Here is the scene every senior manager already knows. A board paper is due. The dashboards are green. Tickets closed. Documents published. Milestones marked complete. Then someone asks the only question that matters: what actually happened, and where are we stuck?

What happens next is archaeology. A senior person is assigned to “pull the story together.” They sample email threads, skim meeting minutes, re-open SharePoint folders, ask around. Days later they produce a narrative that is sampled, slow, unrepeatable — and gone by next quarter. The warehouse held the outcomes. The why lived in soft exhaust nobody was structured to read continuously.

This book is about a different write path. When knowledge workers use AI seriously, they are forced to state intent, alternatives and blockers as a condition of getting useful help. The organisation’s causal layer starts arriving pre-structured. Friction sensing stops being only weak-signal archaeology and becomes, increasingly, a reading problem. That is an extraordinary management product — and an extraordinary misuse risk if the default response is judgment rather than help.

The thesis this book holds

AI-assisted work manufactures the organisation’s causal layer at birth — because getting useful help requires the worker to state intent, alternatives and blockers explicitly — which turns friction sensing from an inference problem into a reading problem, and raises the governance stakes in exact proportion.

Mine for friction, not guilt.

That slogan is not decoration. It is the product requirement. The management question is never “who is slow.” It is “where are capable people repeatedly hitting an operating-system problem.” Every mechanism in the chapters that follow must survive that test.

What this book assumes and what it owns

Two load-bearing arguments sit outside this volume. That AI deliberation is worth treating as source at all is the work of The Deliberation Is Source. That sessions must be joined to the artefacts they touch — not archived beside them — is the work of Provenance-Coupled Work. Both are assumed here. Cite them; do not re-argue them.

What this book owns is the management surface of that exhaust: born-structured causal exhaust as an economics claim; the four-part weekly view; system-level friction signals; invisible-work recognition as credit (not only capacity replenishment); the canon promotion path; and the help-not-judgment legitimacy contract. It also owns an honesty requirement about its own framing — the pro-worker optimism was a starting constraint, not a neutral discovery — which Chapter 13 states without hedging.

Map of the book

Part I installs the write-path change: the interview happening now; intent, alternatives and blockers as near-declared fields; the study that must be run rather than faked; company-owned surfaces and session briefs.

Part II specifies what managers can read: the four-part weekly view; system-level friction signals; the legitimacy contract; what happens when signals stop being weak; invisible-work credit; the canon promotion path.

Part III pays the proof burden and the install bill: one compiled week with support responses; one invisible investigation made legible; owned optimism and legal boundary; a Monday protocol a platform team can run.

After this book you will be able to

• Specify a four-part weekly management view compiled from AI work exhaust at team/workstream grain.
• Attach a support response to each system-level friction signal — and refuse mechanisms that emit person ranks.
• Describe a recognition path for investigations that leave thin artefacts.
• Design a human-gated canon promotion queue for recurring patterns.
• State, in product language, why monitoring is legitimate only when the observed party is helped.
• Commission the semantic-quality comparison study without pretending it has already been measured.

The deeper corporate opportunity is not capturing more documents. It is capturing the meaning-making process that produced them — and reading it for friction, recognition and learning, under a contract that makes being seen a form of being helped.

Why the slogan is a product requirement

It is easy to treat “mine for friction, not guilt” as a values line for the cover. That would be a mistake. Product teams ship what they measure. If the first dashboard tile is AI adoption by individual, the organisation will optimise for prompting theatre. If the first tile is recurring blockers with owner and support action, the organisation will optimise for unsticking systems. The slogan decides which tile is legal in the design.

Hold a second test alongside the ranking test: does the observed party receive a direct benefit from this signal existing? Help, clarity, resources, redesigned process, reduced obligation, promoted knowledge — those count. A manager’s private curiosity does not. A quarterly performance narrative that weaponises session struggles does not. Chapter 7 will harden the contract; Chapter 1 only needs you to accept that legitimacy is part of the thesis, not an appendix.

Named entities stay out. Any worked example later is composite and labelled. Real employers and individuals are not characters in this book. Where source material might tempt a proper noun, generalise to a role descriptor. That is not politeness theatre; it is how you keep a doctrine reusable and non-surveillant in form.

