Survey for AI
Survey for AI
Research knowledge, designed for AI.
A reusable method and Skill for queryable, traceable, maintainable research knowledge. RSI is our worked example.
From a collection of papers
to a usable field of knowledge.
Define the field and reader questions first. Build the evidence, relationships and modules. AI retrieves the relevant knowledge while keeping source versions, conditions and genuine gaps visible.
Method before example.
Adapt, curate, relate, synthesize, answer and maintain. Each stage has its own inputs, outputs and gates.
Take the method to a new field.
Change the vocabulary and evaluation axes; retain identity, version and evidence rules.
RSI ↗
808 families · 26 branches · 38 reader modules. See the method at work.
From a field question to a supported answer.
Seven connected stages, each with explicit inputs, outputs and acceptance gates. Evidence permissions and unresolved questions remain visible.
Make the method your next starting point.
A domain adapter defines terminology, roles, comparison axes and reader needs. The shared workflow preserves identity, versions, evidence and answer rules.
One domain contract. One research workflow.
The adapter declares inclusion, cutoff, role fields, comparison axes, relation predicates, reader questions and budget. The shared Skill manages creation, resumption, evidence, relations, synthesis, answering and maintenance.
Read the Skill ↗Cross-domain effectiveness still needs controlled testing. A portable interface is not validation in every field.
python skill/survey-evidence-ledger/scripts/init_domain.py \ --output ../my-survey --domain-id my-domain \ --name "My field" --scope "My research question"
The AI answer contract
question · corpus_version · answer
claim_ids · citations {url, version, locator}
comparison_conditions · limits · gaps
status: supported | partial | insufficient_evidenceOne real field. Several ways into its knowledge.
Recursive self-improvement is the current worked example. Counts describe a fixed corpus, not global completeness or full-paper certification for every item.
Separate feedback, gains and acceleration.
Identify what changes, feedback, retained state and later use. A fixed model does not by itself negate external feedback; a local structural loop does not prove universal RSI or sustained acceleration.
Strict primary-review plan: 533/570, with 37 source/version/locator exceptions. 509 dossiers have saved passages and 241 have evaluation details.
Publication activity, by month
Read along branches. Connect across them.
26 navigational branches with typed relationships. A branch organizes reading; each relationship has its own evidence and meaning.
See how a paper becomes usable knowledge.
Seven representative cases connect mechanisms, protocols, retained objects, limits and actual sources. Public views show briefs and locators; the full ledger remains access-controlled.
Compare mechanisms and the conditions behind them.
Read distinct systems along shared axes. Formal guarantees, code evolution, parameter training and program search have different evidence boundaries; this is not a cross-protocol leaderboard.
Fields come from saved configurations; missing values remain explicit. Different tasks, metrics, models and budgets cannot be ranked directly by score.
Continue to questions and answers →Start with your question, then retrieve the evidence.
Choose a reader goal and explore existing bounded answers and evidence routes. These are curated module answers, not newly generated unreviewed model claims.
Find a paper. Inspect its evidence depth.
808 work families, filterable by role and branch. The interface and selected case briefs are bilingual; archival records preserve their original curatorial and source language.
Public presentation, explicit evidence and access boundaries.
The method and website are public. The complete RSI knowledge repository remains private. Downloads, versions, coverage and limits are documented separately.
Method & website
Bilingual introduction, Skill, method contracts, website source and public reading views.
GitHub ↗02 / ACCESS REQUIREDComplete RSI knowledge
Full dossiers and canonical ledgers; repository access required. Excludes the website, GitHub configuration, full PDFs and raw caches.
Download when signed in ↓03 / AI RETRIEVALAI reading & citation rules
Authority order, minimal retrieval, evidence permissions and citation boundaries.
Markdown ↓04 / PAPER INDEX808-paper machine index
Stable IDs, roles, branches, dossier paths and saved evidence locators.
JSON ↗v1.1.0 · 2026-10-23
Record progress and exceptions.
Updated 2026-10-23. Fixed scope: 808 families, 26 branches and 38 modules. Strict primary review is 533/570 with 37 remaining exceptions. Candidate and screening rows are not extra admitted papers.
Bilingual guidance, faithful archival records.
The interface, method, selected cases and root-question answers are bilingual. Dossiers retain their original curatorial language and English entry pages label this. Public views omit verbatim source passages and grant no new claim admissions.