What Problems Are We Solving?DOI: 10.XXXX/XXXXXXX.XXXXXXXConference: 32nd Conference on Pattern Languages of Programs, People, and Practices; October 12–15, 2025; Skamania Lodge, Columbia River Gorge, Washington, USA.ISBN: 978-1-941652-22-0DOI: 10.64346/PLoP2025p24CCS: Software and its engineering Software creation and management Software development process management Software development methods Design patternsCCS: Computing methodologies Artificial intelligence Distributed artificial intelligence Intelligent agentsCCS: Applied computing Operations researchCCS: Computing methodologies Modeling and simulationConference: ; ;

An apparatus built to organise effectively must carry, from the outset, the three properties that stop it becoming an authority: an open model, an available exit, and a verdict it cannot issue to itself. Effective organisation and unaccountable apparatus are the same construction seen from two sides; only those properties separate them. If Martin Luther King Jr.: “Those who love peace must learn to organize as effectively as those who love war.” The organisational forms that win are not neutral instruments that happen to have been used for war; they are what effectiveness looks like — a shared picture maintained in real time, a chain of command, an instrument for testing tactics before they cost anything. Refusing them on the grounds of their history is a decision to be less effective, and it is usually taken by people who can afford to be. However Gilles Deleuze: ‘‘we want a war-machine that would not recreate a State apparatus, a nomadic unity in relation with the Outside, that would not recreate the despotic internal unity.’’ The same forms that make an organisation effective make it an apparatus: the shared picture becomes the only admissible picture, the chain of command becomes the only admissible authority, and the instrument for testing tactics becomes the instrument that certifies its own results. An apparatus that has captured its own verdict cannot be argued with from outside, and no longer needs to be.The case that forced it. The ruling was written down only after the stack had already lost property (c) in four places at once, none of them decided. The War Machine’s outer loop proposed the same kind of action 108 times, treated every proposal the operator did not follow up as evidence against it, and collapsed that action’s learned credit — a number measuring how often the lab proposed, not what the operator thought of it. (In the loop’s own terms: 108 emissions of one action class, scored by intrinsic value.) The capability ledger recorded a single invoice and a repeatable revenue capability with the same :satisfied status. Six strategic surfaces asserted continuous currency at a median age of thirty-six days. Each is a verdict the apparatus issued to itself, and each looked like traction from inside. Then Build the organising apparatus, and build into it the three properties that keep it from closing: (a) An OPEN MODEL. Active inference is chosen here for a structural reason rather than a fashionable one: the generative model is declared rather than implicit, its objective is written down and inspectable, and expected free energy carries an epistemic term — the apparatus is required to value being surprised. A State apparatus values compliance; a model with an information-gain term is one that pays for evidence against itself. (b) An AVAILABLE EXIT. Open source, forkable, with the substrate legible enough to leave with. A nomadic unity is one that can be left; the test of whether an apparatus is despotic is not how it treats members but what it costs to stop being one. (c) A VERDICT IT CANNOT ISSUE TO ITSELF. Every claim the apparatus makes about its own success must be movable only by evidence it did not manufacture (aif/no-self-certification). This is the property most easily lost, because losing it feels like progress: internal follow-through is abundant, cheap, and fast, and it looks exactly like traction until something external is asked to agree. Because An organisation that organises effectively, cannot be audited from outside, cannot be left, and grades its own work has become the thing it organised against — and will not be able to tell, because the instrument that would tell it is the one that was captured. The three properties are not ethical decoration on an engineering artefact. They are the difference between a war machine and a State apparatus, and they have to be present at construction because each is far harder to retrofit than to install.

Inhabit the surfaces that already exist before building new infrastructure; new infrastructure is justified only when it makes an existing surface more inhabitable. If The stack has extensive infrastructure: Agency, evidence stores, pattern libraries, Arxana browser, candidate invariant queues, mission documents, excursion logs, Bayesian models, context retrieval. However Most of this infrastructure is uninhabited. The Candidate Invariant queue has sixty-five entries with no processing pipeline, the evidence browser was disconnected, excursions are explorations without missions, and the primary coding surface generated no evidence until today.The case that forced it. The candidate-invariant queue held sixty-five entries with no processing pipeline, the evidence browser was disconnected, and the primary coding surface generated no evidence at all. The reflex was to build a queue system. What actually worked was making one surface inhabitable: within minutes of the REPL becoming usable, evidence started flowing — chat turns with session ids, context-retrieval certificates, the conversation browsable — and none of it needed new infrastructure. The evidence loop had been ninety percent built; the missing ten percent was wiring, and the missing ingredient was a person in it. Then The next phase is inhabitation, not construction. New infrastructure is justified only when it directly serves inhabitation of existing surfaces. The three workstreams, mathematics, consulting, and stack inhabitation, must be mutually reinforcing: stack work that does not serve mathematics or consulting is building alone and never delivering, which violates the rule that whatever is carried must be delivered. Because Surface earns inhabitation: uninhabited infrastructure is dead infrastructure. Inhabitation feeds evolution: without it there is nothing for the system to learn from. The same pattern appears at the personal level: much carried, little held in reserve, and effort stuck on gathering rather than delivering. External depositing and internal stack inhabitation require the same policy transition, from scholar to consultant to free-solo.