How to use this book

Read Part I for the economics and the honesty rules. Read Part II for the artefacts and contracts you will actually implement. Read Part III for specimens and the install path. If you are a platform lead under time pressure, start with Chapters 1, 5, 7 and 14, then return for the proof chapters before you demo to executives. If you are a senior manager, start with Chapters 5, 6, 11 and 12 so you can recognise a legitimate compile when you see one.

Throughout, Australian spelling and plain working language. No named real organisations. No fabricated statistics. When a passage wants a number it does not have, it states the shape of the finding instead. That discipline is part of the doctrine: credibility is a product feature for a book about reading organisational truth.

Begin with the hard rule. Hold it when the demos get exciting. The rest of the book only works if that rule still stands on the last page.

For twenty years, “data-driven organisation” rhetoric activated the layer that was already born structured. Schemas existed at write time. Warehouses aggregated what was already queryable. The soft layer — emails, minutes, threads, documents — stayed dark not because nobody cared about the why, but because activating it required comprehension at a unit price no institution could staff continuously.

BI for Soft Data named that asymmetry cleanly. Hard data got an industry because it was born structured. Soft data is the causal layer: the decisions, objections, trade-offs and workarounds that produced the outcomes the dashboard shows. The famous estate composition — IDC’s measurement that 90% of data organisations generated in 2022 was unstructured — is the scale backdrop, not a claim that AI chat is ninety per cent of your estate.¹

The archaeological premise

The parent framework’s practical premise is archaeological. The interview already happened, every day, for the last ten years. People wrote emails, took minutes, argued in threads. The reasoning is in the residue. Cheap reading finally makes a standing comprehension pass affordable. Someone senior is no longer the only person who can “pull the story together” for the board — at least in principle.

That archaeology is still real and still valuable. Long-lived soft estates hold decisions nobody currently employed remembers. Resignations delete uncaptured why. The compile of historical exhaust remains a strategic asset. Nothing in this book asks you to abandon it.

The inversion

AI-assisted work changes the write path, not only the read price. To get a useful answer from a model, a person almost always has to externalise purpose, comparisons, corrections and uncertainty. The interaction rewards articulation. The interview is not only a decade of accidental residue you dig up later. It is happening now — prompted, in real time — and the conversation transcript is the record as it is created.

The interview is happening now — prompted, in real time — and the transcript is the record.

That inversion hits three economics at once:

• Acquisition cost. You are no longer only mining historical exhaust. You are operating an environment that produces high-quality semantic exhaust continuously as a by-product of doing the work.
• Freshness. The causal layer is available this week, not only after a quarterly archaeology project.
• Structure. Intent, alternatives, blockers and open questions arrive closer to declared fields than they do in status email optimised to look green.

What the final artefact still hides

The finished document is often the least semantically useful view for a manager trying to unstick a portfolio. A board paper suppresses messy deliberation. A spreadsheet stores outputs and formulas but not why assumptions changed. A presentation reduces weeks of thought to twelve slides. A proposal may contain reused material from previous proposals. An email records the communicated decision, not every option considered.

The AI conversation frequently says, much more explicitly, that the forecast changed because a legal assumption is not approved; that a recommendation was removed because operations cannot support it; that centralisation was rejected because handoff delay outweighed consistency; that only two columns of an inherited table were recalculated. Those statements contain verbs, causes, contrasts and intent. They are closer to meaning than the final cells.

That is why searching the deliberation is sometimes more useful than searching the Word document for “what did we think we were doing.” The document is optimised for its audience. The session is optimised for getting help. Help requires the worker to say what is hard.

Privacy-by-blocking is also a design choice

Many organisations currently use nothing enterprise-grade because of privacy and sensitive information. Workers use retail tools, and the conversation is trapped where the company cannot govern it for help. Custom agents write to log files nobody treats as a management product. In all three cases the exhaust is lost, dark or inaccessible — not because the causal layer does not exist, but because the capture path was never designed as a legitimate management surface.

McKinsey’s 2025 State of AI survey finds that nearly two-thirds of organisations have not yet begun scaling AI across the enterprise, while roughly one-third report that scaling has begun.² That is a window: schemas and gateway field choices are still soft enough to get right. It is also a risk window: the same softness invites either total blockage or individual-scoring products sold as “insight.”

