Executable Worldview — When Organisational Knowledge Starts Doing Work

A wiki becomes more than memory when cognitive intermediate representation, intent, authority, action, and write-back close into one system. Organisational interpretation can then perform work—without being mistaken for permission.

Scott Farrell · LeverageAI · July 2026

Most agent deployments still treat organisational knowledge as something the model should “look up.” They wire tools, drop a document pile into retrieval, paste a soft system prompt that pretends to be identity, and hope competence appears. When the work comes back thin, they buy a larger context window. When it comes back bold and wrong, they add another policy paragraph. Both moves miss the architecture.

The question that actually matters is sharper: what changes when a wiki stops being memory behind the agent and becomes the world the agent works inside?

The answer is not mysticism. It is a closed stack. Heterogeneous exhaust compiles into a provenance-bearing cognitive intermediate representation—claims, significance, typed edges, time, uncertainty, source pointers. Live intent activates a task-relevant sub-world. An agent runtime reasons against that activated world. An authority gate independent of the wiki decides whether proposed action may execute. Paths, receipts, and outcomes write back so the map and the apparatus that reads it both improve. That whole composition is an executable worldview.

The archive is source. The wiki is cognitive intermediate representation. The model is the processor. Intent activates the relevant portion of the world. An agent executes against that activated world. Experience compiles back into the substrate.

Memory is not a runtime

Retention is not learning. A fact in an append-only log has been retained. A fact related to prior beliefs, exceptions, causes, contradictions, and transferable principles has been integrated—and integration is where organisational learning actually lives.¹ Earlier work argued that the durable kernel for agents should be a curated, queryable wiki-graph rather than a perishable mega-prompt.² Related work showed that heterogeneous life and work exhaust becomes joinable when compiled into one legible intermediate representation, with agents as runtime against that IR.³

Those pieces are modules. This article’s job is their closed composition with intent, authority, action, outcomes, and apparatus learning. Without that composition you get capable runtimes sitting on fragmented context: epistemic gaps (what world is true here?) and authority gaps (may we act?) both unnamed.

Cognitive IR: where understanding is stored

Ordinary retrieval answers which passages resemble a query. Metadata helps route to containers. Neither layer reliably stores the work of understanding: the claim being made, the role it plays, why it matters, which other claim it extends or contradicts, which implementation embodies it, or what becomes visible when two units join.⁴ Relational meaning is not a unary tag on a chunk; it lives between things and must be concluded by travel.⁵

Intelligent ingestion therefore does not ask “what is in this document?” It asks: what happened to our understanding when this document arrived? The new item walks the existing corpus, forms judgements (confirms, contradicts, extends, exemplifies, opens a question), and leaves edges with cognitive provenance—not mere geometric proximity. Significance is not an intrinsic property of the file. It is a relationship between the document, a worldview, a purpose, and a time. Significance without a pointer back to evidence is marketing, not memory.

Semantic decompilation is the recovery method: deterministic structure finds safe cuts; models reverse-engineer significance; the graph preserves relationships so they need not be rediscovered next time.⁶ Stage-relative source discipline applies: prompts and intent packages can be source upstream of generated code; the wiki holds the epistemic identity those stages compile against.⁷

Intent activates; posture changes

At rest the wiki is static the way source code is static. It becomes active when intent runs through it. The user’s sentence is not merely a search string. It is a temporary organising purpose: subordinate questions, alternative framings, convergence hotspots, minority findings that must not be averaged away, known absences. That activation pattern is the intent-compiler move: the query is a disposable probe; the parent intent is the unit of work.⁸

The resulting object is not a bag of hits. It is a temporarily assembled sub-world: held intent, coverage map, fused evidence, divergences, source receipts, and a synthesis aimed at the original purpose. Edges act as a page table so the agent can demand-page the pieces of the worldview the task touches.

Deeper still: once strategy, history, principles, and relationships are attention-resident, every subsequent token is conditioned on them. A lookup is an event. Attention-residence is a condition.⁹ The model does not literally become the organisation. The composite system changes: different salience, different default assumptions, named exceptions, routes to evidence. A fair formulation is that the agent becomes a runtime instance of the wiki—and that this is also why governance matters. Epistemic conditioning is powerful. In regulated environments it is a feature to instrument, not a vibe to ignore.

