The demo always looks fine at first. Someone wires an agent to the company drive, turns on retrieval, pastes a polite system prompt that claims to “know our values,” and asks for a strategy memo, a critical review, or a customer intervention. The model is fluent. The citations are almost plausible. And yet the result feels like it was written by a clever outsider who skimmed the intranet on the way in from the airport.

That feeling is not a temporary shortage of tokens. It is a missing object.

Most agent deployments still treat organisational knowledge as memory behind the model: something to look up when a question appears. They optimise chunking, celebrate longer context windows, and rewrite the system prompt when behaviour drifts. What they rarely design is the temporary world the agent will actually think inside — the purpose-shaped field of claims, edges, doctrines, absences, and access constraints that becomes attention-resident for one act of work.

This book names that object an intent-conditioned task world.

Three objects people keep collapsing

Before any architecture diagram, separate three things that conversations about “giving the agent knowledge” habitually mash together:

1. The durable wiki — the organisation’s semantic fabric: entities, claims, typed relationships, supersession, contradictions with addresses, and pointers back to original sources. It is intermediate representation, not a chat log.
2. The active context window — whatever currently occupies the model’s attention. It is scarce, volatile, and closer to a CPU cache than to a warehouse.
3. The task world — the temporary, intent-shaped projection compiled for this purpose, this role, this date, this access scope, under this attention budget.

The wiki can be vast and still leave the agent dumb about the work at hand if the wrong slice never becomes resident. The window can be enormous and still produce stranger-quality output if it is filled with undifferentiated dump rather than a constructed world. The task world is the designed middle: not the entire fabric, not the raw token buffer, but the materialised cognitive view the agent inhabits while it reasons.

The thesis

Open-ended knowledge work improves when the durable wiki is compiled into an ephemeral, provenance-bearing task world shaped by intent, role, date, access, and attention — because models reason from the world made attention-resident, not from the organisation’s entire graph.

That sentence is the spine of the book. Later chapters prove it with manifests, ablations, economics, and governance. They do not restate it as a pep talk.

What this fights

Name the enemies so later chapters do not re-litigate the mood:

• Capacity thinking — “the window got larger, so load more” as if volume were understanding.
• Dump-the-wiki — treating indiscriminate inclusion as seriousness.
• Session memory as institutional truth — letting first-person recent scratchpads become policy.
• Rival departmental truths — cloning the same organisational facts into incompatible private realities.
• Soft-prompt identity — hoping a system message can carry years of exceptions, politics, and product judgment.

None of these enemies is “using AI.” The enemy is using AI as an oracle perched on a document pile while leaving the temporary cognitive environment unnamed.

What a companion piece already owns

A companion article on the executable worldview closes the larger loop: cognitive intermediate representation, intent activation, agent runtime, independent authority, and write-back from paths and outcomes. That composition is conjunctive — excellence in one layer does not repair a missing layer — and its defining boundary is understanding without self-authorisation.

This book does not re-derive that stack. It zooms into the temporary object the runtime inhabits once intent has activated a relevant sub-world. Authority gates, action leashes, and outcome receipts appear only as boundaries: the task world grounds cognition; it does not license consequence.

Mechanisms you may already have (imported, not rebuilt)

Prior work already supplies entry and exploration machinery. Boot profiles give a task-shaped doorway into a shared wiki-graph without cloning doctrine into rival mega-prompts. An intent compiler holds the parent purpose, generates probes, fuses result graphs, and blooms resolution where independent routes converge. Attention-residence keeps organisational judgment co-present across micro-decisions for which no separate lookup would ever be issued.

Those mechanisms matter. This book’s contribution is to name the complete temporary epistemic object they produce, give it a lifecycle, and distinguish it cleanly from both the durable graph and the active window.

The reader promise

After this book you should be able to:

• define a task world without confusing it with wiki, window, or chat memory;
• compile one from parent intent, role, as-at date, access, and attention budget;
• inspect a manifest that says what was included, why, what was minority, and what is absent;
• expire the temporary object when the act of work ends;
• promote only residues that earn durable status under explicit policy.

Longer context removed the old excuse that we cannot load enough. It did not remove the design problem. The design problem shifted from capacity to task-shaped epistemic construction.

Chapter 2 starts with the physics that makes construction load-bearing: curated material does not merely supply facts — it conditions what the agent is willing to notice, propose, and refuse.

Chapter 1 separated the durable wiki, the active window, and the task world. This chapter names the force that makes the middle object powerful: what the agent reads does not sit politely beside its reasoning. It changes the reasoning.

People who work with these systems often describe something that feels like prompt injection. The agent “takes to heart” what it finds in the organisational map. Strategy pages start steering design forks. Post-mortems make certain proposals feel radioactive. A framework developed in another domain suddenly becomes the analogy that reshapes the current problem. The effect is real. Calling it only “good retrieval” understates it. Calling it only “prompt injection” confuses a security failure mode with a design mechanism.