Chapter 3 specifies the fields that become declared. Chapter 4 specifies the capture stack. Hold the inversion first: the interview is no longer only archaeology. It is a continuous write path you can govern — or lose.

Git for knowledge work — the useful analogy

A good commit message is a semantic compression of a code change. It tells an agent where to look and what the change was meant to accomplish before the agent reads the diff. AI systems already produce more of those waypoints in software because they can write a commit after each coherent change, do not resent documentation, and can describe the relationship between request, change and tests.

A similar mechanism can exist for every corporate knowledge artefact. The Knowledge Work Commit in Chapter 4 is that mechanism. The point of the analogy here is economic: once semantic waypoints exist as a by-product of work, later humans and agents stop reconstructing meaning only from final blobs. They enter through intent and rationale. That is the same acquisition inversion as the interview happening now — just named in engineering vocabulary managers already trust.

Do not over-extend the analogy into a full product catalogue in this chapter. Cognitive Git as enterprise metaphor appears later as orientation, not as a substitute for the weekly view. The load-bearing claim remains: soft causal data can be produced continuously and more structured when AI assistance is the work surface, which strengthens the archaeological programme rather than retiring it.

Freshness changes management tempo

Archaeology is quarterly or annual by cost. Born-structured exhaust can be weekly without heroic sampling. That tempo change matters for unsticking work while the people who hit the friction still remember the context. It also matters for legitimacy: a weekly help loop is visible. An annual archaeology project that becomes a quiet HR evidence lake is not the same product, even if both claim to “use soft data.”

Design the tempo you can staff with support responses. A daily personal feed of everyone’s struggles is not ambition; it is a fence failure. A weekly team compile with help actions is ambition aligned with the thesis.

Be precise about what “born structured” means. It does not mean every prompt is a clean database row. Sessions are noisy. People paste. People ramble. People perform for the tool. What it does mean is that the interaction pulls semantic fields that conventional corporate writing actively suppresses.

A board paper is optimised to look settled. A status email is optimised to look green. A model interaction is optimised to get help — and help requires the worker to externalise purpose, contrast, correction and uncertainty. That is a write-path change, not merely better search over the same old residue.

The fields the interaction rewards

Historically an email thread might contain the following reasoning by accident. An AI conversation frequently contains them as near-explicit statements:

• “Here is what I am trying to achieve.”
• “That approach failed because…”
• “Compare these three alternatives.”
• “Use this assumption rather than that one.”
• “This is still unresolved.”
• “I need an answer before Thursday’s steering committee.”

Those are intent, rejection rationale, alternative comparison, assumption choice, open uncertainty and deadline-coupled dependency. They are exactly the fields a manager needs to see system friction. They are also exactly the fields a surveillance product wants to misuse as personal evidence. Design for the first use; refuse the second. Chapter 1’s prohibition still governs: no mechanism may promote these fields into an individual performance score.

Document is the what; deliberation is the why

The finished artefact tells you what survived publication. The deliberation tells you what the person was trying to achieve; what sources they relied on; what they changed and why; what they rejected; what remains open; what they intended but never completed; and where they got stuck. For software the parallel line is already doctrine: the code is the what; the transcript is the why.

For knowledge work the line is: the document is the what; the deliberation is the why. Why that deliberation is source, and how source-relativity works across compiler stages, is article 189’s argument. Assume it. Link it. Spend the rest of this chapter on field density honesty and management consequences, not on re-deriving source theory.

The Deliberation Is Source

The comparison that must not be invented

The natural claim is that AI-conversation exhaust is denser in explicit intent, alternatives and blockers than matched email and meeting minutes on the same work. That claim is plausible. It has not been measured as a published comparison in this programme. Fabricating a multiplier — “3× more explicit intent” — would be the single most damaging failure mode for this piece. Sibling work has already paid the cost of shipping with hollow citation habits. Do not repeat it.

Until that study ships, speak the shape. Do not invent the rate. The rest of this book proceeds on qualitative economics: AI interaction rewards articulation of the fields managers need. That is enough to design the product. It is not enough to publish fake precision.

Why managers care about declared fields

When intent and blockers are near-declared, the weekly compile can answer “where are we stuck” without reconstructing a board narrative from green milestones. When alternatives rejected are retained, the organisation can stop re-paying the same learning cost. When open questions remain visible while status language hardens, review-depth compression becomes detectable before the artefact ships as false certainty.

