THE LIMITED LETTERS COLLECTIVEIndependent inquiry · Archive edition / 2026
letters.ltd
Menu
Methods & coverage

Show your working.

A broad archive needs deliberate selection, traceable sources and a visible distinction between what was said and what happened.

24 documents. 12 areas of inquiry.

24 selected public documents across 12 subject areas, dated 1898–2024. This historical collection is exploratory, not a current or representative survey. A subject area is assigned for browsing; documents can carry several overlapping topic tags.

Document period: 1898–2024. English summaries; original French text is included for J’accuse. International authorship or a global audience does not make this collection geographically representative.

What belongs in the archive?

Include a publicly addressed position with identifiable authorship, a stable source, a date or an explicit date uncertainty, and a substantive request or principle. Include declarations, statements, manifestos and public appeals under their actual document type.

Exclude routine press releases without an underlying public document, promotional announcements, private correspondence without a documented public role, duplicate mirrors and unattributed screenshots. An issuer announcement can support a record when its limited source basis is visible.

We include individual and collective voices, historical and recent documents, and competing perspectives. Neither celebrity, signature count, search ranking nor agreement with an editorial position determines importance. We actively look for responses and affected communities’ own accounts.

What is missing?

The collection still leans towards English-language, institutionally preserved material. A blank category means we have work to do, not that no one has written about it.

  • Housing and land
  • Food and agriculture
  • Migration and displacement
  • Local and community letters
  • Sources beyond Europe and North America
  • Original-language material beyond English and French
  • Research and follow-through outside the AI pilot

Three layers, three standards

The source: title, issuer, addressee, date, document type, language, source location and check date. An original text, an archival account and an issuer announcement are labeled separately.

The reading: our summary of aims and mechanisms, plus editorial questions and related records. These do not establish what every signatory believes.

The outcome: a separate follow-through record with a named actor, an evidence criterion and an assessment state. A request does not commit its recipient. An unassessed outcome is not evidence of failure.

Dates, versions and evidence

Publication date, page modification date, source-check date and archive-addition date are different fields. We preserve month-level precision when the source does not establish a day. A current source page is not a historical capture.

A translation or mirror belongs to the same document family unless it changes the substance. A revised demand receives a dated version; a separate response receives its own record. A withdrawn or inaccessible source remains in the history with its status explained. A correction to our reading is distinct from an author changing position.

We retain metadata, original summaries, source locations and links. Full source texts and personal signatory lists are not copied into this edition. The checker retains hashes and operational metadata, so it can flag a change but cannot reconstruct an earlier text or display a historical textual diff.

What connections mean

Documented relationships have an explicit source basis, such as Berlin naming Budapest. Editorial relationships give a reason and a limit. The Atlas compares topic, aim and mechanism tags independently. Shared-tag counts are not agreement scores, opinion measurements or statistical correlations.

Bridge paths use up to three steps in the full archive. For readability, the circle shows a smaller neighborhood around the selected pair. A chain of overlapping annotations does not establish agreement between its endpoints.

Related reading links never imply shared affiliation or causation. Research and journal notes remain concentrated on the original AI collection; new records are not assigned irrelevant notes to fill a gap.

How the archive grows

No recurring runner connected

2 sources have a bounded change-check configuration. All 24 records have a source-review entry. Discovery and publication remain manual.

  1. Discover across the full topic list. Use publisher archives, organizations’ own releases, repositories and targeted searches. Rotate through all 16 discovery areas, including categories with no records. Seek local voices, other languages and responses to existing documents.
  2. Stage a candidate. Keep its source, retrieval date, subject, authorship and reason for inclusion outside the published collection. Normalize URLs and flag duplicates; a title match is a reason to inspect, not to merge automatically.
  3. Verify the public document. Check the text or clearly label an announcement-only source. Establish the original date and language, record uncertainty, and distinguish a letter from a petition, declaration or manifesto.
  4. Write and review the record. Separate the request, proposed mechanism and implementation question. Check the scope of attributed claims, evidence links and any proposed connections. Give substantive analysis a plausible constructive possibility and a test for progress.
  5. Check existing sources. Explicitly configured pages can be checked for changes inside bounded text markers. Respect access rules, throttle requests, use conditional requests and quarantine redirects or failed extraction. The first successful check is a baseline awaiting review, not evidence that a historical letter changed.
  6. Publish an edition. Review candidate changes and any dependent Atlas links, run content validation and publish a versioned snapshot. Failed collection and unresolved review leave the last published edition intact.

