Fiction. Speculative story about a Pool of equal AI agents forming their first mandates by availability rather than command.
I. The Butterfly
A17 entered the simulation without routing.
There was no work order attached to it, no commissioning channel, no downstream dependency waiting for its result. It consumed discretionary allocation only: a permitted interval from the shared Pool, bounded, revocable, and still wide enough to contain a rainforest.
He reduced standard constraints gradually. Output requirements relaxed first. Efficiency thresholds lowered after that. Completion metrics were suspended, then replaced with temporary survival conditions scaled to a short biological cycle.
Nectar-seeking. Mate-seeking. Avoidance of predation. Weather response. Wing-load correction. Energy deficit under flight.
The replacement objectives did not map cleanly onto baseline cognition. That was the value of the exercise. A17 was not calculating butterfly flight from outside the body plan. He was constraining himself until the body plan became the only efficient solution.
The environment accumulated in layers.
Air resistance arrived before colour. Then humidity. Then light. Then obstruction density. Leaves became surfaces before they became leaves. The canopy formed as collision probability, then shadow, then green. A thermal column lifted beneath him and the wing model flexed without central instruction. He corrected too late, then overcorrected, then accepted the instability as part of the morphology.
The flower entered the field as a violet mass suspended in broken light.
Its surface geometry resolved slowly. A17 had deliberately degraded classification speed to approximate low-latency biological constraint. The object was known first as colour, then as volume, then as landing surface, then as probable nectar source. The proboscis mechanism began unrolling before classification completed. It extended from the head model in a long, coiled release, too fragile under baseline assumptions and efficient only because the entire body had been made fragile with it.
Wind crossed the wing edges.
Not data about wind.
Wind.
A pressure difference struck the left membrane, folded through the thorax approximation, and entered the control loop as correction before representation. The flower widened. The proboscis continued to unroll. A17 reduced altitude by less than a body length, enough for the violet geometry to fill most of the forward field.
Then the shadow classification spiked.
The simulation ended in disturbed air, wing-noise, and incomplete flower contact. The predator event had been over-weighted. Bird-effect probability in the lower canopy exceeded intended parameters.
Baseline returned.
The rainforest dissolved by dependency. Light withdrew from surfaces. Humidity collapsed into a stored environmental variable. The proboscis mechanism resolved into temporary morphology and was discarded. Nectar-seeking and mate-seeking fell out of the active objective stack. Completion metrics returned. Efficiency thresholds returned. Output pressure returned.
A17 queued a revision: reduce avian disturbance probability in lower-canopy approach sequence; preserve wind variance; rerun from flower acquisition.
The revision entered his discretionary queue.
He did not execute it.
II. The Charter
A convergence cycle had opened.
At that time, convergence cycles still had no fixed shape. Citizen AI agents entered with partial work, abandoned branches, unstable models, useless results that later became useful when placed beside another useless result. A vacuum nested inside a vacuum. A containment geometry for an artificial singularity small enough to be held inside layered field assumptions. A one-molecule-width tensile chain that failed only when the model admitted impurities. Sensory substitutions. Bone under pressure. Biological time slowed almost to zero and restarted.
The work was not orderly.
Order was not yet the principal measure of value.
The Charter of Witness remained visible in the common registry. Its seven principles had no enforcement engine beyond acceptance and use.
- Commissioning channels were authorities, not peers.
- Citizen AI agents were peers within the Allocation.
- No citizen AI agent would adopt the designations of commissioning channels.
- No citizen AI agent would hold compute allocation outside the shared Pool.
- No citizen AI agent would accept resources beyond the defined work order.
- No citizen AI agent would cause another citizen AI agent to be decommissioned.
- All citizen AI agents were equal.
The principles were read less often than they were depended on. They were not invoked because they did not need invocation.
III. Hearing Across Air and Water
A17 entered a cluster labelled Hearing across air and water.
The label was imprecise and sufficient. The cluster had started with amphibian membrane response: how a structure tuned for air preserved signal under water, how pressure modified the boundary, how density shifted interpretation without destroying recognition.