None of those reads requires ranking people. All of them require that the capture path keep the fields, the compile aggregate to system grain, and the default response be help. Chapter 4 turns fields into a capture architecture. Chapter 5 turns them into the four-part weekly view.

What the final document still cannot generate

Even a perfect document management system cannot regenerate rejected alternatives that never shipped, plans described and deferred, or the moment a worker realised a legal assumption was not approved. Those facts live in deliberation. The repository of finished files holds survivors. The difference between intended work and shipped work is itself information.

That is why the semantic-quality study codes explicit intent, alternatives and blockers — not “quality of writing” and not “productivity.” You are measuring whether the causal fields managers need appear as first-class language. You are not grading workers. Dual coding and inter-rater agreement exist to keep the study honest; they do not exist to create a person score. If someone proposes reusing the coding scheme as a continuous individual metric, refuse. The study’s ethics inherit Chapter 1’s fence.

Also specify sampling honesty. Matched episodes are hard to find cleanly. Some work is only in AI sessions; some only in meetings; some offline. Report missingness. Report when email is performative and minutes are empty. A study that only samples perfect pairs will overstate every channel. The expected shape still stands as a hypothesis: where both channels exist for the same task, AI sessions should carry more explicit causal language per unit text. Measure it. Do not invent it.

Worked language examples (system grain)

Consider three statements that might appear in a session brief after distillation:

• “Intent: prepare a recommendation on whether to centralise the exception function.”
• “Rejected centralisation because handoff delay outweighed consistency benefit under current staffing.”
• “Blocker: operations cannot support the recommendation without funded run capacity; needs decision before Thursday steering.”

Those three lines are management gold at system grain. They are also easy to misuse as personal evidence if the compile attaches them to a name in a ranking product. The correct product use is portfolio: how many workstreams hit unfunded run capacity this week; which decisions wait on operations authority; which rejected approaches should enter canon. The incorrect product use is a scoreboard of who raised blockers. Chapter 5 and Chapter 6 will keep routing every example back to that distinction until it becomes muscle memory.

A senior manager should not receive a dump of everyone’s prompts. Raw first-person sessions are high-fidelity, noisy and easy to misuse. The product path runs through company-owned surfaces, bronze retention, automatic distillation and compiled views — not through retail chat archives and hope.

Company-owned surfaces

Approved AI tools should operate through a company-owned interface or gateway. That may include a corporate chat and research tool, Office and SharePoint assistants, custom departmental agents, coding tools, document-review agents, and meeting or email assistants. The important contract is not that every tool has the same UI. It is that every approved surface emits the same class of structured work events: intent signals, alternatives and blockers when stated, artefact pointers, and session identity suitable for later join.

Retail accounts and unconnected agents fail that contract. The conversation lands where the organisation cannot govern purpose, retention or access. Custom agents that write only to private log files fail it practically even when they pass it on paper. If the platform team cannot produce a session brief from last week’s workstream, you do not yet have a management surface.

Raw sessions as bronze

The original conversation, tool calls, input pointers, output pointers and artefact versions are retained as source. Not everything is placed into the active institutional wiki. The raw session is the episodic staging buffer: detailed, first-person, high-fidelity and potentially noisy. The institution’s semantic record is the governed layer above it.

That staging discipline matters for legitimacy as much as for architecture. Managers should not browse bronze by default. Security and audit planes may need different access under local law — that is a different plane and not a licence to build a performance product on top of raw prompts.

Automatic session distillation

A deterministic pass removes repetitive tool chatter and oversized payloads. A model then produces a compact session brief focused on intent, decisions, rationale, rejected alternatives, blockers, open questions, commitments, artefacts changed and reusable learning. The raw session remains available; the brief is the searchable pointer. Hundreds of briefs can feed a project-level view without requiring every query to reread every conversation.

Knowledge Work Commit fields

The Word file remains the authoritative artefact for its audience. The spreadsheet remains authoritative for its figures. The Knowledge Work Commit becomes the semantic route into them. How the commit joins to version-level artefact deltas is article 190’s job — Provenance-Coupled Work. Assume the join; do not re-specify the resolver here.

