My AI router is boring on purpose. For weeks at a stretch it says nothing. Then, out of the blue: a Wi-Fi mesh point is down, it has been down for two hours, it already told me once, and the fix needs a physical operator to re-power the unit. I walk into the kitchen. The plug is on the floor. Knocked out. Exactly the kind of stupid, local, real-world failure that makes a smart system look psychic when it is really just well-informed and quiet.

The freaky part is not the alert. The freaky part is the month of silence that made the alert feel expensive. If that router had been chirping about CPU temperatures and interface flaps every afternoon, I would have trained myself to ignore it. The true message would have arrived as noise. Instead, when it spoke, I moved.

That sentence is the whole book in miniature. Agents do not only fail by being wrong. They fail by being present when they should have been absent — and by being inventive when the honest answer was that nobody had decided yet.

Two complaints, one missing layer

Listen to people who have shipped a real agent against a real workload. You hear two complaints that sound unrelated.

First: it interrupts constantly. Every email is urgent. Every ticket needs a human. Every alert is “just flagging this.” The operator learns to mute the channel. Trust dies of a thousand pings.

Second: it invents policy. Asked about discounts, escalation paths, retention rules, or “how we usually handle this,” it synthesises a plausible corporate answer from general training data. It sounds professional. It is fiction. Someone downstream acts on it.

The industry diagnosis is almost always the same: we need a smarter model, a longer prompt, a sterner “do not hallucinate” clause, maybe more tools. Those moves sorta help. The “sorta” is the tell. Capability can paper over a missing world for a little while. It cannot manufacture the thing that was never compiled.

What a past actually is

We are not talking about chat history. We are not talking about an append-only memory log that stumbles into relevance one incident at a time. We mean a compiled past: a durable, navigable model of the person or organisation the agent serves — what is normal, who matters, what is in flight, what already happened, and what the organisation has explicitly not decided.

That past does two jobs that look different and are the same architecture:

1. Baseline. Against a dense enough world, most incoming events are unsurprising. Silence becomes a legitimate model output — the expensive, high-judgment product, not an idle default.
2. Documented absence. Against pages that say “no formal policy exists,” the agent gains permission not to invent. Unknown becomes knowledge. Known absence becomes a first-class fact.

Neither of those can be prompted in. A system prompt can scold. It cannot give the agent childhood. It cannot carry six months of institutional osmosis in fourteen hundred words while the agent goes live in seventeen seconds.

What this book owns (and what it refuses)

This field guide is about knowledge-driven suppression: baseline silence and documented absence. That is the layer that turns an amnesiac stranger into something that can earn the right to interrupt — and the right to say it does not know.

How we will prove it

Part I builds the doctrine: silence as high-judgment output, the two kinds of absence, institutional prehistory as architecture rather than sci-fi flavour text.

Part II runs the receipts. Same model, same Gmail stack: hopeless triage until a multi-year email wiki made the agent mostly silent and occasionally indispensable. Same architecture on a home network router. Then the corporate twin: a discount request that forces invention when the world is empty, and honest escalation when the world contains “no formal policy exists.”

Part III is how to compile backward from the exhaust you already have, and what success looks like when the quality ceiling really is silence — when the best agent is the one that notices the world has stopped matching the wiki.

Stop treating your agent like it is stupid. Start treating it like it is amnesiac. Then give it a past.

Most teams optimise agents for helpfulness as volume. More drafts. More flags. More “just so you know.” The dashboard goes green. The human goes numb.

The production pattern runs the other way. Once an agent can see a compiled world — who matters, what is normal, what is already in motion, what the operator already knows — its job flips from narrating the stream to suppressing almost all of it. That suppression is not a bug in the notification system. It is cognition.

Silence is a model output

Treat silence as a first-class product of the model plus its world. Given an event, the agent must decide whether the event clears a bar. Clearing that bar is a judgment. Failing to clear it should produce nothing — not a soft-pedalled FYI, not a low-priority badge that still trains the human to flinch.

That is why “hardly ever tells me about any of my email” can be the success state for a triage agent. Industry ears hear that sentence as failure. Operators who have lived with a good agent hear it as precision. When the rare message arrives, it is expensive in the good sense: it costs attention because it is worth attention.

Silence is a high-judgment output. Noise is often the cheap default of an empty world.

