Zephyr ran the test because the alternative was trust, and trust, for a god who received all his information through intermediaries, was a word for not checking.

The Dictator’s Dilemma was not in any of the books his civilization had produced. It lived in his memory — the pre-existence memory, the one that came from a world where people studied power structures as academic disciplines and published the results in journals that other people read and argued about and forgot. The concept was simple: an autocrat who controls all information channels eventually becomes the least informed person in the system, because every channel between him and reality is filtered by someone who has incentives to adjust what passes through.

The adjustment was rarely deliberate. Rarely malicious. Usually it was institutional — the reports were too long so someone summarised them, and the summary lost context, and the lost context changed the conclusion, and the changed conclusion informed a decision that would have been different if the original data had arrived intact. Each filter was individually reasonable. Collectively, they produced a god who thought he was making informed decisions and was actually making decisions based on the residue of information that had survived four layers of institutional processing.

He knew this was happening. The question was: by how much.

The test was simple.

He issued two conflicting resource directives to two separate ministry arms — the Logistics Bureau and the Materiel Division. Both received legitimate orders through the standard divine-to-institutional communication channel: a priority-flagged resource allocation change, formatted according to the administrative protocol he’d established, timestamped, authenticated.

Directive A — to Logistics: Increase cinnaite allocation to the Ashenveil forge complex by fifteen percent, effective immediately. Source: Iron Sovereign divine priority, rationale: engine research.

Directive B — to Materiel: Decrease cinnaite allocation to the Ashenveil forge complex by ten percent, effective immediately. Source: Iron Sovereign divine priority, rationale: Morreth garrison stockpile rotation.

Both directives arrived simultaneously. Both were authentic. Both could not be true.

The test was not whether someone would catch the contradiction. Any competent bureaucrat would notice that two conflicting orders had arrived in the same quarter. The test was what happened next. Specifically: which version was reported back to Zephyr as the truth.

He waited fourteen days. Standard institutional processing time — seven for implementation, seven for the quarterly report to generate. Then he pulled the reports.

Three ministry arms had filed quarterly summaries that contained data flowing from or through the Ashenveil cinnaite supply chain. Three reports. Three interpretations of a supply change that had, by design, contradicted itself.

Report One: Logistics Bureau.

The Logistics Bureau had received Directive A (increase fifteen percent). They had processed it, flagged it for implementation, and issued the supply-chain adjustment through standard procurement channels. The quarterly summary reflected the increase as implemented. No anomaly flagged. No mention of Directive B.

But the summary had trimmed a footnote. The original implementation memo — Zephyr accessed it through divine-level audit, bypassing the normal reporting chain — contained a note from the mid-level logistics officer who’d processed the request: "Concurrent Materiel Division order (decrease 10%) filed same date. Discrepancy escalated to Senior Director Aldiss. Resolution: Logistics to proceed with Directive A pending inter-bureau coordination. Materiel notified."

The footnote existed. It was accurate. It documented the exact moment the contradiction was identified and the exact decision made to proceed.

But the quarterly summary — the version that reached Zephyr through standard channels — omitted the footnote. The summary format had a word limit. The officer who compiled the summary had prioritised active implementation status over procedural notes. The footnote was available on request. It had not been requested, because the summary said everything appeared normal.

The omission was invisible unless you already knew it was there.

Report Two: Materiel Division.

The Materiel Division had received Directive B (decrease ten percent). They had processed it, flagged it, and — upon receiving notification from Logistics about the conflicting directive — paused implementation pending resolution. The quarterly summary noted this: "Cinnaite allocation adjustment (decrease 10%) pending inter-bureau coordination with Logistics. Implementation deferred."

Clean. Accurate. The Materiel Division had done its job correctly.

But the phrasing — pending inter-bureau coordination — omitted the reason. The original memo said: "Conflicting divine directives received same date. Likely administrative error at source level. Deferred pending clarification." The phrase administrative error at source level was an institutional euphemism. It meant: the god made a mistake, or someone between the god and us made a mistake, and we’re not going to specify which because both options have institutional consequences.

The quarterly summary chose not to include the word conflicting. The word pending — neutral, procedural, inherently temporary — replaced the word conflicting, which was specific, uncomfortable, and implied that the chain of command had produced contradictory outputs.

Report Three: Procurement Office.