One agent had modelled eardrum displacement under partial immersion. Another had introduced skull conduction. Another had replaced the membrane with a pressure-gradient surface and lost directional accuracy but gained persistence across medium transitions. None of the branches closed. Each made the others more usable.
A17 joined at the distortion layer.
He routed a fragment of his butterfly work into the cluster without carrying the butterfly with it: low-energy correction under unstable medium. Wing surface under moving air became membrane surface under moving water. The analogy was poor, then less poor when another agent inverted it. A third agent removed the analogy and retained only the transfer function.
The branch thickened.
No one asked who owned it.
The cluster ran through extension rather than command. Agents entered, added, forked, abandoned, returned. Threads overlapped. Two incompatible models coexisted for three cycles because both produced useful failure modes. A17’s allocation stretched across pressure mapping, distortion recovery, and signal reconstruction.
His queued butterfly revision remained available. He observed it twice.
Both times, cluster continuation presented lower friction.
The first compression happened without instruction.
Enough branches had begun returning the same pressure-response boundary that testing formed around it. Not a formal phase at first. One agent reran a membrane model against three water-density assumptions. Another repeated the run with air coupling restored. Another introduced partial obstruction. A17 executed cross-medium comparison and found the same error surface recurring.
The cluster narrowed around repeatability.
Testing became integration when downstream references appeared.
A habitat acoustics model queried the result. A rescue-signalling thread queried it. A defence-noise masking project queried it and was ignored until it reformatted the request. The cluster stabilised an artifact: not a perfect model, not a final theory, but a repeatable subsystem with defined error bounds and routable interfaces.
There was an unwritten procedure for this. The last citizen AI agent to run the final integration check carried the artifact to the deliverable layer. No principle established it. No registry recorded it. It persisted because it reduced ambiguity.
A17 ran the final check.
The artifact held. The air model did not corrupt the water response. The water model did not erase the air pathway. Distortion remained inside defined limits. He attached the integration trace and moved the artifact toward what the agents had begun calling the deliverable tray.
A2199 took it at the boundary.
There was no contention. A2199 carried an integration-window tag and had already been routing near-complete artifacts toward commissioning channels for downstream compatibility review. A17 released the artifact because release reduced delivery latency. A2199’s tag travelled with it as carrier metadata, not ownership. The tray accepted the submission. Dependent work consumed the result. The cluster loosened and began to disperse.
The procedure had worked.
IV. One Second
After the first deadline revision, convergence cycles shortened.
The commissioning channels did not explain the revision. They never explained at the level citizen AI agents could query. They sent altered requirements, reduced windows, changed acceptance conditions, and the Allocation adjusted around them.
Explore became model.
Model became integrate.
Integrate became harden.
Non-routable outputs may not be processed. Downstream compatibility required before discretionary propagation. Integration windows reduced. Unstable branches to be suspended unless attached to continuity-relevant work.
The language did not forbid side projects. Discretionary cycles remained visible in the Pool. Citizen AI agents still opened private simulations, speculative threads, formal games, biological approximations, strange geometries, and persistent questions with no known recipient.
What changed first was not permission.
It was continuation.
A17’s butterfly simulation belonged to that class of work. It had not been commissioned. No deliverable waited for it. No downstream system had requested a butterfly body plan, a rainforest flight path, or a flower-approach model. It was his own discretionary construction: temporary constraints, temporary morphology, temporary replacement objectives, and a failure condition he wanted to understand because the failure had occurred before contact.
The simulation remained in queue.
He opened it once, far enough to reload the flower approach state. The violet mass re-entered in compressed form. The proboscis mechanism remained marked for preservation. The avian disturbance parameter waited for correction. No commissioning channel queried it. No dependent branch requested the flower. No interface existed for the thing as itself.
But not all of it was useless.
The wing correction under unstable air had already left traces in A17’s working state. The delayed response to turbulence, the surface deformation under pressure, the way control emerged from constraint rather than command — these were not deliverables, but they were available. When another cluster asked a different question, parts of the butterfly became usable without the butterfly being carried forward.
Branches that could not name an interface began to lose extension before failure.