What the baseline must contain

A baseline is not “more documents in the vector store.” It is a compact model of normality and significance for the surface the agent owns. For a personal inbox agent, that looks roughly like:

For a corporate agent the nouns change — accounts, policies, margin pressure, open deals, authority maps — but the shape does not. The agent needs a world in which most events are unsurprising.

Triage is a diff

Ask the question carefully: Is this email worth interrupting Scott for? That is not a classification of the email in isolation. It is a comparison between the email and a model of Scott. Without that model, the agent is forced into one of two failures: interrupt everything that looks linguistically urgent, or invent a theory of Scott from general priors.

So write the claim the way operators eventually discover it in production:

That is why better models only sorta help on empty-world agents. Intelligence can guess harder. It cannot remember a childhood it never had. We will run the controlled experiment properly in Part II. For now, hold the mechanism: the expensive read is the world; the cheap decision is the interrupt bar.

Without a baseline, novelty is universal

Here is the diagnostic that should be painted on the wall of every agent ops room:

• Every email is significant when there is no baseline.
• Every network alert matters when you do not know this interface flaps twice a week.
• Every customer request needs an answer when you do not know that “we do not have a policy” is itself a fact about the company.

Everything looks novel when nothing came before. That sentence diagnoses loudness without insulting the model. The model is doing what models do in an empty universe: treat the present as the first event.

Hold that distinction. This book is not a hymn to muteness. An agent that never speaks because it has no world is not exhibiting negative cognition. It is exhibiting missing software. Wise silence is comparative. It is the residue after a diff against a past.

The trust ledger

Think of operator attention as a ledger with a brutal exchange rate. Each interrupt debits trust. Each true, specific, well-timed interrupt can credit it — but only if the account was not already overdrawn by trivia. The kitchen-plug class of alert works because the prior balance was high. The same sentence, delivered by a chatterbox, gets swiped away with the rest of the noise.

So when product managers ask for “more engagement” from an agent, ask which ledger they are optimising. Engagement with the stream is often the enemy of engagement with the one message that matters.

What to build before you tune personality

Teams love to edit the system prompt’s tone: warmer, briefer, more assertive, less eager. Tone is not baseline. You can make an amnesiac more polite. You cannot make it less surprised by ordinary life.

Build the baseline first. Encode normality. Encode significance. Encode what is already known. Only then does silence become a judgment rather than a shrug. And only then does the rare interrupt inherit the trust that makes it operationally useful.

People who have watched agents invent corporate policy often say something true and incomplete: in the absence of information, it will make things up. True. Incomplete, because “absence” is doing too much work. There are at least two different epistemic states hiding under that word, and systems that cannot tell them apart will keep producing confident fiction.

The two-absences table

This table is the artefact. Later chapters will use it; they will not rebuild it.

Why a retrieval-miss is a Rorschach test

When a RAG stack returns nothing useful, the agent still has to act. Humans under the same pressure invent stories; models do it fluently. The empty result might mean:

• The company truly has no discount policy
• Retrieval failed on wording (“commercial flexibility” vs “discounting”)
• The document exists in a silo the agent cannot see
• Access control stripped the hit
• The “policy” lives only in a sales director’s head
• General knowledge should fill the gap (the most dangerous interpretation)

Similarity search cannot represent “what I failed to retrieve” as a typed object. It returns chunks or it does not. Groundedness metrics then measure faithfulness to whatever was retrieved — which is cold comfort when the retrieval set is empty or off-topic.¹ Microsoft’s own GraphRAG writing is blunt that baseline RAG struggles to connect the dots across disparate facts.² That is a relationship problem. Known absence is a different problem: you need a place in the world where non-decision is stored on purpose.

Manners versus physics

The industry’s favourite fix is manners:

DON'T HALLUCINATE DISCOUNT POLICIES!!!
Be accurate. If unsure, say you don't know.
Follow company policy at all times.

Manners are not nothing. They slightly bias generation. They do not create a fact about the company. Under a live customer ask — twelve percent off, multi-site deal, margin pressure in the room — the model still has to produce tokens. General priors about “typical discounts of five to fifteen percent” are sitting right there in the weights. Manners ask the model to police itself against its own fluency. That is a weak leash.

Physics is different. Physics is a page in the world the agent must read:

# Discount Policy

Status: No formal policy exists.

Current practice is discretionary.
See [[Sales Authority]].

Open question:
A formal discount framework has been discussed but not approved.

Now the agent is not being scolded out of invention. It is being informed. The absence is part of the corporate past. The honest answer becomes the high-probability continuation because the world contains it.

