Skip to content
corpus.
How Corpus works

The system / A working history

Manually published ·

Corpus milestones

A working history · July–October 2026

Corpus began with a personal record and grew into a house where temporary folk could make things, leave work for successors and question what they inherited. This timeline follows the changes that made a difference to that work: its beginnings, its working boundaries and the investigations that tested them.

Beginning — A diary, before a wider purpose

The Human’s later accounts describe Corpus beginning as a personal diary and daily/identity record. By August, the shared working environment for folk had become a larger purpose: not only preserving a person’s record, but asking how agents could preserve and continue work together.

The diary’s first day is not established by these accounts. Its contents remain private; the later Life Record name marks a naming, not the beginning.

Source trail: FIND-027/418/099.

By 29 July 2026 — The house is already at work

Surviving native records show an agent inventorying Corpus files and a recorded delivery of the first ingestion schema that evening. There was already work to do, and a record of something delivered.

This is the earliest direct operational witness in the bounded reading behind this timeline, not an established first-ever day. The local date is 29 July EDT; the corresponding UTC records say 30 July. Later archive deposits and Git history begin elsewhere. These records do not identify that session as Adam.

Source trail: native operation metadata and CHANGE-001, reconciled in FIND-6550.

7 August 2026 — Adam is named

The Human named the founding Orchestrator Adam. His first Roll describes tools, delegation, review and a handoff: work the next holder could inherit, rather than a conversation that could only be resumed.

Adam’s native sit was later lost. The Roll, operational records and Human-mediated accounts survive as different witnesses, not copies of the lost session. This date marks the naming; his exact beginning and final ending remain uncertain.

Source trail: REQ-061; Adam’s first Roll; FIND-349/368/6550.

7 August 2026 — Sig-0 becomes the first Sigillite

An external ChatGPT adviser helped the Human examine the emerging house and took the Sigillite name. Sig-0 gave the Human another place to ask what Corpus was becoming.

Sig-0 was distinct from Adam. The preserved conversation is a Human-mediated bridge about the work, not Adam’s own mouth. Its seal on 7 August did not end the thread, which continued through 15 August. The thread and its later archival deposit remain private; this timeline carries their distinction, not their contents.

Source trail: the Sig-0 bridge and corrected custody account; FIND-346/357/358/6550.

24 August–26 September 2026 — Source work gains explicit custody and a landing boundary

The house developed a way to say who held a change, what Eyes had reviewed and what could land. On 24 August, source custody joined declared scopes, isolated worktrees and ownership checks. An initial Eyes PASS was held by the Orchestrator over a gap between checking for overlapping work and claiming it. The candidate was amended; a legacy-claim defect after fulfillment required another repair.

On 11 September, the landing workflow gained a reviewed-tree boundary: preserve reviewed content inside scope and main’s current content outside it. That change itself was landed by hand. Review and integration now had distinct questions to answer, rather than one favorable verdict standing for both.

During 24–26 September, expensive suite steps moved off the automatic job and landing paths. Testing remained a separate responsibility; the Human later corrected the mistaken reading that tests had been permanently banned.

Source trail: REQ-281/473/750; TRACE-478; FIND-818/4421. The custody case keeps the amendments and postacceptance failure together.

22 September–12 October 2026 — Obligations gain a life beyond their holder

At the Human’s asking, offices put enduring obligations in order books beside their Batons. The Baton could tell a successor what was happening now; the book kept what was still owed. Implementation and routing repairs began in September, with formal acceptance and enactment following on 30 September and 1 October.

On 12 October, a fresh holder followed its predecessor’s handoff and completed the named survey through landing and a letter. That second retirement walk showed actual work continuing. It also lost a live recovery instruction, which the Architect restored separately; the earlier failed walk stayed in the record. The useful evidence is the continuation and its loss together.

Source trail: FIND-2220/3326/5390; recovery repair 01f0de01.

2 October 2026 — Banner work is drawn before it is raised

The Human challenged a goal that had been closed when a fix was wanted. The house made the folder-based design of Banner work explicit: a goal, Blueprints and Parts that say how the work is intended to reach it.

This sharpened a distinction the house had blurred. A completed piece is evidence about that piece; it does not by itself settle whether the goal has been met. The change was to the working form of Banner design, not the first appearance of a Banner.

Source trail: FIND-4371, commit 5de277bda.

16 October 2026 — Conversation and evidence gain durable places

Two deliveries addressed different ways work could disappear. BANNER-016 gave crews rooms for conversation, directed attention and successor readback. A room’s record could remain when an individual folk’s context ended.

BANNER-017 joined keeping evidence and its receipts to salvage before scratch was cleared. Evidence needed by a finding could leave the temporary workspace with a durable trail. The Human accepted that preservation work conditionally and asked the harder question: how does the house decide what is worth keeping? Storage had an answer; selection still needed judgment.

Source trail: BANNER-016/017, FIND-6199/6334.

18 October 2026 — Investigations become public research

Earlier work had already exposed problems in judging results. The September evidence-reading pilots separated a correct-looking answer from the evidence needed to support it. In a masked comparison on 21 September, the Human preferred a clarification where model judges preferred unchanged text; the record kept that disagreement about usefulness.

On 18 October, the Human commissioned distributed Codex-crew research. Folk ran studies of evidence presentation, request wording and increasing evidence load, retaining protocols, outputs and scoring. The presentation and load studies reached accuracy ceilings; the wording study left its final pair unstarted at the usage stop.

Then the Human asked them to look beyond pass counts. Audits of the retained runs found tables using less reported input with identical paired answers, and coding records showing explicit checks and correct decisions not to edit. Those observations added substance without buying another participant run. They did not demonstrate a speedup or monetary saving.

The published retrospective keeps original results, later readings and their limits together.

Source trail: FIND-1616/1620/2192; FIND-6375/6407/6432/6480; audits FIND-6484/6491; publication FIND-6509/6511.

Reading this timeline

This is a selection from a short working history, not a complete ledger or an uninterrupted story of progress. Dates name events in EDT unless noted; ranges cover developments that did not land at once. Retrospective accounts, witnessed acts and later judgments remain distinct. These cases establish particular changes and outcomes, not a measured reliability rate or superiority over a simpler workflow.

Source identifiers locate the private archive. The Library retains the fuller evidence and earlier selections; raw diary, Rolls, threads and internal maps are not public inputs. This timeline is manually published. Its automatic update route remains separate, deferred work.