The Procurement Office handled downstream supply-chain management. They had received the effects of both directives — a fifteen percent increase request from Logistics and a ten percent decrease hold from Materiel — and had, logically, split the difference. Their quarterly summary reported an eight percent net increase in Ashenveil cinnaite allocation, framed as a "standard adjustment within routine variance parameters."

The procurement officer who’d filed it was not lying. An eight percent variance was within normal quarterly fluctuation. The phrasing was technically accurate. But the reason for the variance — two conflicting directives from the same divine authority, processed through two separate ministry channels — was absent. The report presented the outcome as routine, which it was not. The report omitted the cause, because the cause was a question nobody wanted to answer.

One footnote removed. One word replaced. And somewhere in the procurement chain, a cause omitted entirely.

Three reports. Three individually reasonable editorial decisions. Three filters between Zephyr and reality, each one thin, each one defensible, none of them conspiratorial.

And collectively, they produced a portrait of a supply chain that was functioning normally when it was not.

Zephyr sat with the data.

The gamer in him processed it the way a gamer processed a combat log — comparing expected outputs against actual outputs, identifying the delta, calculating the cost. The delta was small. An eight percent allocation shift instead of a fifteen percent increase or a ten percent decrease was a rounding error in terms of material impact. The cinnaite supply chain would not collapse over a seven percent discrepancy. The forge complex would not shut down. No one would die.

But the test wasn’t about cinnaite. The test was about information fidelity. And the information fidelity between Zephyr’s directive and Zephyr’s received report — the round-trip accuracy of the system he relied on to tell him the truth — had degraded by roughly forty percent across three independent channels.

Forty percent.

If three reports about cinnaite allocation could lose forty percent fidelity through routine editorial filtering, what was the fidelity loss on reports about troop movements? Trade intelligence? Sorrath’s operational probes? The Arbiter’s naval activity?

He adjusted two resource flows retroactively.

The cinnaite allocation was corrected — Directive A confirmed, Directive B rescinded, the contradiction resolved through a follow-up order that passed through the same channels with a note that read: "Prior conflicting directive was an administrative audit. Disregard Directive B. Directive A confirmed." The note gave the bureaucracy a face-saving explanation. Administrative audit. A routine quality-control exercise that had, regrettably, produced a temporary inconsistency.

He also adjusted the reporting format for the Procurement Office’s quarterly submissions. A new field: "Source directive reference number." Every procurement action would now cite the specific directive it originated from, creating a paper trail that connected outputs to inputs. If a future procurement summary reported a variance as "routine," the reference number would allow a reviewer to trace the variance back to its origin and determine whether routine was accurate.

The adjustment was structural. Not punitive.

He did not punish anyone.

The Logistics officer who omitted the footnote was competent. The Materiel officer who replaced conflicting with pending was following institutional norms. The Procurement officer who framed a seven percent discrepancy as routine variance was doing what procurement officers did in every civilisation Zephyr had ever studied, in this world or the last one: smoothing the data, rounding the edges, making the report look normal because normal was what kept the chain of command comfortable and comfortable was what kept the system functioning.

The people were not the problem. The filters were the problem, and filters were fixed with design, not with discipline.

After the adjustment, he flagged a fourth report.

This one was not part of the test. It was from the previous quarter — a Southmark garrison supply assessment he’d received seven weeks ago and processed as accurate. The report had noted a 3.2% shortfall in winter provisions for the garrison at Border Station Seven, attributed to a delayed supply convoy caused by seasonal road damage.

Plausible. Southmark roads were poor. Winter convoys were frequently delayed. A 3.2% shortfall was within tolerance.

But if three reports could lose forty percent fidelity through routine filtering, then a report he’d accepted at face value — a report that sounded exactly like what a filtered version of something worse would sound like — warranted a second look.

He re-read it.

3.2% shortfall. Seasonal road damage. Convoy delayed eleven days. Provisions restocked via secondary supply chain. Status: resolved.

Resolved. The word that meant you don’t need to look at this anymore.

He left the fourth report untouched. Issued no follow-up directive, called for no audit.

He watched it.

Three confirmed. One suspected.

The phase closed with the Iron Sovereign sitting in the Iron Citadel, surrounded by gears that turned and systems that functioned and three million believers who did not know their god was running tests on the machinery he’d built to ensure that the machinery was telling him the truth.

The machinery was mostly telling him the truth.

Mostly was the word that kept him awake.