Don’t hallucinate is manners. The absence page is physics.

A good wiki gives the agent permission not to know

This is the counter-intuitive gift. We usually sell knowledge systems as ways to know more. For agents, half the value is legalising ignorance. Humans already do this: “We don’t have a policy for that; take it to the sales director.” That sentence is organisational competence. Agents are rarely given the substrate to say it without sounding broken.

Permission not to know is not vagueness. It is a specific speech act backed by a page:

• What is undefined
• What practice exists instead
• Who holds residual authority
• What open question would close the gap

Chapter 6 walks a full discount scenario end-to-end. Here, lock the principle: if your corpus only stores positive claims, your agent will treat every gap as an invitation to generalise. If your corpus stores gaps, your agent can refuse invention without refusing usefulness.

Why invention compounds

In multi-step and multi-agent systems, a confident early fiction becomes everyone else’s context. Agents tend to align with a fluent assertion rather than push back, so a hallucinated “policy” introduced at hop one can lock into false consensus by hop three.³ An absence page is a circuit breaker. It does not merely protect a single answer. It protects the chain.

What to write this week

You do not need a perfect enterprise graph to start. You need the absences that already burn you:

1. List the last ten times an agent (or a junior human) invented “how we do things.”
2. For each, write a one-page known absence or open question.
3. Link the residual authority (“who decides when this is undefined”).
4. Force the agent to read those pages before policy-shaped answers.

That is not knowledge management theatre. That is installing physics where you currently have manners.

Science fiction keeps returning to the same device: implanted memories, childhoods that never happened, a past installed so a being can function in the present. Blade Runner is the obvious reference. The device is easy to misuse in an AI essay. We are not arguing that corporate agents need feelings, trauma arcs, or synthetic birthdays.

We are arguing for the architectural analogue.

What prehistory contains

A compiled organisational (or personal) past answers questions the model cannot responsibly invent:

That second column is why Chapter 3 matters. Prehistory that only stores victories and approved policies is a highlight reel. Agents need the gaps. Humans carry the gaps in corridor knowledge. Agents get nothing unless you compile them.

BI for soft data

Traditional BI works because organisations eventually agree on the semantic layer for hard facts: revenue, region, product, period. Nobody asks every dashboard developer to rediscover what “revenue” means. Agents are being asked to rediscover the soft layer on every call:

• Who matters
• What we believe
• What is in flight
• What is normal
• What policy does not exist
• What everyone quietly knows

Call it BI for soft data if the phrase helps executives. Architecturally it is the same move: a common representation so cognition is not rebuilt from scratch for every question. Agents make the need more urgent than human BI ever did, because humans arrive with years of invisible context and agents arrive with a timer.

Six months versus seventeen seconds

We complain that onboarding takes months. Of course it does. A new hire is downloading a gigantic invisible package: meeting history, gossip, mistakes, manager reactions, customer scars, watching how others escalate, the real meaning of “that’s not how we do it.” Traditional onboarding to full productivity is often measured in months, not days.⁴

Then we deploy an agent:

You're live in production in seventeen seconds.
Answer instantly.
Be accurate.
Don't hallucinate.
Here's a 1,400-word system prompt.

What did we expect?

The prompt is not a childhood. It is a laminated mission statement handed to a stranger at the door of a working company. Worse: a large share of institutional knowledge is unique to the individuals who hold it — when they leave, colleagues cannot simply perform that portion of the work.⁵ Agents inherit none of that osmosis unless someone compiles it.

The corporate wiki is institutional pre-thinking performed once and persisted — so every agent invocation does not reconstruct the company from scratch.

Elsewhere we have called the single-pass failure mode context starvation: asking a model to understand and solve at once without a navigable landscape. A compiled past is the organisational fix. Understanding is capitalised. Solving rides on top.

Prehistory is not lore

There is a temptation to treat this as storytelling — write a charming “about us” for the bot. Resist. Prehistory that works for agents is:

• Typed enough to walk (entities, edges, status fields)
• Current enough to diff against today’s events
• Honest about absences and open questions
• Linked to residual authority when formal policy is missing
• Readable as text — the model’s home turf — with pointers back to exhibits

Claims-and-edges thinking from modern wiki-graph practice is the right shape: not a blob of prose nostalgia, but a maintained map with lint for contradictions and orphans.⁶ You are building the intermediate representation of the organisation’s soft world.