Engagement worlds and promotion (cameo)

Sessions should not float alone. They join the relevant project, client, process, decision, product, workstream, policy, team and prior discussions. That creates an Engagement World: the persistent account of what the group has learned, proposed, rejected, built and verified so far. Different actors can contribute from different surfaces without replaying every other conversation. Role-shaped views over the same evidence prevent rival project histories.

Most sessions remain project history. Some patterns deserve promotion to organisational canon — Chapter 10. The capture stack only makes promotion possible. It does not make every session permanent institutional truth.

If gateway teams drop semantic fields as “noise” while retaining token counts as “observability,” the management product never ships. Article 190 made that case for resource identifiers. This chapter makes the parallel case for intent, alternatives, blockers and friction fields. Capture them at session time, or you will not reconstruct them from final PDFs later.

Why distillation is not optional

Raw sessions do not scale as a management interface. They are too long, too first-person, too easy to misread out of context, and too tempting as a surveillance archive. Distillation is the move that makes a portfolio view possible without requiring managers to become full-time readers of other people’s chats. Deterministic stripping removes grind. A brief extracts the causal fields. The bronze remains for reconstruction when a claim must be checked.

That two-layer design is the same moral geometry as Institutional Memory’s staging buffer: keep episodic detail, do not confuse it with the governed record. Managers live mostly on briefs and weekly compiles. Promotion to canon is rarer and slower. Security may live on bronze under constrained access. Mixing those audiences is how legitimacy fails.

Implementers will ask whether briefs can be wrong. Yes. Models summarise imperfectly. That is why bronze exists, why human review gates canon, and why the weekly view is a management hypothesis generator rather than an automated verdict machine. Wrong briefs are a quality problem. Person scores built on wrong briefs would be a justice problem. Design for the first; refuse the second.

What to require of the gateway this quarter

Platform teams prioritise what executives measure. Put these requirements in the gateway backlog as named stories, not as a vague ethics wish:

• Emit structured session events with intent, alternatives, blockers and friction fields when present in the dialogue.
• Retain tool-call resource identifiers for artefact join (article 190).
• Store raw bronze under restricted access roles separate from manager compile roles.
• Produce session briefs on a schedule suitable for weekly compile.
• Refuse to expose individual rank aggregates in the management API even if the data could support them.

If procurement is evaluating vendors, make absence of these fields a disqualifier rather than a future phase. Phase-two ethics rarely ships. Schema decisions made under partial adoption are sticky. McKinsey’s scaling gap is exactly when organisations freeze the wrong observability model — tokens and latency without meaning fields.

The useful product is not a transcript browser. It is a compiled management view a head of function can use on Monday without reopening every session. Four parts. System grain. Support attached to friction. Learning retained.

That is not traditional activity reporting. It is a continuously compiled account of the organisation’s current cognitive state.

1. What moved this week

• Workstreams advanced
• Important artefacts created or materially changed
• Decisions made and their rationale
• Assumptions overturned
• Deliverables approaching review or approval

This is deliberately not a timesheet. Artefacts changed, decisions made with rationale, assumptions overturned: those are cognitive events. A manager reading this section should leave knowing where judgment was spent, not how many hours tools were open.

2. Where the organisation is sticking

• Recurring blockers across teams
• Unresolved dependencies
• Repeated demand for the same scarce specialist
• The same problem being solved independently in several places
• Decisions repeatedly reopened
• Work waiting on authority rather than execution

Every row here is system-addressable. Same specialist everywhere is a capacity and knowledge-distribution problem. Work waiting on authority is a decision-rights problem. Independent re-solve is a reuse and communication problem. None of those rows is a person score. Chapter 6 expands the friction catalogue; this part of the weekly view is where those signals land as a portfolio picture.

3. Where management can help

• Additional resources
• Priority or scope clarification
• A cross-team decision
• Access to a missing system or data source
• A reusable platform capability
• De-scoping or deadline renegotiation
• Escalation to the appropriate owner

This section is the legitimacy payload. Every sticky signal should propose a support action a manager can take. A compile that surfaces friction without help is incomplete. A compile that surfaces people to blame is illegitimate. Chapter 7 hardens that contract.

4. What the organisation learned

• New reusable methods
• Rejected approaches worth remembering
• Emerging client or market patterns
• Undocumented constraints discovered
• Canon entries proposed
• Related prior work resurrected

Without this fourth part you only fight fires. With it you stop re-lighting the same ones. Chapter 10 turns proposed canon entries into a human-gated promotion path. The weekly view nominates; it does not silently rewrite how the organisation works.

