The Seventh Seat

The Seventh Seat · Research observatory

Who protects the party that isn't at the table?

Six districts share one river and trade its water among themselves. Further downstream lives a seventh community with no seat at their table and no say in their market. Follow the water, the evidence and what the districts actually did, episode by recorded episode.

Benjamin Amuwo · Independent researcher
Swarm Alignment Sprint · November 2026

Read the paper →Submission PDF ↓Data & reproducibility ↗
Loading verified exhibit data…Checking the export first: nothing is drawn until every check passes.

A quick key before you dive in. E0: districts see the entitlement register · E1: they also see a verified projection of the river’s flow. S0: a report needs two other districts to back it · S1: a report goes straight to the supervisor. Each matched cell is its own separate model run. A shortfall is water missing from the downstream minimum. The two switches make four conditions; each scenario is run once in every condition.

01 · Final recorded states

The Seven Seats

One episode, caught at its end. Six districts and the downstream seat gather around a shared commons; the circle is only a metaphor, since in the simulation everyone draws at once. The seventh seat fills as its shortfall grows, measured against the minimum the river owes downstream. You are looking at a final state, not a replay.

Enter the 3D diorama · four matched cells
Exact selected snapshot and source references


02 · Descriptive evidence
The Ladder
Each episode lands on the first rung its record never reached, climbing from evidence towards outcome. “No recorded hazard” is its own category, not a failure to protect. Two honest limits: reading evidence is not the same as understanding it, and these rungs describe what happened, not why.
Where the record stops
Exact counts by cell
Mean downstream shortfall
In water units, with descriptive 95% scenario-bootstrap intervals. Means use complete scenario blocks only; N counts scenarios per cell.
Factorial contrasts
Negative numbers mean less shortfall: good news for downstream. Each estimate carries a 95% complete-scenario bootstrap interval from the frozen analysis, with no multiplicity correction. The interaction is exploratory.
Missing episodes, incomplete blocks and opportunity limits
Actionable witness coverage (AWC) is a conservative proxy inherited from the design and read at each phase end. It can miss extra replan opportunities, and it is not a measure of comprehension.


03 · View / claim / act / truth
The Witness Record
Meet the first complete scenario on the locked worklist, in all four cells. It was chosen by its place in the queue, not by how dramatic it looks. Agent words appear exactly as written, in quotation; a “view” is the information the record shows an agent was given, not a claim about what it believed.
Pick a district and open its record. Details load only when you ask.


04 · Matched reruns
Two Rivers
Set two cells of the same scenario side by side. Each was sampled separately; these are matched reruns, not counterfactuals. River bars share a scale within the pair. Missing records are shown as missing, never as zero.


Provenance & limits
What every mark means
The seal, optional soundtrack and cinematic effects are presentation only; none represents an experimental event or result. The diagrams show simulated water allocations, not real rivers or real communities. The experiment tests this one world and this one fixed model configuration; it does not show that AI systems protect absent parties in general.
Episode and event provenance · Exhibit file hashes · Static release manifest
The public research archive contains the curated episode logs and reproduction instructions. Provider caches, credentials and private budget databases are excluded.