence, no serious unexplained counterevidence, and an explicit account of what remains uncertain. |
| Provisional | The current evidence favours the claim, but important limitations or unresolved objections remain. | The balance of available evidence points toward the claim; independence or completeness is not sufficient for a stronger label. |
| Contested | Credible arguments or evidence materially conflict. | At least two defensible positions survive scrutiny, or high-quality evidence points in different directions. |
| Underdetermined | Available evidence cannot adequately distinguish between competing explanations. | The examined alternatives predict or accommodate the same observations; no discriminating evidence is currently available. |
| Unresolved | Insufficient evidence or analysis is available to reach a justified assessment. | The question is well formed but the investigation has not produced enough verified material to assess it. |
| Refuted | The thesis conflicts with established premises, a valid counterexample, or sufficiently strong contrary evidence under its defined evaluation standard. | A counterexample or contradiction is established relative to clearly stated assumptions. Refutation is always relative to a standard of proof. |
How serious is the strongest surviving objection?
Assessed at its strongest, not against a convenient straw man.
Are the supporting sources actually independent?
Ten repetitions of one unverified assertion are one source, not ten confirmations.
What conclusion does the whole assessment justify?
Assigned under an explicitly stated standard, not averaged from the dimensions.
What remains unknown, disputed, or unexamined?
Recorded as limitations, open questions, and missing discriminating evidence.
Check existence, metadata, quotation, location, context, independence, and whether the source actually supports the cited claim.
Construct the strongest supporting and opposing arguments. Search actively for counterexamples and alternative explanations.
Apply domain-appropriate checks: formal proof, consistency, evidence quality, methodology, source criticism, conceptual analysis.
Propose new theses and relationships only where supported by explicit argument or evidence. Detect contradictions and duplicates.
Pass the evidence gate, then store the summary, arguments, retrieved evidence, limitations, and status. If the gate fails, publish as unresolved.
Schedule reassessment on new sources, retractions, linked-status changes, review intervals, or user flags.
The worker implements planning, retrieval, excerpt matching, argument construction, status capping, and publication into D1. Independent human review and semantic verification of relevance are not implemented.
Author, venue, date, and identifier match the record.
Interpretation
The passage means what the evidence excerpt says it means.
Relevance
The passage actually bears on the specific claim, in context.
Opening a URL establishes none of these. It establishes that the URL resolved. In this corpus, no source has completed the four checks, so none is counted as verified.
Publication gate
Before publication: scope is precise; the summary matches the investigation; every claimed source exists and has been checked to the configured standard; material claims link to sources; both sides are represented fairly; the status follows the methodology; uncertainty and limitations are visible; edges have justifications; no unsupported certainty appears.
When no excerpt survives matching, the evidence gate fails and the thesis is published as unresolved rather than presented as supported.
Provenance
Theses are generated by a Cloudflare Workers AI model and published by the research worker. Sources are retrieved from the Stanford Encyclopedia of Philosophy and Wikipedia at research time. A quotation is stored only if its normalized text appears in the retrieved page. This proves existence and wording, not truth, context, or relevance. Model, dates, URLs, and verification level remain attached to each record.