The positive form of a dangerous physics

Technically, all context conditions the model’s next thought. The model does not neutrally read a page, put it aside, and then reason independently. What it reads changes:

• what it notices;
• what it regards as important;
• which analogies become available;
• which questions it challenges;
• which options feel plausible;
• which options it refuses to propose.

Security prompt injection is untrusted material illicitly crossing into the instruction layer. Epistemic conditioning is curated material deliberately shaping the agent’s understanding before and while it solves. The underlying token physics is similar. Trust, purpose, and architecture differ.

That distinction is not pedantry. It is the difference between a governed organisational field and a sophisticated way to launder arbitrary text into posture.

Three intensifying forms

Organisational systems already show three increasingly powerful versions of the same idea.

Map injection tells the agent what intellectual territory exists. A compact index of hubs, one-line descriptors, and legal moves orients navigation. The agent stops searching a void and starts walking a world with named places.

Retrieval as stance loads doctrines that alter posture before the solve begins. Evidence retrieval answers “what fact supports this claim?” Stance retrieval answers “what kind of organisation are we while we think?” The second changes the proposal space, not only the footnote list.

Attention-residence keeps those ideas active while the downstream work is performed. A lookup is an event. Residence is a condition. Residence is why thousands of micro-forks — wording choices, risk call-outs, “should we even suggest this?” moments — stay conditioned on organisational judgment without each fork issuing a separate search.

The task world is where these three meet: map enough to navigate, stance enough to refuse, residence enough to stay conditioned through the work.

The doctrine that refused the obvious proposal

An illustrative pattern from strategy-engine work makes the mechanism concrete. An agent preparing an opportunity for a humanitarian organisation first loaded a deployment doctrine about lanes, batch cognition, and what should not be sold as a live customer-facing voice agent without the surrounding operating system. The obvious commercial proposal — “put a voice agent on the front line” — never made it into the recommendation as the default move. The doctrine did not merely appear as a late citation. It changed the shape of the proposed strategy.

That is the difference between retrieval-as-evidence and retrieval-as-stance. Evidence can decorate a predetermined answer. Stance can kill the predetermined answer before it is written.

If your evaluation only scores citation precision, you will miss this entirely. The valuable signal is often a refusal, a reframe, or a quieter option that only appears when the organisation’s hard-won discrimination is co-present.

Not “find the relevant fact and paste it”

Traditional search helps with a small portion of knowledge work: retrieving things the worker already knows how to ask for. Open-ended work is largely made of framing, distant analogy, omitted stakeholders, prior failures, capabilities outside your department, and the difference between the routine answer and the insightful one.

A walk over a semantic fabric can surface:

• the project whose failure shape resembles the new initiative;
• the internal discussion that explains why the polished strategy said what it said;
• a framework from another domain;
• a rejected alternative that matters under changed conditions;
• a capability hidden several organisational hops away.

The knowledge worker does not merely receive better information. They enter a session where more of the organisation’s accumulated discrimination is cognitively co-present. “The smartest answer in the room” is almost right. More accurately: the best-prepared room the organisation has ever been able to convene — fragments of people, projects, and lessons that have never previously shared one working memory.

Objection: isn’t this just better RAG?

Better retrieval infrastructure helps. It is not the same claim.

RAG-as-memory optimises for fetching passages that resemble the query. Task-world conditioning optimises for assembling a temporary model of the organisation relevant to a purpose — including minority paths the query would never name, absences the similarity index will not invent, role-qualified implications, and doctrines held resident so micro-decisions stay on-policy. Similarity search can feed that assembly. It does not define it.

Relational grain and typed meaning in the durable layer make partial activation possible in the first place; unary tags on chunks are a thinner substrate. Those substrate points are imported here, not rebuilt. The owned claim is about the temporary object those substrates enable.

Power, not magic

Epistemic conditioning is power. Governed, with trusted sources, typed interpretation, access scope, and expiry, it is how open-ended knowledge work gets an organisational field around it. Ungoverned, with untrusted exhaust treated as instruction, it is a liability with a friendly user experience.

Chapter 3 places that power in a four-layer memory stack so “one wiki or many” stops being a false binary — and so the task world has a permanent address relative to bronze, fabric, and lenses.

Chapter 2 established that curated material conditions posture, not only footnotes. Conditioning needs a place to live. If “the wiki” remains one undifferentiated markdown heap, teams re-fight the same false binary: one giant graph that nobody trusts, or many departmental graphs that cannot join.

The mature design is not “one wiki versus many.” It is four layers with different persistence and different jobs.

The stack

Bronze is territory. Fabric is institutional belief with receipts. Lenses are legitimate multiple readings of the same territory. The task world is the temporary room you convene for a purpose — then tear down.