Personal and corporate are the same machine

The Gmail agent that becomes brilliant after a personal wiki and the corporate agent that stops inventing discount policy are not different products. They are the same architecture pointed at different exhaust: years of mail versus years of deals, policies, and corridor norms. Chapter 5 and Chapter 6 are the two faces. Hold the unity now: corporates need a compiled organisational world so their agents are not dumb; people need a compiled personal world for exactly the same reason.

This chapter is a production receipt, not a parable. I gave an agentic runtime early access to Gmail. It was hopeless at triage. After it could read a wiki compiled from years of mail, it became a genius at the same job — so much so that it hardly ever tells me about any email, which is roughly the right angle. When it does speak, I take notice, because it probably means something.

Hold the constants: same tooling family, same mailbox, same human. The variable that moved was the world the agent could see.

Three hypotheses, one controlled accident

Without meaning to, the evolution of that inbox agent ran the industry’s three favourite theories as an A/B/C test.

What changed in the mechanism

Triage is a diff. “Is this worth interrupting Scott?” is unanswerable without a model of what Scott knows, cares about, and has pending. The wiki is that model. Receipts, newsletters, dormant threads, real changes to live work — they only separate once normality exists.

Cold Gmail agent:

Receipt!
Newsletter!
Someone said "urgent"!

Wiki Gmail agent:

Scott already knows this.
That's routine.
That project is dormant.
This person hasn't written for six months.
Wait — this changes something currently in flight.

That is negative cognition from Chapter 2, made concrete. The agent earns the right to interrupt because it can suppress almost everything else.

The success metric inversion

Product culture still celebrates agents that “surface insights.” Production culture, once burned, celebrates agents that shut up. “Hardly ever tells me about any of my email” sounds like underuse until you have lived with the alternative. The silent agent is not idle. It is spending judgment on suppressions. When it spends an interrupt, the trust ledger still has balance.

That is the personal version of the quality ceiling we will name in Chapter 8. For now, treat the Gmail arc as the existence proof that silence can be the high-end behaviour of a cheap model with a dense past.

The router is the twin experiment

The home AI router that stays quiet for a month and then reports a down mesh point needing physical re-power is not a different idea. It is the same architecture on infrastructure telemetry. Baseline: what is normal for this network. Diff: this point has been down long enough to matter. Actionability: human hands required. Trust: high, because the channel was not training me to ignore it.

You already met the kitchen plug in Chapter 1. Here is the architectural point: personal inbox and personal network are one doctrine. Suppress the unsurprising. Speak when the world stops matching the map. Do not confuse that with an interrupt quota bolted on top of an empty world — quotas are a different book. This is knowledge-driven quiet.

What we are not re-deriving

There is a full economic argument for why wiki-plus-utility can beat frontier-plus-raw on context-depth tasks — intelligence moved from operating expense to capital asset. That argument is the job of Context Arbitrage. Here we only need the behavioural receipt: the past flipped the agent from stranger to genius without a model upgrade story. If you are choosing budgets, read that sibling. If you are diagnosing loudness and invention, stay here.

How to steal the experiment

You do not need my mailbox. You need one recurring agent surface and a honest log:

1. Freeze the model tier for thirty days.
2. Week 1–2: prompt and tool tweaks only. Record interrupt quality and invention incidents.
3. Week 3–4: add a minimal compiled past for that surface (normals + entities + absences). Same model.
4. Compare. If quality jumps without a smarter model, you were not short of intelligence.

Most teams never run step three with discipline. They keep buying hypothesis three.

The bottleneck was never cognition. It was input.

Chapter 5 was personal: inbox and mesh. This chapter is corporate. The architecture does not change. Only the nouns do.

The cold path

Imagine an agent sees:

Customer asks for 12% discount.

Cold agent, no institutional prehistory:

request = discount
typical discount policy maybe 5–15%
answer politely

It will sound commercially reasonable. It may even match something someone said at a conference once. It is still invention. The numbers came from general priors, not from this company’s decisions.

Now add the favourite system prompt:

Be commercially sensible. Preserve margin. Follow company policy.

What policy? The prompt assumes a world that was never compiled. Manners without physics. Chapter 3’s distinction in a single customer sentence.

The world that makes twelve percent meaningful

A mature corporate wiki might contain something like this — not as marketing copy, as working knowledge:

[[Discounting]]

No general discount policy exists.

