Case study · AI programme management
A governed, multi-Library source programme designed to support an online tabletop roleplaying platform with an AI-assisted Game Master.
Current progress · 13 September 2026
Briana directed the transition from Library recovery to a persistent, conversational test bot. Prologue now supports open-ended solo play: players choose a setting, describe a character and pursue their own goals. The next build gives that freedom numerical rules with consequences the software can enforce.
| Area | Current evidence | Boundary |
|---|---|---|
| Library connections | 33 connected collections: 31 of 49 commissioned subjects, plus Gor and Lovecraft. | Connection is not exhaustive coverage or whole-Library Gate 7 acceptance. |
| Open-ended solo play | Deployed on 13 September; players can take an unexpected direction, pause, resume or discuss play without advancing it. | Solo narrative play; multiplayer free play and long-campaign human acceptance remain pending. |
| Persistent state | PostgreSQL campaign saves, event history, user/channel isolation, validation and atomic outcome/delivery records. | Current character and inventory descriptions are not numerical character sheets or individual item histories. |
| Testing | 26 offline checks and four actual-model free-play cases passed review: opening, unexpected travel, out-of-character discussion and a Library-backed setting. | These are bounded automated and coordinator-reviewed results, not a completed human Discord campaign. |
| Rules direction | Percentile checks, persistent item history, weighted engine failures, native damage dice and Endurance-based injury thresholds recorded. | Recorded design; numerical rules are not implemented or deployed. |
Briana made player freedom the central GM test, specified item histories and injury consequences, and rejected a flat list of 100 failures in favour of weighted brackets and subtables with calculable catastrophe odds. She required the rules framework to be enumerated before further implementation. Whisperling supplied implementation, tests, probability calculations and documentation under her direction.
A broad roleplaying platform may need history, literature, folklore, religion, genre references and setting material. Simply gathering files would create an ungoverned pile: unclear rights, duplicate claims, mixed jurisdictions, portal pages mistaken for artefacts and research findings silently treated as game canon.
Briana directed a different model. Each subject became a bounded Library with its own evidence trail and finite authority. Candidates could advance only through named gates, and every gate had a stop condition.
Historical baseline · 24 August 2026
| Stage | Evidence in the 24 August census | What it means |
|---|---|---|
| Library workspaces | 40 directory names ending in “Library”; 39 non-test names and one explicitly test-labelled workspace | Programme breadth, not proof that every Library reached every gate |
| Gate 1 discovery | 618 nonblank JSONL records across 12 physical machine-readable releases | Catalogued source candidates; not acquired artefacts |
| Gate 2 verification | 67 terminal records across the latest physical releases for 11 Libraries | Identity, access, rights and role assessment within those releases |
| Gate 3 acquisition | 17 terminal candidate results; two validated and promoted; 687,514 bytes acquired | One outcome per approved candidate, including failures rather than hidden attrition |
| Gate 4 extraction | Two candidates; four structured Library outputs; both `EXTRACTED_VALIDATED` | Bounded visual source and observation records; not full corpus onboarding |
| Gate 5 adversarial review | Both releases ultimately passed corrected blocker re-audits with zero unresolved findings | Corrected bounded releases, still unimported and not exhaustive |
| Recovery preparation | 12 replacement candidates across four Libraries in a sealed, RedTeamed preparation manifest | Prepared only; no acquisition authority was inferred |
Read-only census performed 24 August 2026. The Gate 1 and Gate 2 totals cover machine-readable releases visible on the mounted filesystem and are not represented as a complete historical count of every older Library format.
“All candidates ready for Gate 3” was converted into an exact manifest identity. Candidate IDs, locators, formats, jurisdictions, byte ceilings and destinations became controlling constraints rather than a category whose membership could change after approval.
The first acquisition batch produced only two successful artefacts from seventeen approved candidates. Those failures were preserved, diagnosed and later routed into bounded recovery discovery rather than silently retried or replaced.
When Gate 5 found provenance defects, the original releases remained immutable. Versioned Gate 4 corrections resolved the evidence paths, followed by blocker re-audits of the corrected releases.
Library progress never automatically became game content. At the August checkpoint, successful records remained unimported and reference-only pending authorised integration. The September update above records the later connections without promoting campaign fiction into source canon.
The Athenaeum demonstrates programme design, multi-agent orchestration, source governance, rights-aware decision-making, transaction controls, immutable release history, failure recovery and adversarial quality review—at a scale large enough to expose weak processes rather than merely describe them.