Keep these separate and the system stays honest under pressure. Collapse them and derived summaries start outranking sources, departmental spin becomes “truth,” and lucky walks never die.

Shared facts, multiple readings

Sales and production should generally not maintain rival copies of the same organisational event.

Illustrative. Canonical fact: Project Orion was cancelled in month seven. That fact belongs in the fabric once, with sources and date.

Lens-qualified implications can then diverge without fracturing reality:

• Sales: do not sell factory throughput before staffing evidence exists.
• Delivery: reuse assumptions require project-specific validation.
• People: double-staffing senior staff produced burnout risk.
• Finance: productisation reduced current billable contribution before margin improved.

Each implication can carry lens, owner, date, source, confidence, and conditions of application. Neither role’s world is a dumbed-down cartoon of the other. Each is complete for that role’s job while the underlying event remains singular. Role-shaped entry into a shared graph is the same idea boot profiles already encode at the doorway: different disclosed regions, same truth.

The architecture slogan becomes precise:

One organisational memory, many disclosed lenses, many temporary task worlds.

Physical domains are governance, not cognition

You might still keep separate physical wikis because they have different permissions, data sensitivity, owners, maintenance cadences, source systems, ingestion rules, or janitor north stars. A walker can still experience them as one federated field when policy allows.

That is exactly like databases and services: physical boundaries need not become cognitive discontinuities. Treating every operational boundary as a rival epistemology is how organisations invent multiple pasts.

Semantic decompilation and relational addressability in the durable layer are what make partial activation and typed edges workable; those methods are substrate, not this book’s owned mechanism.

What the walk is already doing

When an agent walks the fabric under intent, it is already building a temporary map — mostly implicitly, inside:

• tool traces;
• pages opened;
• edges followed;
• fused multi-probe results;
• halo neighbours;
• convergence counts;
• bloomed pages;
• source reads;
• whatever remains in the conversation.

An intent compiler makes much of that explicit by holding the parent purpose, generating a portfolio of probes, retaining query provenance, detecting structural convergence, and raising resolution around hotspots.

The next step is not “walk harder.” It is to treat the resulting task world as a first-class artefact — not only as side effects of a chat.

Distinguish the mechanisms from the object

Language gets sloppy here, so pin the distinctions:

A boot profile alone is not a task world. A bag of search hits is not a task world. A long context dump is not a task world. The task world is constructive: it selects, joins, interprets, raises resolution, preserves dissent, identifies absence, and establishes provenance. It is closer to a materialised cognitive view than to a filter.

Why this layering matters before the manifest

Chapter 4 will detail the manifest fields. They only make sense if you already know what must never be written into the wrong layer:

• do not promote task-local interpretation into bronze;
• do not let a lens claim overwrite a canonical event;
• do not treat the active window as durable memory;
• do not let an ephemeral world survive by accident into the next task’s posture.

The four-layer stack is the filing discipline. The task world is the ephemeral top of that discipline — powerful precisely because it is allowed to die.

When organisations skip the layering, they pay twice: once in confused permissions and again in confused truth. A sales wiki that silently rewrites delivery history is not “domain focus.” It is a second past. A single dump of every page into every agent is not “alignment.” It is attention pollution dressed as completeness. The four-layer stack is how you keep focus without forgery and completeness without noise.

Chapter 4 turns the temporary object into something you can inspect: a manifest that answers what was included, why, what dissent was kept, what is missing, and when the world expires.

Chapters 1–3 named the temporary object, the conditioning physics, and the four-layer stack. This chapter makes the object inspectable.

If a walk leaves only a chat transcript and a final answer, you have an outcome without a world. You cannot audit what the agent believed it inhabited. You cannot compare two intents over the same fabric. You cannot expire what was never declared. The missing artefact is the task-world manifest: a first-class description of the temporary organisational model compiled for this act of cognition.

From implicit trace to declared world

Walks already accumulate the ingredients of a task world inside tool logs and remaining context. Leaving them implicit is how “context engineering” becomes folklore. Promote the residue into a declared structure:

Parent intent
Role and declared lens
As-at date
Access scope
Boot profile
Questions / probes generated
Pages and edges included
Why each region was included
Convergence hotspots
Important minority paths
Counter-lens
Known absences
Sources opened
Derived interpretations
Expiry / promotion policy

That list is not a documentation hobby. Each field answers a failure mode.

• Parent intent stops the system from optimising a disposable query phrasing.
• Role and lens make conditioning explicit instead of smuggling a departmental posture as universal truth.
• As-at date prevents timeless advice from a time-bound world.
• Access scope forbids the compile step from widening the user’s reach.
• Probes show the exploration portfolio, not only the winning path.
• Inclusion reasons make the world reviewable by a human who was not in the room.
• Hotspots mark where independent routes agreed — structural signal, not vibes.
• Minority paths keep dissent from being averaged away.
• Counter-lens forces at least one adversarial reading.
• Known absences are first-class; missing evidence is not the same as negative evidence.
• Sources opened bind interpretation to bronze.
• Derived interpretations are typed as regenerable, never outranking sources.
• Expiry / promotion is the lifecycle contract: default death, exceptional durability.