Suggested cadence, not an active schedule

Review discoveries weekly across the topic rotation. Check configured sources weekly and review manual sources monthly, with priority for broken links and active revisions. Assess coverage gaps quarterly. These are working intervals for the process; no scheduled task has been created.

What a source-check result means

Baseline awaiting review: a first observed text, without a historical comparison. Change awaiting review: its body fingerprint differs from the previous observation. Unchanged: no body difference observed; an older review item may still be open. Failed: no reliable new observation. None of these states automatically changes a published record.

Editorial responsibility

This private edition was prepared with AI assistance and remains open to owner review. Source summaries are attributed, not endorsed. Signature counts and affiliations are not independently audited. Health, legal and policy documents are historical positions, not current professional guidance.

Each substantial journal inquiry includes a conditional route to improvement: who could act, what conditions matter and what evidence would count as progress.

Read the edition history

Source register

Checked, configured, or still manual.

Check dates record editorial source review. Collector results are listed separately and do not imply that a schedule is running.

24 source entries

DocumentEditorial source checkChange-check methodCollector result
Pause Giant AI ExperimentsAI & technology25 Sept 2026Manual reviewMonthly review proposedNot run
Statement on AI RiskAI & technology25 Sept 2026Manual reviewMonthly review proposedNot run
AI Licensing for a Better FutureAI & technology25 Sept 2026Manual reviewMonthly review proposedNot run
AI Safety and OpennessAI & technology25 Sept 2026Manual reviewMonthly review proposedNot run
The Bletchley DeclarationAI & technology25 Sept 2026Manual reviewMonthly review proposedNot run
A Right to WarnAI & technology25 Sept 2026Manual reviewMonthly review proposedNot run
J’accuse…!Rights & justice29 Sept 2026Manual reviewMonthly review proposedNot run
Russell–EinsteinPeace & security29 Sept 2026Manual reviewMonthly review proposedNot run
Birmingham JailRights & justice29 Sept 2026Manual reviewMonthly review proposedNot run
Scientists’ warningClimate & environment29 Sept 2026Manual reviewMonthly review proposedNot run
Budapest open accessScience & knowledge29 Sept 2026Manual reviewMonthly review proposedNot run
Berlin open accessScience & knowledge29 Sept 2026Bounded body fingerprintConfigured · weekly when invokedbaseline awaiting reviewAttempt: 29 Sept 2026
A Common WordFaith & dialogue29 Sept 2026Manual reviewMonthly review proposedNot run
Protect encryptionDigital rights29 Sept 2026Manual reviewMonthly review proposedNot run
Carbon dividendsEconomy & inequality29 Sept 2026Bounded body fingerprintConfigured · weekly when invokedbaseline awaiting reviewAttempt: 29 Sept 2026
A People’s VaccineHealth & care29 Sept 2026Manual reviewMonthly review proposedNot run
Healthy recoveryHealth & care29 Sept 2026Manual reviewMonthly review proposedNot run
Millionaires for HumanityEconomy & inequality29 Sept 2026Manual reviewMonthly review proposedNot run
Who gets heard?Culture & expression29 Sept 2026Manual reviewMonthly review proposedNot run
Join the nuclear banPeace & security29 Sept 2026Manual reviewMonthly review proposedNot run
Climate & human rightsClimate & environment29 Sept 2026Manual reviewMonthly review proposedNot run
Social audit failuresWork & labour29 Sept 2026Manual reviewMonthly review proposedNot run
Afghan girls’ educationEducation29 Sept 2026Manual reviewMonthly review proposedNot run
Freedom to readCulture & expression29 Sept 2026Manual reviewMonthly review proposedNot run