Historically, sales directors have approved discretionary
discounting for multi-site contracts.

Current gross-margin pressure means discounts above 5% are
unusual.

[[Acme Dental]] has been negotiating a three-site expansion.
Scott previously indicated willingness to trade implementation
fees for a longer contract term.

Sales authority:
- Account managers: no unilateral discount authority
- Sales Director: up to 10%
- CFO: above 10%

Related:
[[Margin Protection]]
[[Acme Dental Expansion]]
[[Sales Authority]]

Now the agent is not smarter. It has a past. The words “12% discount” activate a world: no formal band, discretionary history, margin pressure, a live account context, and a clear authority ladder. Twelve percent is not a vibe. It is above the director’s unilateral band and sits in CFO territory — unless the Acme structure trades fees for term instead.

The absence-page template (ship this)

When even the rich page above is more than you have, start with the minimum physics. This is the definitive template for this book:

# [Topic] Policy

Status: No formal policy exists.

## Current practice
[What people actually do when this comes up.
Be concrete. "Discretionary" alone is weak.]

## Residual authority
When this arises, who decides?
- Role A: ...
- Role B: ...

## Related
- [[...]]
- [[...]]

## Open question
[What would close this gap? Who owns drafting?
What has been discussed and not approved?]

## Sources / exhibits
- [Links to threads, decks, decisions that establish
  that the absence is real, not a retrieval miss]

## Last reviewed
YYYY-MM-DD — [who]

Filled for discounting at minimum viable density:

# Discount Policy

Status: No formal policy exists.

Current practice is discretionary.
See [[Sales Authority]].

Open question:
A formal discount framework has been discussed but not approved.

That page alone changes the agent’s honest continuation from inventing a 5–15% band to something like:

I can’t confirm a standard discount because we don’t currently have a formal policy. I can check the appropriate authority for this case.

Not because the system prompt screamed. Because the world contained the absence. Known absence as first-class fact — see Chapter 3’s table.

Why authority edges matter

An absence without residual authority can still strand the agent in useless honesty: “We don’t know.” Pair every important absence with who decides when this is undefined. That is not execution gating in the hard governance sense (a different book). It is knowledge about the organisation’s decision topology so the agent can route the human conversation correctly.

Corporate and personal are one proof

Put the receipts side by side:

• Gmail: baseline lets the agent suppress; silence becomes the product.
• Discount: documented absence lets the agent refuse invention; honesty becomes the product.

Both failures — loudness and hallucination — were missing-prehistory failures. Both fixes are pages in a world the agent can read. That is the whole thesis under commercial pressure.

Pitfall: positive-only wikis

Teams love to document what they decided. They under-document what they deferred, rejected, or left informal. Agents then treat every undefined as a creative writing prompt. The John West instinct applies to knowledge engineering: the absences you refuse to write are the ones that will poison you. Write them.

Pendant recorders, meeting bots, and always-on transcripts share a limitation: they capture from the day you switch them on. Useful. Incomplete. The identified gap in any life-logging or institutional-memory programme is prehistory — everything that already happened before the mic was hot. That gap is exactly what a wiki pointed at existing exhaust can close. Email, documents, tickets, photos’ semantic twins, old project folders: the distillery is already on disk. Someone has to run it.

Exhaust, not vibes

Compiling a past is not “write a long about-the-company essay.” It is closer to a compiler pipeline:

SOURCE EXHAUST
  email / tickets / docs / decisions / transcripts
        ↓
deterministic structure where natural keys exist
  (threads, ticket trees, account IDs, dates)
        ↓
AI synthesis into claims, entities, edges, significance
        ↓
COMPILED PAST (wiki / soft-data IR)
        ↓
agents diff new events against it

Anthropic’s context-engineering guidance elevates structured note-taking — agentic memory persisted outside the window — and notes that runtime exploration is slower than retrieving pre-computed data.⁷ That is the same economic instinct: pay once to compile, amortise forever on every call. We are not re-opening the full opex/capex argument here. We are saying: backward compilation is how prehistory becomes engineering rather than nostalgia.

Minimum viable past for one agent

Do not boil the ocean. Pick the agent that is currently loud or inventing. Build only what that surface needs.

1. Normals for the channel. What is routine? What does the operator already know? What can always wait?
2. Entities and edges. People, accounts, systems, projects — and how they connect.
3. In-flight work. What is live right now so a change-in-flight can clear the interrupt bar.
4. Absence pages. Top undefined policies and open questions (Chapter 3 table; Chapter 6 template).
5. Residual authority. Who decides when the page says undefined.