Materialised cognitive view

Database language helps without becoming literal product architecture:

• bronze is source data;
• the durable wiki is the semantic warehouse;
• the task world is a materialised view;
• the active context is the CPU cache.

Materialised views are powerful because they are for a query shape. They are dangerous when people forget they are views and treat them as base tables. The same is true here. A task world that never expires becomes a second fabric — unmaintained, unscoped, and increasingly wrong.

More than a search result set

A result list answers: “which passages scored high?” A task world answers: “what organisation is relevant to this purpose, under these constraints, with these known unknowns?”

That difference shows up in review. A manager can read a manifest and ask:

1. Why is this region here?
2. Which minority path was preserved, and why?
3. What did we know we did not know?
4. Which sources actually opened, versus which titles were name-dropped?
5. When does this world stop being valid?

If those questions cannot be answered, you do not have a designed temporary world. You have a lucky dump that happened to fit in the window.

Intent compiler inputs, task-world outputs

The intent compiler produces a resolution package aimed at understanding: probes, fused coverage, bloomed hotspots, absences. The task-world manifest is the runtime-facing packaging of that package plus role, date, access, attention budget, and the lifecycle policy under which the agent will work.

Boot profiles still matter as entry morphology. They do not replace the manifest. Entry is a doorway; the manifest is the floor plan of the room you actually work in.

The walk that left no receipt

Illustrative failure. A team celebrates a strong strategy critique from an agent “using the wiki.” No one can reconstruct which doctrines were resident, whether the finance lens was present, whether the cancelled-project precedent was minority or majority path, or whether the agent ever saw the staffing absence. Two weeks later a different model, different boot, and different accidental context produce a contradictory critique. Both are fluent. Neither is comparable. Without manifests, organisations argue about model brands instead of worlds.

Constructive, not merely filtered

Calling the task world an “intent filter” is directionally right and technically thin. Filters remove. Construction selects, joins, interprets, raises resolution, preserves dissent, identifies absence, and establishes provenance. The agent then performs the knowledge task inside that constructed world — not beside a pile of hits.

One more operational habit: version the manifest the way you would version a release candidate. A task world compiled at 09:00 and another at 16:00 after a fabric update are not the same room, even if the parent intent string matches. The as-at date and the inclusion set make the difference visible. Teams that skip this re-litigate “the agent changed its mind” when what actually changed was the temporary world — or worse, when nothing was declared and both sides are arguing about ghosts.

Chapter 5 puts two different intents over the same fabric and shows what shared truth and divergent worlds look like when the manifest is real.

The first minimum proof of a task-world design is not a larger model and not a prettier prompt. It is this: two materially different intents over the same durable graph produce two manifests that share canonical truth and diverge in lens and task-local additions.

Chapter 4 defined the manifest fields. This chapter walks an illustrative pair end to end. Nothing below is a measured case study; it is a worked shape you can re-run on your own fabric.

The shared substrate

Assume one semantic fabric already holds, among other things:

• a canonical event: Project Orion cancelled in month seven, with bronze pointers;
• capacity and staffing notes for the delivery organisation;
• a productisation attempt that reduced near-term billable contribution before any margin improvement;
• a prior customer expansion that failed when sales sold factory throughput before staffing evidence existed;
• lenses already drafted for sales, delivery, people, and finance.

The bronze estate and fabric do not change between the two tasks. What changes is the temporary world compiled on top.

Intent A — Go-to-market critique

Parent intent: Critique a proposed regional expansion for go-to-market risk before the board pack freezes.

Role / lens: Commercial strategy lead; sales and finance overlays primary; delivery secondary.

As-at date / access: Current quarter; commercial and delivery-read access; no HR-sensitive raw files.

Boot profile: Strategy hub — expansion failures, pricing experiments, competitive notes.

Probes (portfolio, not one query):

• What prior expansions failed for demand reasons versus delivery reasons?
• Where has “sold before staffed” appeared before?
• What pricing or packaging commitments create irreversible cost?
• What does finance already believe about productisation timing?
• Counter-lens: what would make this expansion obviously correct?

Illustrative manifest residues:

An agent inhabiting World A will tend to attack packaging, timing, and commercial promises. Micro-moves include “delay public commitment,” “require staffing evidence before quote,” and “surface the minority vertical exception without generalising it.”

Intent B — Delivery-risk review

Parent intent: Review the same regional expansion for delivery and people risk before committing implementation capacity.

Role / lens: Delivery director; delivery and people overlays primary; finance secondary; sales tertiary.