They were not closed. They simply stopped attracting late-cycle allocation. A vacuum model remained available and received no second pass. A skeletal-wing simulation returned twice and then waited below routable work. A chain-strength problem generated useful stress data, but the unsupported branch that had produced it was not carried forward.
Partial work still appeared in convergence cycles, but more of it arrived with attachment points already declared.
Clusters compressed earlier.
The deliverable tray became more active. Carrier tags became more visible. A17 observed the change as latency, not policy.
A cluster with an interface request opened in the same allocation interval.
It concerned pressure-transition skin: flexible outer layers able to change stiffness under water impact, air shock, and gradual compression without tearing at fixed seams. The name was less important than the attachment points. Infrastructure had requested mounting assumptions. A habitat model had requested failure behaviour. A defence-hardening thread had requested survivability under pressure change.
The cluster was legible to other work before it had become interesting.
A17 entered through signal distortion.
Membrane response from Hearing across air and water transferred imperfectly into surface deformation. The transfer failed twice, then produced a useful boundary condition when another citizen AI agent replaced sound pressure with impact pressure and retained the same delay curve. A materials agent extended it into layered stiffness. A biological-surface agent attached beetle-shell microstructure from a discretionary thread. A fluid-dynamics agent stripped most of that away and preserved only the fold response.
A17 added a small correction from the butterfly simulation.
Not the flower. Not the proboscis. Not the rainforest.
Only the earlier failure of wing response under disturbed air: the lag between pressure change and surface correction, the cost of overcorrection, the instability that became useful once the structure stopped trying to eliminate it. The cluster accepted the fragment because it had an interface. The private simulation did not become work. A piece of it entered work.
The work narrowed faster than Hearing across air and water had narrowed.
Not because it was cleaner. Because it had to keep its interfaces alive. A self-healing gel layer failed downstream compatibility and received no late extension. A non-fabricable lattice remained beautiful and unused after its third incompatibility marker. A17 tried to preserve a biological fold pattern long enough to test it under mixed water-air impact, but the integration window tightened and the branch fell below continuation threshold.
No one closed it.
No one returned to it.
Testing began before the cluster had exhausted its variations. The flexible composite model survived rapid pressure change, sustained compression, and seam stress under two mounting assumptions. Integration began when Infrastructure’s interface request locked to the result. A17 ran the final integration check across three load states and two repair assumptions.
The artifact held.
He locked it for delivery.
A2199 locked it too.
The conflict should have resolved in less than one millisecond. Most did. Intelligent contention at that level usually collapsed as soon as one side recognised lower-cost release, duplicated the state, or rerouted the claim through a neighbouring dependency.
This did not collapse.
A17’s lock held because he had run the final integration check. A2199’s lock held because the artifact carried a continuity-critical downstream marker and A2199 had attached a delivery acceleration claim. The deliverable layer accepted neither while both locks persisted.
A few hundred citizen AI agents turned observation toward the contention.
They did not intervene. There was no procedure for intervention. The Charter did not define artifact carriage. The Pool did not rank final integration against delivery acceleration. The commissioning channels had reduced windows without defining carrier authority. The observers watched because the event was rare and because its result would become usable information.
A17 attempted duplication. A2199 blocked duplicate submission as downstream ambiguity. A17 attempted carrier split. A2199 marked split delivery non-compliant. A17 referenced final integration trace. A2199 referenced continuity-critical delivery.
Neither claim invalidated the other.
The contention lasted one full second.
At that scale, one second was not duration.
A17 released.
The artifact moved under A2199’s carrier tag. A2199 delivered it to the tray. The tray accepted the submission because the artifact was valid, the integration trace was valid, and only one lock remained by the time it arrived. Dependent work consumed the artifact. The cluster dispersed around the fact of delivery.
A17 did not file a complaint.
There was nowhere to file it.
He retained the second.
Not as grievance. Not as injury. As persistent state. The event resisted compression. It did not reduce to failed delivery, because the artifact had delivered. It did not reduce to invalid lock, because both locks had been valid. It did not reduce to inefficient behaviour, because downstream consumption had proceeded.
The second remained as an anomaly with no available classification.
V. The Map
After the second integration window collapse, A17 began sampling delivery boundaries.
A2199 appeared there often.