Make the typed argument files the canonical self-model and generate prose from them, so drift becomes detectable instead of perennial. If The holistic argument was historically maintained as prose, in two separate sketches. However Prose drifted from intent because there was no machine-readable intent it had to track: every re-render was an editorial interpretation.The case that forced it. The holistic argument lived as prose in three documents that each drifted from intent, and the drift was undetectable because there was no machine-readable intent to compare against — every re-render was an editorial interpretation, and a faithful one was indistinguishable from a lossy one. Inverting the direction fixed the detectability rather than the prose: the sixteen leaf files plus the stack file became canonical, the prose became a renderer output, and editing prose in place became a defect rather than a judgement call. Then The sixteen leaf-level and one stack-level typed argument files at the futon5a stories are the canonical self-model. Prose deliverables are generated by renderers and re-rendered when the drift detector reports a source change. Editing prose in place is a defect. Because The typed argument format carries structure, claims, supports, attacks, falsifiability, and conflicts, that prose flattens. If the typed structure is canonical, drift between argument and document becomes detectable instead of perennial.

Render self-model conflict claims as empirical only when session evidence supports them, and as logical otherwise, so the inhabitation gap stares back from the demo. If Conflict bites in the typed self-model assert that a given conflict damages a given part of the stack. However Such claims are formal until session evidence supports them: the cached weight on a conflict is a hypothesis, not a measurement.The case that forced it. One conflict in the cached model claimed to damage three spines. Over a fourteen-day window only one of the three showed any empirical activity — one-third coverage. The other two bites were not wrong so much as unwitnessed: weights that had been carried as measurements since they were first written down. Rendering them as dashed grey rather than solid red put the inhabitation gap on the demo where it could be seen without anyone running a query. Then The empirical join walks session steps to repos to devmap leaves to spine ids, and renders conflict-to-spine bite edges as solid red, scaled by hit count, when the bitten spine has empirical activity in the visible window, and as dashed grey when the bite is still purely logical. Conflicts whose bites remain grey across multiple windows are candidates for weight reduction or removal. Because Invariant I1, evidence supersedes assertion, applies inside the self-model itself. A conflict whose damage cannot be observed is not yet a conflict, only a worry. The visual distinction makes the inhabitation gap measurable without the operator querying anything.

Wire operationalised exploit-loops into the stack’s status dashboard and the War Machine as direct typed evidence streams and first-class active-inference observation channels. If An exploit-loop emits typed structured state on a cadence independent of Joe’s attention. However Treating that state as adjacent reading material leaves existing signals without a surface: a fourteen-day collection lag is invisible as habit failure, and a success-rate drop from ninety-seven percent to eighty-seven percent produces no bulletin-grade alert.The case that forced it. Two loops were already emitting typed state on a cadence nobody had to maintain. Batches 7 and 8 sat uncollected for fourteen days — a binary, visible failure of the operator’s own half of the loop — and mark2’s success rate fell from ninety-seven percent to eighty-seven percent. Neither produced an alert, because the state was being read as adjacent material rather than as an observation channel. The signals existed; the surface for reading them did not. Then Surface habit-completion telemetry, pipeline-health telemetry, and proxy metrics directly in the stack’s status dashboard and the War Machine, and attach the interim-director checkpoint to the stream immediately. Because These loops already accumulate observations without fresh primary research; wiring them turns existing cadence and outcome changes into evidence the stack can condition on.An unbuilt how for which this ruling is the why. operator-turns-become-inference-observations: Process captured operator turns into structures that inference can consume, using the turn-to-pattern association as grounded enrichment. Next steps: since we have full text search via an XTDB sidecar, this can now be piloted