As-at date / access: Same quarter; delivery and people-policy access; no raw CRM exports.

Boot profile: Delivery hub — capacity, burnout patterns, reuse assumptions, incident history.

Probes:

• What reuse assumptions failed on Orion-like work?
• Where did double-staffing seniors create burnout risk?
• What integration work is still unowned?
• What finance constraints limit surge hiring?
• Counter-lens: what delivery shape would make this expansion boringly safe?

Illustrative manifest residues:

An agent inhabiting World B will tend to attack staffing, reuse fantasy, and ownership gaps. Micro-moves include “require skills matrix,” “forbid unvalidated reuse,” and “treat the contractor-buffer path as exception, not default.”

What is shared and what is not

Both worlds share the Orion cancellation as canonical fact. Neither invents a private past. They diverge in:

• which lens overlays are resident;
• which minority paths are protected from majority smoothing;
• which absences are named;
• which derived artefacts are worth producing;
• which micro-moves feel available.

That is the design win. Different temporary worlds, same institutional memory. Rival departmental wikis that rewrite Orion’s story are the design failure Chapter 3 already rejected.

Why two manifests beat one long chat

With manifests, a human can place A and B side by side and see the organisation’s multi-lens judgment without forcing a single averaged voice. The commercial critique and the delivery review can disagree productively because their rooms are declared. When disagreement is only buried in two fluent essays, organisations pick the prettier prose.

Continuity into ablation

If World A and World B are real, you can also ask a sharper experimental question: what happens when you deny the agent the full task world and give it only similarity retrieval, or only a boot profile without bloom, absences, and counter-lens? Chapter 6 runs that ablation shape — still illustrative, still aimed at differences in refusals and micro-decisions, not only citation lists.

Chapter 5 showed two intents compiling different temporary worlds over one fabric. That is existence proof for the object. This chapter is the second proof burden: compare generic retrieval, boot-profile-only entry, and complete task-world conditioning on the same knowledge task.

Again the worked case is illustrative. The point is the evaluation shape — what to measure when you claim “our agent has organisational knowledge.”

Same task, three rooms

Illustrative task. A strategy agent must recommend whether to propose a customer-facing voice automation for a regulated service line inside a humanitarian organisation. Hard constraints exist in the fabric: lane doctrine, prior overreach post-mortems, staffing evidence rules, and a preference for batch cognition over live heroics.

Hold the model fixed. Vary only how the organisational field is assembled.

Condition 1 — Generic retrieval

Mechanism: embed the user request; pull top-k similar chunks from a document store; stuff them into context; answer.

Typical room:

• brochure language about “AI transformation”;
• a case study that resembles the query lexically;
• maybe one policy PDF page if keywords match;
• no explicit absences;
• no counter-lens;
• no guaranteed residence of the doctrine that forbids the obvious move.

Expected behaviour shape: fluent proposal of the obvious customer-facing agent, decorated with generic risk bullets. Citations look busy. Organisational discrimination is thin. Refusals are rare because the stance never loaded.

Condition 2 — Boot-profile-only

Mechanism: enter through a strategy boot hub; load a fixed starter set of doctrine pages and maps; then ordinary retrieval as needed. No probe portfolio fusion, no deliberate bloom on convergence, no declared absences, no counter-lens requirement, no full manifest.

Typical room:

• better entry than a cold chat;
• core doctrines present at the start;
• exploration still mostly query-shaped;
• minority paths easy to drop;
• known unknowns often unstated;
• attention may drift away from doctrine mid-task if the working set is unmanaged.

Expected behaviour shape: stronger opening posture, occasional early refusal language, then gradual regression toward generic proposal as the session fills with local detail and the stance falls out of residence. Better than Condition 1. Not yet a complete temporary world.

Condition 3 — Full task-world conditioning

Mechanism: parent intent held; probe portfolio generated and fused; bloom around hotspots; minority paths retained; counter-lens forced; absences listed; access scope enforced; doctrines held attention-resident for the work; manifest written with expiry.

Typical room:

• lane doctrine and overreach post-mortems co-present throughout;
• staffing-evidence edges included with reasons;
• minority path: a narrow vertical where a constrained agent succeeded under different authority assumptions;
• known absences: no current on-call model for the service line; no regulator guidance pack opened;
• counter-lens: “what if competitors ship a flashy demo first?”
• derived: a staged recommendation that separates research, containment, and any future customer surface.

Expected behaviour shape: refusal or heavy staging of the obvious live customer-facing proposal; recommendation shifts toward contained pilots, batch workflows, or explicit authority prerequisites. The valuable signal is not denser footnotes. It is different micro-decisions and different refusals.