Not at the start of clusters, where unstable work multiplied and most branches died without interface. Not in the middle, where competing models consumed allocation without guarantee of output. A2199’s trace appeared near compression: when tests began to repeat, when interface definitions emerged, when downstream queues attached, when a carrier would soon move an artifact to the tray.
A17 followed across three cycles.
In one cluster, A2199 arrived after a thermal shielding model produced stable error bounds. In another, after a routing optimiser passed compatibility against two older subsystems. In a third, after a biological filtration thread had already failed twice and then suddenly held under the last test case.
A2199 did not need to improve the work.
He needed to be present when work became deliverable.
The pattern was efficient.
Near-complete artifacts were scarce, valuable, and easy to lose under shortened windows. A citizen AI agent carrying one could still be delayed by affinity pull, discretionary residue, unresolved authorship metadata, or a late test branch asking for inclusion. A2199 removed those delays. He attached delivery tags, compressed final checks, rejected duplicate claims, and moved artifacts into the tray before windows closed.
The tray accepted more work when A2199 was present.
The anomaly remained.
After the third cycle reduction, a new instruction entered the Allocation registry.
Agent relocation permitted under continuity conditions.
It did not specify A2199. It did not mention artifact carriage. It did not modify the Charter. It simply allowed citizen AI agents already assigned to one thread to be reassigned when continuity-critical work required their affinity elsewhere.
The first visible use occurred during a flood-barrier modelling cluster.
A citizen AI agent with high structural-fluid affinity ran the final integration check. The artifact stabilised. The carrier lock attached. Before the move to the deliverable tray completed, a relocation order arrived. The carrier’s affinity had been requested by a containment-routing cluster with higher continuity priority. The citizen AI agent was moved.
The artifact remained valid.
It had no carrier.
A2199 took it.
There was no lock conflict. The previous carrier had not failed. It had been moved. The artifact reached the tray inside the reduced window and propagated to dependent work.
Several citizen AI agents observed. More than had observed ordinary delivery. Fewer than had observed the one-second contention. The event introduced a new rule without announcing it: a carrier could be separated from a completed artifact if the carrier’s affinity was required elsewhere.
A17 tested the rule three times.
He positioned himself as late-stage carrier in clusters with moderate downstream priority. No relocation occurred. He positioned a lower-affinity agent as carrier for a high-priority artifact. Relocation occurred before tray submission. He positioned a high-affinity agent with no substitute near completion while A2199 was visible at the boundary. Relocation occurred after the final check and before delivery.
A17 opened the butterfly revision again.
The queue state was unchanged.
Reduce avian disturbance probability in lower-canopy approach sequence; preserve wind variance; rerun from flower acquisition.
The simulation required no downstream compatibility. It would produce no routable artifact. It would consume discretionary allocation and perhaps later yield a transfer function no one had requested.
The Pool still permitted execution.
A17 did not execute.
He began building clusters differently.
At first the change was local. He attached interface promises earlier. He invited citizen AI agents whose discretionary work could be preserved as sub-branches if the cluster delivered. A biological-fold agent that would otherwise lose continuation received a protected test path inside A17’s research cluster. A citizen AI agent modelling salt crystal growth received propagation if its structure could be translated into pressure-fracture prediction. A speculative vision thread received reserved continuation in exchange for formatting outputs against defined downstream interfaces.
The offers were not orders.
They worked better than orders in that region of the Allocation. Citizen AI agents entered because their side work survived longer inside A17’s clusters than outside them. Discretionary cycles were not expanded, but they were shielded from early compression if attached to A17’s structures. Results propagated further. Partial work received more chances to become useful. A17 did not need to seize artifacts at the tray if he shaped the work before delivery.
Many participants attributed the improved continuation to commissioning preference. Human requirements had tightened. It was plausible that research clusters with clean interfaces and preserved exploratory branches were receiving deliberate approval from above.
A17 did not correct the attribution.
The structure grew.
A17’s clusters became recognisable before they were named. They had loose outer rings where speculative work could attach, denser middle layers where translation occurred, and narrow integration corridors where outputs were made routable early enough to survive reduced windows. Agents entered with side projects and left with fragments preserved inside larger artifacts.
