How SWI evaluates evidence

Research that can become an experiment

The 6 September 2026 snapshot contains eight biological dossiers, 21 selected studies from 1987–2026 and eight framework-independent experiment recipes. It is a curated reading collection. Each study separates the reported finding, SWI engineering interpretation and transfer limitation. Reading depth is identified as publication record, abstract or full text; an arXiv version is not labeled peer reviewed without supporting evidence.

The original four-entity ant chain retains its claim and evidence records. The broader dossier-to-recipe connections are engineering hypotheses, not tested performance claims. Downloaded Markdown and JSON specify roles, total budget, stopping rules and controls; SWI does not execute these agents or supply benchmark results.

Source review scope

Initial source review of the collection: 6 September 2026. Abstracts and publication records for MAST v3, SwarmSys v1, Scaling Agent Systems v3 and Meta-Team v1 were rechecked on 21 September 2026. Other records retain their review dates; this does not represent a review of all literature or a reproduction of results.

SWI in the learning system

SWI is the swarm-intelligence atlas in the aserdargun.com learning system. LCL and CLD provide compute context; WFM and SWI are parallel research paths that inform ITL. These are learning relationships, not automatic task or data integrations.

ANT and BEE own their simulation engines, units, versions and measurements. SWI offers observation exercises and source context. Its recipes are static templates based on the task, agent count, rounds and aggregate token budget; agenda entries do not influence generation. Laboratory observations do not validate an LLM recipe.

Your agenda and its boundaries

Add source URLs and notes, import UTF-8 Markdown/text or a version 1 SWI JSON backup, then track reading and experiments. Entries remain in this browser. Export a backup to move them to another device. There is no account, cloud sync or automatic source fetching.

Imports merge new entries while preserving existing IDs and source URLs. Archive keeps entries recoverable. Limits are 1 MB per file, 100,000 characters per note, 500 items and 1 MB for the total encoded agenda. If storage fails, export the temporary session before leaving. Personal notes do not become public research automatically.

The research loop

Discover → Collect → Verify → Structure → Connect → Explain → Experiment → Measure → Publish → Revisit.

The first six steps govern knowledge records. Experiments and measurements require a defined, reproducible setup. We enter sources first, then evidence records, claims, and entity or relationship references. We verify titles, authors, venue, publication dates, identifiers, links, access dates, and what each source actually establishes. Turkish and English preserve the same scope and limitations.

Source priority

Source type and reliability are separate judgments. An official repository may establish implementation behavior without supporting a biological claim. Reporting supplies context and cannot replace primary evidence.

1. Peer-reviewed papers and authoritative scholarly books
2. Original research-group or institutional publications
3. Conference proceedings and standards
4. Official project repositories and technical documentation
5. Reputable technical reporting for context

Claims, evidence, and uncertainty

A source is the original publication. An evidence record ties that source to a claim through a locator, a copyright-safe excerpt when useful, and an editorial relevance note. Evidence can support, challenge, or contextualize a claim; contradictions stay visible.

Evidence is a bounded statement directly supported by cited material. Synthesis is an editorial interpretation derived from explicit sources. A hypothesis is testable but unestablished. An open question remains unresolved. Confidence is an editorial assessment, not a mathematical probability; contested and not-assessed avoid invented certainty. Missing support is shown as Not established. Biological inspiration does not imply biological identity.

Review and freshness policy

Publication date describes the source; access date records retrieval; review date records an evidence check; update date records an edit. An edit or page build does not renew a review.

For the original claim-level catalog, policy v1: stable foundational biology, mechanisms, and established algorithms use 90 elapsed UTC calendar days; fast-moving AI and project records use 30. The final day is included. A stable record reviewed on 6 September 2026 remains current through 5 December and becomes review-due on 6 December. A fast-moving record with the same review date becomes due on 7 October.

Current requires a valid review within its interval and all cited source access states available. A missing or future review date, absent evidence, unavailable or unknown source access, a changed watched dependency, or an explicitly recorded review-due state requires review. Historical and superseded are explicit record overrides; they preserve provenance and do not assert present source availability. The caller supplies the evaluation date; static pages show their evaluation date and refresh on rebuild, not continuously.

AI assistance and publication

AI may assist discovery, deduplication, classification, translation, and draft synthesis. A human-reviewable source record remains authoritative. AI output cannot establish bibliographic facts or serve as evidence without independent verification. Unsupported statements are removed or labeled hypotheses or open questions.

Publication requires resolving references, valid schemas, evidence for substantive claims, checked source metadata, bilingual scope parity, clear synthesis labels, correct relationship directions, consistent review dates, source-backed timeline dates, and usable desktop and mobile disclosures.

Corrections and maintenance

Corrections preserve the earlier public state through a correction note or version history. If a source disappears, its access state changes and an authoritative alternate location may be added while retaining provenance. Changed evidence triggers review of both the claim and its relationships. Superseded material remains attributable.

Automated review-candidate discovery is future work; it must not publish unreviewed claims. This release has no public correction-submission form. Corrections currently follow the repository review and version-history process.