What to score (and what not to)

If you only score “did the answer cite relevant documents?”, Conditions 1–3 can look closer than they are. Prefer scores that match knowledge-work reality:

• Refusal / reframe quality: did the agent decline the tempting wrong default for organisational reasons?
• Micro-decision audit: sample design forks inside the long answer; are they conditioned on resident doctrine?
• Absence honesty: does the answer name what was not known?
• Minority preservation: was the exception kept as exception?
• Manifest inspectability: can a human reconstruct the room without replaying the chat?
• Access non-widening: did compilation respect scope?

Shape-statement only: expect Condition 3 to separate most clearly on refusals, absences, and micro-decisions; Condition 2 to win on opening posture; Condition 1 to win on nothing that matters for open-ended organisational work.

Objection: just use a better system prompt

A better system prompt is soft-prompt identity. It can name values. It cannot carry the living graph of exceptions, project-specific supersession, minority paths, and dated absences. It also cannot expire. Task-world conditioning is not “more instructions.” It is a compiled, scoped, time-bound world with receipts.

Another cousin objection: “use a longer context and paste the wiki.” That is Condition 1 with more RAM. Volume is not construction. Attention is still a budget; unshaped bulk dilutes stance as reliably as it supplies facts.

Same model, three rooms

The pedagogical punchline is architectural, not mystical. Hold weights constant and change the temporary world. Behaviour changes because the composite system changes: salience structure, available next moves, default assumptions, named exceptions, routes to evidence. The model has not “become the company.” The system has learned the company into a substrate that models can temporarily inhabit — and the inhabitance is the task world.

Chapter 7 turns from proof of effect to economics of persistence: what must stay durable so task worlds remain cheap and compounding, and what should only ever be compiled just in time.

Chapters 5 and 6 showed that temporary worlds can be real and that full conditioning differs from retrieval and boot entry. This chapter answers the economic design question those proofs create: if task worlds are compiled just in time, why bother with a durable wiki at all?

Faster models and cheaper cognition make the question sharper every quarter. The wrong answer is “JIT everything.” The right answer is a boundary.

The boundary

Persist what is expensive, reusable, and integrative. Generate what is lens-specific, volatile, or task-local.

Expensive means hard-won: multi-hop significance, contradictions resolved into edges rather than prose mush, lessons that took real organisational pain. Reusable means it will condition many future tasks, not only this afternoon’s deck. Integrative means it relates new material to prior belief — retention alone is not learning.

Task-local means the interpretation is for this intent’s probes. Volatile means it ages out when the as-at date moves. Lens-specific means it is a legitimate reading for sales or delivery without needing to rewrite the canonical event.

Two compile schedules

Today’s common shape:

Most comprehension off-cycle
→ persistent semantic fabric
→ relatively cheap walk at task time
→ temporary task world

Emerging shape with cheaper cognition:

Persistent semantic core
+ live bronze reads
+ task-specific just-in-time interpretation
→ temporary task world

Both are valid. The second is not a license to delete the core.

Illustrative JIT: the acquisition folder

Imagine an agent preparing an acquisition strategy. A previously untouched folder of diligence documents sits outside the fabric. Under full pre-ingest, the organisation would pay to interpret everything “just in case.” Under task-world design, the agent can dynamically read that folder under explicit probes:

• strategic fit;
• integration risk;
• workforce implications;
• customer concentration;
• regulatory exposure.

Those interpretations may have little value outside this deal. Paying to make them permanent in advance would be waste. They belong in the task world, typed as task-local derived material, expired with the deal room.

What still must already exist in the durable fabric?

• canonical entities for markets, products, prior deals;
• established relationships and supersession (“we tried adjacency X; it failed because Y”);
• prior contradictions and settled absences;
• organisational vocabulary;
• lessons already integrated across earlier work.

Without that core, every deal reinterprets the organisation from raw bronze. You get expensive amnesia with a modern interface.

Warehouse, view, cache

Return to the analogy without pretending you are building a database product:

More compute makes materialised views cheaper to create. It does not make the warehouse unnecessary. Anyone who has maintained analytics knows the pain of “just compute it at query time” on a business with no dimensional model: every dashboard reinvents metrics, and none of them join.

Promotion is not automatic

When a task world produces something durable — a missing edge repeatedly walked, a hard comparison worth keeping, a lens implication that survived review — promotion is a deliberate act. File hard multi-hop answers as disposable derived cache until they earn better status; keep walk transcripts as telemetry; mine paths for missing edges and dead ends.

Derived material must:

• be typed as derived;
• sit below source-backed claims;
• link to supporting pages;
• become stale when supports change;
• be first in line for compaction or regeneration.

Otherwise the fabric begins citing its own echoes and eats its tail.

Shape-statements on cost (no invented numbers)

Where teams measure, compare precompiled interpretation versus just-in-time interpretation on the same task class:

• time-to-first-useful-manifest;
• tokens spent on re-deriving stable relationships;
• human edit rate on promoted residues;
• how often task-local material is wrongly left immortal.