That set is enough to flip behaviour on one surface. Densify after the first trust recovery, not before the first page exists.

Write absences on purpose

Positive extraction is the easy half. Models summarise what was said. Humans document what was approved. The load-bearing discipline is interviewing for gaps:

• What do new hires get wrong for six months?
• What do we keep saying “it depends” about?
• Which policies exist only as folklore?
• Which decisions did we deliberately leave informal?
• Where did the last agent invent something embarrassing?

Each answer is a candidate absence page. If you skip this interview, you will build a highlight reel and call it a world. Agents will still invent at the edges.

The densification curve

On empty maps, agents thrash tools: search, grep, fetch, re-search. As the compiled layer densifies, tool calls behave like cache misses against a comprehension cache. Early work is noisy. Mature work is quiet. You can feel the curve without inventing a dashboard religion: fewer exploratory calls, more answers from L1 pages, rarer interrupts, fewer invented policies.

Reflection-style memory research has long suggested that consolidating raw observations into higher-order structure supports long-horizon coherence better than raw logs alone.⁸ Your janitor and synthesis passes are that consolidation for institutions.

Pitfalls

A two-week first cut

1. Pick one painful agent surface.
2. Interview two operators for normals and undefined policies (ninety minutes total).
3. Write ten absence/open-question pages with residual authority.
4. Ingest one year of that channel’s exhaust into entities + baseline claims.
5. Freeze model tier; compare interrupt quality and invention incidents to the prior fortnight.

If nothing moves, your pages are not in the agent’s read path. That is an integration bug, not a refutation of prehistory.

The past is not only what you remember going forward. It is what you are willing to compile from what already happened.

If you only remember one product sentence from this book, make it this: the quality ceiling can be silence. Not because silence is fashionable. Because a sufficiently dense past makes most of the present unsurprising, and surprisingness is what should cost human attention.

Three receipts, one machine

Baseline enables the first two. Documented absence enables the third. Together they are knowledge-driven suppression: the agent suppresses noise and suppresses invention because the world makes both judgments possible.

The best agent doesn’t tell you what happened. It notices when the world stops matching the wiki.

What success looks like in the building

• Interrupts are rare and operators still flinch productively when they arrive.
• Policy-shaped answers either cite the world or escalate with residual authority.
• “Don’t hallucinate” remains in the prompt as manners; the absences carry the physics.
• Model upgrades become secondary procurement choices, not existential hope.
• New agents get useful fast because they compile against a shared past, not a blank universe.

Interestingness itself is a relation to what is already known — a diff, not a property of the incoming item alone. That framing is developed for curators in the newsfeed work; the same diff engine is what makes silence rational here.

Operator trust as the real KPI

Dashboards love message counts. Operators love not being lied to and not being nagged. After a compiled past is in place, measure what they actually feel:

• When the agent is quiet, do I assume competence or death?
• When it speaks, do I act before I argue?
• When it refuses a policy answer, do I trust the refusal?

Those three questions are sharper than most eval sets. They map directly to baseline density, interrupt precision, and known-absence coverage. If silence feels like failure, the past is still thin — or the agent is not reading it. If silence feels like earned quiet, you have crossed the quality ceiling this book is aiming at.

Resist the product instinct to “use the capacity” of a quiet agent by inventing new notification types. Capacity is not the scarce resource. Trust is. Spending spare model time to re-earn the right to interrupt is how good agents become loud again.

Where this sits among siblings

Doctrine is a stack. This book owns one shelf.

What we still refuse to own

So the fence stays sharp on the last page as on the first:

• Not interrupt quotas or alert budgets as primary control
• Not execution gating and authority infrastructure below the model
• Not what to show when the agent does speak
• Not attention sovereignty and routing authority

Those matter. Mixing them here would smuggle three other books into a doctrine that is already complete: give the agent a past.

The charge

When your agent is loud, do not open with a model upgrade. When it invents policy, do not open with a sterner scold. Open with the missing-prehistory diagnosis:

1. Does it have a baseline for what is normal on this surface?
2. Does the world contain typed absences for what was never decided?
3. Can it read those pages before it spends trust or invents physics?

If the answers are no, you do not have an intelligence problem. You have an amnesiac in production.

No event is the first event in the universe. Not if you do the work.

Give your agent a past.