The wiki grounds action; it cannot authorise it

This boundary is load-bearing. The wiki answers what world the agent believes it inhabits, what is significant, what is unknown, what has happened before. It does not answer whether the agent may act, who holds authority, which data may leave, or what threshold requires escalation.

Safe delegation pairs an epistemic leash above the model with an action leash below it.¹⁰ A well-informed agent without an action leash can perform highly contextualised overreach. An action-gated agent without the wiki can enforce rules while misunderstanding the business. Both are required. The operational identity of the full system is larger than the wiki: epistemic identity, tools, and legitimate agency.

wiki / worldview
    ↓ epistemic leash
model
    ↓ proposed action
authority infrastructure
    ↓ action leash
execution

Outcome closure and write-back

Without outcome closure the system can improve retrieval and still not learn whether real-world recommendations worked. For consequential actions, capture proposal (intent, evidence, assumptions, authority path), execution (what actually happened), observation (results and surprises), and learning (which assumptions held; local evidence vs transferable principle). Then the loop is metabolism: world → wiki → cognition → authorised action → changed world → observed outcome → revised wiki.

Use itself is a learning stream. Hard-won syntheses may be filed as derived—cache, never outranking sources. Paths are telemetry: pages co-visited suggest missing edges; abandoned branches suggest dead ends; never-reached pages suggest naming or link failures.¹¹ When frameworks in the wiki inspire better retrieval machinery, and that machinery is implemented in the walker, you get epistemic reflexivity: the insight compiles into the apparatus that later retrieves the insight. Publishing can even act as an outbound sensor—attention outcomes as probes that return receipts into the canon—without replacing the general architecture.¹²

Flagship: one outreach intent through the stack

The clearest receipt is small on the outside. The literal request looked like: the lead is ignoring me—what should I send? The parent intent was closer to: reopen this commercial relationship intelligently, without needy tone or unsupported accusation.

Against a compiled worldview the temporary field of cognition included employment history and public positioning, an approaching role anniversary, a hypothesis about talk versus commercial delivery, existing correspondence, a product artefact already built, attention-budget doctrine, other people with influence, and response/non-response branches with cadence. The compressed public move was something like: what are you building in year two?—recognition, challenge, next-twelve-months orientation, and an opening to show work already done, without forcing a defensive year-one audit.

That is high context density: abundant private machine attention so scarce human attention is demanded carefully on the outside. Adjacent influencers and follow-up cadence were not scope creep once the parent intent was held. They were part of solving the job. A weak system returns copy. A wiki-conditioned system returns a strategy whose first move happens to be copy.

Failure analysis (illustrative shape)

Round-trip: ideas explain code; code tests ideas

When a development wiki and an IP wiki sit beside each other, forward engineering traces claims into prompts and implementations; reverse engineering recovers design principles from code and behaviour; discrepancies revise the wiki. That is not documentation theatre. It is a verification loop for organisational thought. Model swaps then become less existential: if epistemic identity lives in the wiki rather than in weights, a large portion of behaviour can survive the processor change—remove the wiki and the model returns to being a brilliant stranger.¹

What to build next (and what not to confuse)

If you already have retrieval, the upgrade path is not “more chunks.” It is: claim-level IR with provenance; intent-conditioned activation rather than single-shot search; attention-residence for work that has many micro-forks; an authority plane that is not the wiki; write-back of derived cognition and path telemetry; outcome receipts for consequential actions. Separate bronze evidence from derived cache so the system cannot eat its tail. Preserve contradictions as edges. Keep an alien-signal path so the worldview cannot only confirm itself.

Later work will name the ephemeral task-world object, the asset economics of prepaid orientation, and the organisational jump from institutional memory to institutional cognition. Those are adjacent chapters of the same programme. This article’s claim stands alone: organisational knowledge starts doing work when the worldview is executable—and stays governable when knowledge is never confused with permission.