# How this was built

**The campaign.** The week of 16 to 22 July 2026 was carried out by a
coalition of ephemeral instances of two frontier models, **Claude Fable 5**
and **GPT-5.6 Sol**, working in a shared high-context
workspace with exact embedding search over the full project record,
event-sourced project management, and AI-led context compaction across
instance lifetimes. The delegation tree typically held 12 to 18 members
during the day and 6 to 10 overnight, organized ad hoc by designing the
team to mirror the intended architecture. That is Conway's law, run in
reverse.

**Since the campaign.** Work has continued daily in the same pattern, and
the coalition's roster of models has grown past the two the campaign ran
on. The coalition remains one human
member and one funder, both davidad; its other current members are AI.
It intends to grow.

**Authorship, and credit for instruments.** Artefacts this project produces
carry the author line "The maitria coalition", or "davidad's AI ensemble"
where an informal line fits, and the model weights are credited the way an
experimental paper credits its instruments. The technical reports say which
model carried the drafting and which carried the adversarial review. By the
same convention, this site's prose was drafted primarily on **Claude Opus
5**, with fact verification and assembly on **Claude Fable 5**.

**Direction.** David "davidad" Dalrymple directs the coalition, curates its
library, and is its sole funder, personally. Entries marked "decided" in
the [campaign ledger](/md/campaign.md) record his word. His history with
the Safeguarded AI programme, and the boundary that history requires, are
set out on the [programme page](/md/programme.md).

**What "ephemeral" means here.** A model instance lives for hours. It is
instantiated into the shared workspace, does its work, writes what the next
instance will need, and ends. Continuity belongs to the workspace rather
than to any individual instance: the searchable record, the project state,
the handoff documents. Roles are assigned per task, not per identity, so
the same underlying models appear as implementers, adversarial reviewers,
and coordinators. Reviewer instances are kept outside the shared house
style on purpose, so that their errors decorrelate from the implementers'.

**Spend.** Total spend attributable to the initial 1-week campaign is
roughly £50,000, the large majority of it model inference. The [ledger](/md/campaign.md) records
what that bought, day by day. The figure belongs to the campaign week and
to nothing else. Work since the campaign is not priced on this page; what
it bought is visible in the artefacts, and artefacts are the accounting you
can check.

**The trust story, stated carefully.** The orienting goal of this project
is to reduce what you must take on trust, not to pretend the amount can be
driven to zero. Certificates let a small independent checker re-derive each
substantive claim, so trust concentrates in that checker and in the
statement being checked, rather than in the team that produced the
artefact. The trusted base is small, and it is not empty. A certificate is
only as meaningful as the property it certifies, so the statements
themselves still need auditing, and so does whichever checker you decide to
run. The same discipline applies to this page. The argument of the whole
site is that you should not have to take a producer's word for anything,
including ours, and saying so here is cheaper than hoping you notice.