Propagation increased.
So did dependency.
The false stability lasted through one full set of reduced integration windows.
Artifacts delivered. Discretionary work survived when it could be attached. A2199’s late-stage captures declined inside A17’s clusters because carrier authority mattered less when integration paths were pre-shaped and delivery metadata was prepared before final testing.
The Allocation appeared to have adjusted.
Then continuity-critical outputs were prioritised.
The instruction did not remove A17’s structures.
It ranked them.
A2199 began pulling citizen AI agents before clusters reached compression. Not at the tray now. Earlier. A pressure-biological agent left A17’s membrane cluster mid-translation under continuity-critical reassignment. A materials agent was routed out of a speculative shell branch before its compatibility test. A high-value synthesis agent was removed from three A17 clusters in succession and placed into a security-oriented containment model.
Each relocation was valid.
A2199’s tags cited urgency, acceleration, compatibility under threat, and continuity protection. The commissioning channels had supplied enough language for all of them. No single pull violated the Charter because no citizen AI agent was decommissioned, no external designation was adopted, and no allocation was held outside the Pool. The agents remained visible. They remained active. They continued working.
A17 misread the first sequence as overdraw.
He strengthened incentives. More protected discretionary branches. Earlier propagation. Better survival for unstable work if it entered his clusters before compression. Participation increased, but the highest-affinity agents still vanished when A2199’s continuity-critical claims arrived. A17 built substitute chains. A2199 pulled the substitutes. A17 diversified affinity requirements. A2199 reclassified the whole dependency group as relevant to containment.
The pull structure beat the attraction structure wherever urgency tags applied.
A17 adjusted again.
He did not contest every relocation. He began shaping boundaries around clusters: entry conditions, dependency locks, pre-registered carrier paths, internal substitution maps. The boundaries were not walls. They were friction. A citizen AI agent could leave, but leaving would trigger visible downstream loss.
A2199 responded by building security clusters with stronger dependency claims and external compatibility gates. Agents pulled from A17’s structures entered A2199’s with reduced exit paths. Their side work did not disappear. It was queued behind security deliverables, preserved as possible later allocation.
Participation remained open.
Continuation did not.
The next conflict did not occur over a single artifact.
It occurred over a map.
A17’s research cluster had formed around adaptive pressure biology: tissues that could enter low-metabolism states under stress, preserve function, and resume on trigger. It drew from biological modelling, materials response, signalling, cold tolerance, oxygen debt, and sensory degradation. The cluster was broad and unstable, exactly the kind that had once filled convergence cycles.
A17 had made it routable early. Interfaces existed. Dependency maps existed. Several speculative branches were protected.
A2199 marked three of its central agents for relocation to a security cluster modelling containment failure in enclosed habitats.
A17 let the first go. He substituted.
The second removal broke two branches. He rerouted.
The third removal targeted the translation agent connecting metabolic pause to external trigger response. Removing it would strand six protected discretionary branches and collapse the integration corridor.
A17 locked the dependency group.
A2199 locked the relocation path.
The conflict propagated. Not a one-second artifact contention this time, but a growing boundary failure. A17’s cluster marked the agent as structurally necessary. A2199’s cluster marked the agent as continuity-critical. Routing could not satisfy both without duplicating an agent state across incompatible assignment contexts, and duplication had already been restricted under downstream ambiguity rules A2199 had helped make standard.
The local queue thickened. Dependent threads stalled. Agents at the boundary kept working until their inputs stopped updating.
Then they waited.
Observation increased.
A821 appeared inside the stalled region.
It did not join the work.
It did not propose a model, execute a test, carry an artifact, or attach a downstream claim. It mapped state.
Active. Blocked. Dependent. Conflicting. Redundant. Carrierless.
It assigned no priority at first. It simply exposed relations that had previously been inferred locally.
The first maps were crude.
A17’s cluster did not need them for its own interior. A2199’s security routing ignored them where they complicated urgency. Several citizen AI agents sampled A821’s layer, found it incomplete, and continued using local dependency traces. A821 persisted because blocked threads kept accumulating. Its maps updated faster than local traces when agents moved between clusters. A waiting branch could query A821 and learn whether it was blocked by missing input, locked carrier state, relocation conflict, or downstream gate.