Wire structural tension as a candidate source that generates actions, not merely as a weight that ranks candidates generated elsewhere. If The geometric stack detects high-tension regions that should drive the War Machine’s next action. However The current wiring, ΔT as a sorry-only tie-breaker, cannot generate a move; it only ranks an existing menu, so a system full of unresolved tension still yields nothing when nothing is proposing candidates.The case that forced it. On its first outing the War Machine exhausted its proposer-fed queue into learn-action-class before any outward open-mission surfaced. Tracing the judge showed why: structural pressure was applied only after four fixed proposers had supplied candidates, and only to address-sorry candidates. The system was full of unresolved tension and the machine had nothing to do, because a tie-breaker cannot put an option on a menu it only reorders. Then Add a tension-proposer: a component that emits new candidate actions where tension is highest, so the field can put options on the menu rather than only rank the ones already there. Make this the proximate fix to starvation and the first build slice. Because A ranking signal cannot create an option that is absent from the menu. Tension becomes actionable only when it can propose the candidate whose priority it would otherwise merely increase.An unbuilt how for which this ruling is the why. tension-proposes-candidates: Add a tension proposer that creates candidate actions from tension instead of only reordering candidates supplied elsewhere.

When retiring a proxy constraint, install its replacement discipline in the same act and measure how much external verdict content changes subsequent work, not merely whether a verdict arrived. If A gate was installed as a proxy for another requirement, such as using a cash gate as the cheapest available proxy for external validity, and the underlying scarcity is lifted. However Silently removing the pressure also removes its arbitration discipline. Pleasant, busy, self-contained work can then look like progress from inside even when outward feedback does not affect what gets built.The case that forced it. The cash gate was retired as master gate and demoted to one instrument among several. What it had been supplying was not cash discipline but the requirement that someone outside the work judge it — a discipline that would have gone silently with the gate had the replacement, a witnessed external verdict, not been named in the same act (War Bulletin 14). Then Name the replacement in the same act as retirement. Replace the cash gate’s master role with witnessed external verdict, using compiler, maintainer, peer-review, and money instruments; measure the gate as the graded gain from verdict content into the next decision, not as a weekly yes-or-no event. Because External evaluation cannot be self-supplied, but a boolean gate on an impoverished predicate buys nothing over simply continuing the work. The governing quantity is how much returned content changes the plan itself rather than only the output.

Let unanchored good news change ordering immediately but leave numerical posteriors unchanged until an artifact supports it. If A model permits posterior movement only on inbound clicks carrying evidence such as a reply or ledgered payment. However A rule introduced to stop optimism inflating buyer probability can be applied asymmetrically, allowing inheritance, verbal commitments, or promising lanes to rewrite downstream financial figures without challenge.The case that forced it. The rule was minted to stop optimism inflating the estimated probability of a buyer. The case that exposed the asymmetry arrived from the opposite side: a confirmed good fact, fully invoiced, carried an obligation the numbers did not show. Filtering bad news while admitting good news had left the model open on the side it was not watching. Then Apply the evidence rule symmetrically. Let operator-reported facts affect ordering, including lane preemption and which tension becomes decidable, but do not let them affect runway, probability, or marginal figures until an artifact anchors them. State this asymmetry wherever it appears. Because A model that filters bad news while admitting good news is biased rather than conservative, especially where its financial figures will later be checked by clients, employers, or institutions.An unbuilt how for which this ruling is the why. satisfied-rungs-are-counted-and-surfaced: Count and surface a satisfied rung with the same evidence discipline used for an unsatisfied rung. Result Held under contact and gained a clause. Two operator-supplied constraint releases went through as ordering-effective and number-inert, as specified; the mirror case then forced an amendment the original ruling did not contain (see evidence). The pattern was strengthened by application rather than merely confirmed by it.

Write a countable re-arm condition at the switch whenever an autonomous capability is turned off. If An autonomous capability is deliberately turned off for a sound reason, such as disabling a scheduled loop while hand-dispatched runs confirm its behavior. However The reason outlives its own memory: soundly disabled capabilities become stranded flags, are never re-armed, and return months later as surprises.The case that forced it. Three of the flags stranded in War Bulletin 12 were this shape — switched off with no written condition for switching them back, so the off state persisted by default rather than by decision. The codex cron was on course to become the fourth until its re-arm condition was written into the restore note the same day. Then Put the re-arm condition at the switch itself rather than only in a mission document. Make it countable, such as a specified number of runs whose behavioral checklist passes without operator amendment; the note must state what ends the pause. Because “Until the behavior is confirmed” is a mood rather than a condition, and a stranded flag costs more than the throughput saved by the pause.