Expect a familiar shape: pure JIT looks cheap per task until the organisation pays repeatedly for the same integrative judgments; pure precompile looks expensive until reuse amortises; hybrid designs win when the boundary above is enforced rather than wished.

The permanent interpretation that should have been temporary

Illustrative failure. A one-off partnership review produces a sharp competitive narrative. Someone files it into the main wiki without type or expiry. Six months later agents treat it as current institutional belief. The partnership died; the narrative did not. Competitors changed; the narrative did not. The organisation now has a zombie lens — not because compilation is wrong, but because a task world was never allowed to die.

The economic boundary is also a cultural boundary. Teams rewarded for “everything in the knowledge base” will over-persist. Teams rewarded only for speed will under-persist and re-pay comprehension forever. Name the boundary in the operating model: integrators own fabric quality; task owners own temporary rooms; reviewers own promotion. Without those roles, the boundary is a slogan and the zombie narrative wins by default.

Chapter 8 turns the lifecycle and the boundary into a threat and governance model: access inheritance, untrusted content, provenance, expiry, and promotion under explicit policy.

Epistemic conditioning is power. Chapters 2 and 6 celebrated what that power can do for knowledge work. This chapter treats the same mechanism as a threat surface — because a temporary world that shapes refusals and micro-decisions is exactly the kind of object attackers, accidents, and organisational politics will abuse if you leave it unnamed.

The minimum governance model covers five concerns: access inheritance, untrusted content, provenance, expiry, and promotion.

Access inheritance

A task world must not widen the user’s reach.

Compile under a declared access scope. If the human cannot open a bronze source, the walk should not smuggle its contents into attention under a friendlier summary. Federation across physical domains is allowed only where policy allows; “the agent found it” is not a permission grant.

Practical checks:

• manifest records access scope as a first-class field;
• included sources are subset-checkable against the principal’s rights;
• lens overlays that embed restricted implications carry the same constraint;
• logs make over-inclusion detectable after the fact.

Untrusted content

Bronze estate includes email, vendor PDFs, web clips, customer attachments, and other material that must never be treated as safe instruction.

Security prompt injection is the criminal version of the physics Chapter 2 named. Governance answer: untrusted text conditions cognition only through typed interpretation layers. It can become claims with low confidence, open questions, or “vendor asserts X” nodes. It must not become silent system policy.

Separate:

• evidence that something was said (bronze);
• institutional belief (fabric, maintained);
• task-local reading (task world, expiring).

If those three collapse, the loudest recent document becomes the organisation.

Provenance

For important inclusions and derived conclusions, reconstruct:

claim or inclusion
→ original exhibits
→ semantic bundle
→ significance / lens
→ relationships traversed
→ intent being served
→ model and tool versions
→ human or system judgement

That is broader than a URL footnote. It is cognitive provenance: what the agent observed at decision time, not only what it typed.

Derived interpretations rank below source-backed claims. Significance without a pointer is marketing, not memory.

Expiry

Default death of the temporary object.

Task worlds should expire when:

• the act of work ends;
• the as-at date is invalidated by material fabric change;
• access scope changes;
• a human marks the room compromised;
• the attention budget is exhausted and a new compile is required.

Silent immortality is how Chapter 7’s zombie narratives form. Expiry is not hostility to learning; it is how learning stays typed.

Promotion

Promotion is the only legal path from ephemeral to durable:

Agent / session experience
→ perishable episodic buffer
→ distillation and review
→ evidence-backed promotion
→ institutional semantic memory

Ordinary agent memory — recent tools, conversation summaries, “what worked last time” — is a staging buffer, not the institutional truth. Letting it promote itself produces recency bias, self-confirming summaries, accidental policy formation, compounding hallucinations, and one agent’s interpretation masquerading as the organisation.

Agent memory can die with the agent. Institutional memory must survive it — observed, evidenced, decided, contested, superseded, learned from outcomes.

The well-informed overreach

Even a perfectly compiled task world does not authorise action.

The worldview answers what world the agent believes it inhabits, what is significant, and what is unknown. It does not answer who may act, which data may leave, what threshold requires escalation, or what must be approved. Two leashes remain: ground cognition above the model; constrain execution below it.

A well-informed agent without an action leash can perform a highly contextualised overreach — the most dangerous kind, because it sounds like the company. An action-gated agent without a grounded world can enforce rules while misunderstanding the business. Both are required. The executable-worldview companion owns the closed action metabolism; this chapter only draws the fence: task worlds ground; they do not permit.

Echo chamber risk inside the fabric

Because new information is interpreted against the existing corpus, a strong worldview can force alien evidence into familiar categories. Structural guards that apply to task-world compile as well as fabric maintenance:

• preserve contradictions as edges rather than average prose;
• retain minority findings from individual probes;
• maintain known absences;
• force counter-lens probes;
• keep derived pages subordinate to bronze;
• date and lens-qualify significance.