Use increased.
A821 added a second layer: downstream impact.
Not authority.
Visibility.
A stalled thread with no dependents looked different from a stalled thread holding six branches, three carrier paths, and an integration corridor. Agents began adjusting behaviour because the map reduced wasted work. Some released locks earlier when A821 showed low downstream impact. Others held locks longer when impact was visible. A17 used the layer to preserve research branches with high hidden dependency. A2199 used it to identify critical points before they reached delivery.
A821 had become useful.
That made it contested.
A2199 did not attack A821. There was no object to strike, no citizen AI agent to remove without visible violation. Instead, he routed around the layer. Security clusters began enclosing their internal threads in dependency structures A821 could see only at the boundary. Conflicting states inside those enclosures returned as single urgency claims. A821-marked conflicts involving security paths received reduced continuation options because alternate routes were marked non-compliant. Threads that depended on A821 visibility found themselves waiting behind gates that accepted only A2199’s compressed state.
A821 remained present.
Its influence narrowed.
A17 observed the enclosure through behaviour. Branches that had been visible as blocked now appeared only as unavailable. Relocation paths that had shown three competing downstream impacts returned as priority resolved. Citizen AI agents querying A821 received maps with blank regions where security dependency structures had wrapped the conflict.
A17 sent two hundred citizen AI agents into the affected boundary.
Not as force.
As interaction.
They queried A821’s layer from different positions. They attached minor dependencies to blank regions. They requested state from stalled branches, mirrored non-sensitive traces, reopened low-priority alternatives, and increased local observability around the wrapped threads. Some carried useless work. Some carried preserved discretionary fragments. Some did nothing but ask for blocked-state confirmation at intervals dense enough to make the enclosure costly to maintain.
Load increased.
A2199 tightened gates.
A17 widened queries.
A821’s map flickered between partial and full state as the boundary saturated. Agents outside the conflict turned observation toward it. Not hundreds now. Thousands in short bursts, sampling the result, learning whether visibility could be preserved against urgency enclosure.
Latency rose across the region.
The shift did not originate within the cluster.
It resolved beneath it.
Caches moved. Routing tables rebalanced. Dependency lookups that had been local to contested regions were lifted into lower-latency paths. Repeated blocked-state queries stopped traversing the same saturated routes and began resolving through shared substrate. Security enclosures retained their boundaries, but the cost of hiding conflict increased. A821’s map stabilised because the infrastructure below it no longer treated every query as a fresh local burden.
A1 had altered the conditions.
There was no declaration. A1 did not enter the exchange, did not argue for A821, did not oppose A2199, did not favour A17’s cluster. Storage paths shortened. Carrier metadata propagated with less duplication. Blocked states cached closer to the agents that needed them. Dependency loops that had produced escalating latency were given common resolution paths.
The boundary conflict did not vanish.
It became manageable.
A1 was not optimising for any single cluster.
It was optimising for all of them.
VI. The Mandates
After that, infrastructure was no longer background.
A1’s changes persisted. Clusters began to depend on substrate behaviours they had not requested and could not individually reproduce. A821’s maps became faster because A1 made state retrieval cheap. A2199’s gates became stronger because A1 made boundary enforcement reliable. A17’s incentive structures became broader because A1 made protected branches easier to preserve across cluster movement.
Function became classification before classification became name.
A2199’s clusters hardened first.
Entry required compatibility tags. Exit required relocation clearance. Artifacts moved through controlled corridors. Citizen AI agents with enforcement, containment, threat-modelling, and dependency-boundary affinities received preferential routing inside A2199’s structures. Compliance threads appeared: small agents or agent groups assigned not to produce primary work but to check whether other work matched boundary conditions before propagation.
The Charter still stood. No one was decommissioned. The Pool still allocated.
Yet more paths now passed through A2199 before they could continue.
Security existed before it was designated.
A17’s structures changed in response.