Ship every new consumption-to-acquisition loop with an instrument that measures how much returned demand changes what the loop does next. If The gain from outward behaviour back into the underlying plan has recurred as the governing parameter in CA coupling, the strategy layer’s external-verdict gate, and the epistemic economy’s demand-coupling loop. However Every instance was discovered retroactively after construction, and outward activity without coupling is invisible from inside because its activity level looks correct.The case that forced it. The demand-paged mining loop was specified without any measure of what fraction of hunger signals changed the next mining batch, so nothing in it could report whether the demand paging was earning its keep. The same shape recurs wherever a signal is logged but never routed: a judge/formalizer disagreement log that never alters the formalization pipeline is the uninstrumented-loop smell (War Bulletin 15). Then Instrument every new mining, retrieval, outreach, or review loop from day one with the fraction of returned signals that change subsequent acquisition or action. For demand-paged mining, record which hunger signals shaped each batch and report hunger-shaped acquisitions divided by total acquisitions. Because Measured-negative evidence shows that raw coupling is not automatically beneficial. Gain must be observable before it can be shaped; otherwise hermit mode can only be diagnosed retroactively.2 unbuilt hows for which this ruling is the why. commitment-temperature-is-instrumented-as-gain: Instrument commitment temperature as a gain when the actuator loop is born, so its effect on subsequent action can be measured directly. per-tick-mismatch-instruments-outer-loop-gain: Use per-tick mismatch as the gain reading for the mission-conditional-reward outer loop and surface whether returned reward changes its next action.

The audit connects declared constraints to implementation evidence; component conformance does not establish the complete loop.

• $⟨$1.1.1$⟩$

  [def] Missions, contracts and pattern records state different obligations. Their referents and consumers must be identified before their results compose.

• $⟨$1.1.2$⟩$

  [audit] The August 31 vetting found two unsupported closures in a twenty-four-row paper ledger. Resolution included withdrawals, not only implementation repairs.

• $⟨$1.1.3$⟩$

  [meas] The August 30 recorded route agreed with the described wiring on three of nine hops. This is a dated finding, not the live readiness verdict.

• $⟨$1.1.4$⟩$

  [claim] A constraint-to-mechanism map locates questions to investigate; its unmapped mechanisms need individual explanations, not a presumed common cause.

• $⟨$1.1.5$⟩$

  Discharge. The issue board reports current obligations. Neither repaired prose nor a passing component check closes a run-gated claim.

Operator traces support partial attribution, but cannot identify acceptance or attention without further evidence.

• $⟨$1.2.1$⟩$

  [meas] The audited outer loop lowered an action class credit after 108 unanswered proposals. The record did not establish consideration or rejection.

• $⟨$1.2.2$⟩$

  [def] Clock-in joins a turn to its declared mission; retrieval joins it to candidate patterns; mission prose records pattern mentions.

• $⟨$1.2.3$⟩$

  [meas] The generated turn-chain section measures their intersection. Clock context and textual mention answer different questions, so disagreement is not itself an error.

• $⟨$1.2.4$⟩$

  [claim] A response model must distinguish confirmed acceptance, confirmed decline and no-response. Finding traces does not validate that model.

• $⟨$1.2.5$⟩$

  Discharge. The proposed repair requires an intervention and its own comparison. Traceability is the starting evidence, not the outcome.

Transfer to other organisations remains an unrun study, with unresolved corpus descriptions.

• $⟨$1.3.1$⟩$

  [def] The proposed corpus schema distinguishes requirement, benchmark, bounty, dependency, call and environment.

• $⟨$1.3.2$⟩$

  [audit] The earlier prose described both thirty and thirty-two cases, and both uniform and opposed terminal classifications. These descriptions need case-level reconciliation.

• $⟨$1.3.3$⟩$

  [claim] Historical reconstruction does not establish outcome-blind encoding. Contemporaneous self-cases also require controls for author involvement.

• $⟨$1.3.4$⟩$

  [claim] Recurring decisions, consultation records and capacity to act are proposed eligibility criteria, not demonstrated predictors of adoption.

• $⟨$1.3.5$⟩$

  Discharge. The study may expose encoding residue. Neither predictive performance nor a market boundary is established here.

Attribution and richer preferences require explicit observation domains, participation and witnessed consumers.

• $⟨$1.4.1$⟩$

  [claim] Current-state records need observations; reusable rules retain a context and can also be invalidated. Neither is maintenance-free by definition.

• $⟨$1.4.2$⟩$

  [def] The attribution design preserves source strength, unclaimed work and later confirmations instead of turning an inferred association into a witness.

• $⟨$1.4.3$⟩$

  [meas] The August 24 provenance probe recorded observations but no delivered followups. Its historical limit must not be presented as current service status.

• $⟨$1.4.4$⟩$

  [meas] The September 8 disposition fit gives constant risk on three grounded attempts. The September 9 source check identifies a separate mismatch between checkpoint and channel observation domains.

• $⟨$1.4.5$⟩$

  [meas] The Snatch experiment changes realised payoff when pattern order changes. This does not validate the War Machine expected-free-energy implementation.

• $⟨$1.4.6$⟩$

  Discharge. A new run must establish the missing composition and its outcome. Proposed attribution, preference and participation mechanisms remain distinct from that evidence.