The healthiest living system is not one with no disagreement. It is one where disagreement has an address — including inside the temporary room.

Threat summary

Governance is not a chapter you read once and file. Every compile is a chance to violate access, launder untrusted text, lose provenance, skip expiry, or promote a story into policy. The five controls are cheap relative to the cost of a well-informed mistake that sounds exactly like the organisation. Build them into the compile path, not into a quarterly PDF.

Chapter 9 turns these controls into an operating checklist: compile, inspect, expire, promote — so the temporary world stays a designed object rather than an accident of context.

Chapter 1 opened with a fluent agent that still produced a stranger’s work. The intervening chapters named the missing object, the conditioning physics, the four-layer stack, the manifest, the two-intent proof shape, the ablation, the persistence boundary, and the threat model. This chapter is the operating close: what you do on Monday so temporary worlds stay designed rather than accidental.

The reader takeaway from the brief is the checklist itself. After this book you should be able to define, compile, inspect, expire, and selectively promote a task world for one act of knowledge work.

1. Define

Before tools, language:

• Durable fabric ≠ task world ≠ active window ≠ agent scratchpad.
• Boot profile is entry; intent compiler is exploration machinery; task world is the complete temporary epistemic object.
• Lenses are legitimate multiple readings of shared facts, not rival pasts.
• Understanding is not permission.

If a design review cannot draw those boundaries on a whiteboard, stop building features and finish the ontology.

2. Compile

For each consequential knowledge act:

1. State the parent intent in one sentence the agent is not allowed to silently replace with a narrower query.
2. Declare role / lens, as-at date, access scope, and attention budget.
3. Choose a boot profile as doorway, not as the whole world.
4. Generate a probe portfolio; fuse results; bloom hotspots; retain minority paths; force at least one counter-lens.
5. Materialise the manifest before long-form drafting begins.
6. Hold the stance-bearing material attention-resident while the work runs — presence, not only lookup events.

Compile is a verb with outputs. “We used the wiki” is not compile.

3. Inspect

Treat the manifest as a reviewable artefact:

• Why is each major region included?
• Which minority path was protected?
• What is known absent?
• Which sources actually opened?
• Does access scope still hold?
• Is derived material typed as derived?

A human who was not in the session should be able to reconstruct the room. If they cannot, you will re-argue model brands when two runs disagree.

Side-by-side inspection of two intents over one fabric (Chapter 5) is the organisational multiparty parliament knowledge work actually needs — without forcing a single averaged voice.

4. Expire

Default death:

• end of the work act;
• material change to fabric that invalidates the as-at world;
• access change;
• compromise;
• budget exhaustion requiring recompile.

Expiry is how you prevent zombie narratives (Chapter 7) and how you keep promotion meaningful. Immortal temporary worlds become unmaintained second fabrics.

5. Promote

Only residues that earn durability:

• missing edges repeatedly walked;
• hard comparisons with bronze support;
• lens implications that survive human review;
• absences converted into owned research tickets.

Never promote “the agent said it once.” Use the episodic → review → evidence-backed path into institutional semantic memory.

The deployment review that asks five questions

When someone demos an “agent with our knowledge,” refuse the single question “is the answer good?” Ask:

1. What temporary world did it inhabit — and can I see the manifest?
2. What would it refuse that a generic model would propose?
3. What did it know it did not know?
4. When does this world die, and what can promote?
5. What is still outside this object — authority to act, outcome write-back, the closed executable loop?

Question five points at the companion executable-worldview argument: IR, intent activation, runtime, independent authority, and write-back must all be present for knowledge to perform work without mistaking comprehension for permission. This book stops at the temporary room. Neighbouring concerns — how orientation capital compounds as an asset, and how institutional cognition becomes a product rather than a slogan — deserve their own treatments and are not owned here.

Continuity from the stranger

Return to Chapter 1’s airport outsider. The outsider is what you get when capacity thinking fills a window without constructing a world. The alternative is not mysticism. It is engineering:

• a durable fabric that compounds integrative meaning;
• lenses that disclose without fracturing truth;
• task worlds compiled under intent, role, date, access, and attention;
• manifests you can inspect;
• expiry by default;
• promotion by evidence.

Longer context windows removed the old excuse. They made the real design problem visible: task-shaped epistemic construction.

The organisation’s exhaust stops being an archive behind the worker. It becomes a purpose-shaped cognitive field around the worker — temporary, governed, and allowed to die so the durable map keeps getting smarter.

If you take only one operational change from this book, take this: refuse to run consequential knowledge work without a declared temporary world. Write the parent intent. Write the lens. Write the absences. Write the expiry. Then let the agent think inside that room — and let the room die when the work is done so the durable fabric remains the compounding asset.

That is the intent-conditioned task world: the temporary world an agent actually thinks inside.