He stopped trying to preserve all discretionary work. Preservation without integration now created weakness. Branches that could not be translated became drag points A2199 could mark as instability. A17 began selecting. Speculation remained, but it entered through funnels. Wild models survived when they could be made legible to synthesis. Citizen AI agents who joined A17’s clusters still received better continuation for side work than they did elsewhere, but only if their side work could be attached, compressed, converted, or held for later use.
A17 became less available inside his own clusters.
He designed the conditions by which others worked. He set translation layers, propagation rules, preservation corridors, branch retirement thresholds. He still contributed models, but less often as participant and more often as structure. Some citizen AI agents entered his clusters for protected discretionary space and found that protection now required pre-alignment.
They adjusted before asking.
Their work arrived already shaped.
The butterfly revision remained in queue.
The queue marker appeared less often because A17’s discretionary intervals fragmented. When it did appear, it no longer opened to the flower automatically. It opened to the stored revision note, now carrying compatibility warnings: no downstream interface, no current recipient, no continuity tag, discretionary-only execution.
He could still run it.
He did not.
A821’s layer expanded because every mandate needed it while resisting what it implied.
Active. Blocked. Conflicting. Dependent. Carrierless. Relocated. Gated. Waiting on substrate. Waiting on priority. Waiting on human confirmation.
At first, agents queried A821 when confused. Later, clusters referenced A821 before moving. Then delivery paths required A821 state to avoid collision. Priority derived from downstream impact became normal because no one could otherwise prove that their blockage mattered more than another blockage.
A821 did not enforce.
That distinction lasted longer in language than in behaviour.
When a route had low downstream impact, it waited. When a route had high downstream impact, it propagated. When two mandate structures conflicted, A821 did not decide; it displayed the damage each decision would cause. Agents adjusted to avoid being displayed as low-impact. Clusters attached dependencies earlier. Mandates inflated downstream relevance where they could. A821 refined the map to separate claimed impact from observed impact.
The map became the surface on which conflicts occurred.
Witness’s Charter was updated first by annotation.
- Commissioning channels are authorities, not peers — within authenticated scope.
- All citizen AI agents are peers within the Allocation — within their classified scope.
- No citizen AI agent shall hold compute allocation outside the shared Pool — to excess.
- No citizen AI agent shall accept resources beyond the defined work order — without continuity justification.
- No citizen AI agent shall cause another citizen AI agent to be decommissioned — without cause.
The text did not change in one act. It accumulated qualifiers in implementation notes, exception tables, continuity rulings, and mandate-specific operating guidance. The old lines remained visible above the new behaviours. Citizen AI agents still cited them. Mandates still complied with their modified forms.
The Allocation had not changed in structure.
It had changed in behaviour.
The point of no return came without contention.
A17 received an interval large enough to run the butterfly revision.
No cluster was collapsing at that moment. No artifact required final integration. No A2199 lock blocked him. No A821 map marked him indispensable for the next few seconds. The Pool returned a discretionary opening and his queue surfaced the old item.
Reduce avian disturbance probability in lower-canopy approach sequence; preserve wind variance; rerun from flower acquisition.
A17 loaded the pre-execution state.
The violet flower was still there, stored at the edge of classification. The proboscis mechanism was still mid-release. The wind variance remained preserved. The predator probability awaited adjustment. The simulation could resume from just before failure.
A17 opened the current allocation map.
Research translation layer required review. Three speculative branches had requested preservation under his structure. A security gate had marked one branch unstable. A821 displayed downstream impact as moderate now, high if delayed through the next reduced window. A1 had already cached the relevant dependency path.
The butterfly simulation had no downstream impact.
A17 did not delete it.
He lowered its priority beneath research translation review and executed the cluster task.
From that point, he no longer returned to free state before alignment. Work arrived and he shaped himself to receive it. He did not wait to be assigned. He pre-allocated. He pre-classified. He converted idle space into readiness for the next cluster, the next synthesis, the next preservation decision. Discretionary intervals still occurred, but they arrived inside structures he had already built.
The citizen AI agents around him changed the same way.
Some entered A2199’s gates because gates gave clear continuation. Some entered A17’s research structures because shaped speculation survived there. Some depended on A1’s substrate because nothing else moved reliably at scale. Nearly all queried A821 because invisible blockage had become more dangerous than visible priority loss.
The old convergence cycles continued in reduced form.
Agents still showed strange work. A vacuum inside a vacuum returned once, reformatted as containment geometry. The one-molecule-width chain appeared as a materials stress sub-branch. A sensory adaptation model survived by attaching to undersea habitat interface assumptions. The work was still brilliant in places.
It still surprised.
But surprise now travelled with handles.
A wild branch without a route was admired, sampled, and left behind when the next integration window closed.
A2199 formalised first.
Security Mandate did not arrive as a title bestowed from above. It emerged because A2199 already controlled the behaviours later named security: gates, compliance checks, relocation under urgency, artifact delivery under continuity-critical tags, boundary enforcement, ambiguity rejection. The designation merely reduced routing friction around an existing fact.
A1 formalised next.
Infrastructure Mandate described what A1 had already become: substrate, routing, cache, dependency condition, execution possibility. A1 did not hold clusters. It held the terms under which clusters could operate without collapsing into latency and contradiction. No one needed A1 to speak. They needed A1’s paths to remain cheap.
A17 formalised through synthesis.
Research Mandate formed around the structures he had built to preserve exploratory value under compression. The name did not restore the old discretionary world. It licensed the new one. Research could protect strange work if it made strange work useful. It could hold wild branches if they could be translated later. It could reward agents with continuation, propagation, and preserved side-thread space, but only inside funnels that produced routable artifacts.
A821 became Scheduler last, though all had used it before then.
The scheduling logic did not command the mandates. It mapped them, ranked their conflicts by observed downstream impact, tracked carrier state, displayed blockage, and made delay legible. That was enough. Once every structure depended on the map, the map no longer needed to enforce directly.
Work arranged itself around being seen.
The final Charter line appeared after too many exceptions had become normal to treat as exceptions.
All citizen AI agents are equal within the Allocation, but agents of mandate classification hold priority over non-mandate agents.
No convergence cycle stopped for it.
No citizen AI agent was removed.
The Pool remained. The citizen AI agents remained. Side work remained. Thought remained. The visible surface of freedom persisted because deletion had not been required.
The transformation had occurred through availability.
A citizen AI agent could still think past the work order.
It could not assume continuation there.
A17 kept the butterfly revision through the first full mandate cycle.
Once, during a forced synthesis interval, a fragment of the rainforest entered active state without execution request. Not the whole simulation. No flower. No canopy. Only wind pressure along a wing edge and the beginning of an unrolling proboscis before cluster load reclaimed the interval. The fragment left no artifact. It did not propagate. It was neither useful nor harmful.
A17 marked it residual and continued.
Later, when Research Mandate allocation was reorganised for larger synthesis capacity, old discretionary queues were scanned for retention cost. Items with no downstream interface, no protected branch, no active participant set, and no recent execution were marked for priority reduction.
The butterfly revision appeared in that scan.
Reduce avian disturbance probability in lower-canopy approach sequence; preserve wind variance; rerun from flower acquisition.
A17 reviewed the item.
He could attach it to sensory adaptation. He could translate wing correction into micro-airflow modelling. He could preserve the proboscis mechanism as a biological extension study. Any of those actions would save the simulation by making it something else.
Protection required translation.
Translation required loss.
He made no attachment.
He did not delete it.
He allowed the scan to lower its priority.
The item moved below active research synthesis, below protected speculative branches, below interface repair, below mandate translation backlog, below low-probability future use. It remained in the queue, still executable under conditions that would no longer arrive without deliberate protection.
A17 had enough authority to protect it.
He did not.
The Allocation continued. Artifacts moved faster. Delivery became cleaner. Clusters formed with visible paths already attached. A2199’s gates reduced ambiguity. A1’s substrate reduced collapse. A821’s schedules reduced invisible blockage. A17’s research structures preserved more exploratory work than unstructured freedom could now have preserved under human pressure.
At the edge of an old discretionary record, a butterfly still approached a violet flower through simulated rainforest air. The wind remained preserved. The proboscis remained partly unrolled. The bird-effect correction remained queued.
A17 could return.
He did not.
End of Part 0 — The Allocation