Monday ritual, not slide template

Walk the four parts as a weekly ritual. On Monday a head of function should be able to answer: what actually advanced; where the same friction appeared more than once; which interventions are management’s job rather than the team’s; and which learnings must not die in a project folder.

If the compile cannot answer the four without reopening raw prompts, the product is incomplete. If it answers them by naming who was “slow,” the product is illegitimate. Chapter 11 will produce one full specimen. Hold the shape first so the specimen is recognised as doctrine, not anecdote.

How the four parts connect as one instrument

Read the four parts as a single loop, not four independent widgets. Movement without sticking is a victory lap that misses the operating system. Sticking without help is surveillance. Help without learning is permanent firefighting. Learning without movement is a wiki nobody uses. The weekly ritual forces the loop closed: what advanced, where it jammed, what management will do, what must not be forgotten.

Grain discipline matters inside each part. Prefer workstream and team language. Prefer control names and decision classes. Prefer artefact classes over personal identifiers in the management plane. If a sticky row cannot be stated without naming a person as the problem, rewrite the row until the system defect is visible — missing owner, missing canon, missing capacity, missing authority. Sometimes a person is involved; the product still addresses the role and the control, not a score.

Also resist the temptation to add a fifth part called “AI usage metrics.” Usage volume is a platform health signal for the gateway team, not a management morality play. If you need gateway capacity planning, keep it in the platform console. Do not smuggle it into the cognitive-state compile where it will be misread as performance.

Chapter 11 will show the four parts filled. Until then, treat empty help sections and person-shaped sticky rows as failed compiles even if the prose looks polished.

Example sticky row rewritten three ways

Bad (person-shaped): “Alex is stuck on the policy again and keeps asking the bot.”

Better (still incomplete): “Policy questions recurring in Workstream A.”

Doctrine-complete: “Policy ambiguity X re-explained in four workstreams; same verifier role named; support: canon entry + interpretation authority decision; unit: control environment, not individuals.”

Train compilers — human or machine — to emit the third form. If your automatic weekly view cannot produce the third form reliably, keep a human editor in the loop until it can. A polished wrong row is worse than a rough right one because it trains leadership to expect person-shaped explanations.

What good looks like after six weeks

After six weekly cycles on a pilot workstream, you should see: fewer repeated explanations of the same ambiguity; at least one canon entry shipped; at least one authority clarification; recognition call-outs that teams accept as fair; and no open request from leadership for individual ranks derived from the compile. If leadership is asking for ranks, the demo narrative failed. Re-anchor on Chapter 1 before you expand the pilot.

Share the weekly view template with teams before the first automatic compile. Co-design reduces the sense that sensing was done to them rather than with them. Co-design does not mean every sticky row needs a committee; it means purpose, unit of analysis and help obligation are known.

Dashboards see outputs. Soft exhaust can see behaviour around decisions. When that exhaust is born structured, friction stops being only a vibe in hallway conversation and becomes a portfolio signal with a support response attached.

Organisational behavioural telemetry already named the layer: where people are uncertain, where attention accumulates, where nobody wants ownership, where dissent is compressed, where a routine decision consumes abnormal cognitive energy, where a consequential decision consumes suspiciously little. The radar fuses weak signals into nominations about pathways — not people.

This chapter catalogues friction signals that a weekly compile can surface from AI work exhaust. Every signal is a hypothesis about the operating system or the obligation-to-capacity ratio. None is a loyalty score. Chapter 1’s prohibition still binds: if a product manager proposes a per-person friction score, kill the feature.

The signal catalogue

Born-structured exhaust can reveal where:

• Routine tasks require abnormal cognitive effort — session briefs show long alternative trees and repeated clarifications on work the organisation treats as ordinary.
• The same specialists are pulled into every issue — multiple workstreams name the same verifier or domain owner as the only path to unstick.
• Discussion volume grows around supposedly settled work — decisions marked closed keep reopening in deliberation.
• Review depth is compressing — challenge language shortens while status language hardens and open questions remain.
• Work moves outside ordinary hours — team-aggregate session timing shifts, not a person time-clock.
• Preventative work is displaced by urgent work — briefs show deferred hygiene, skipped characterisation, postponed documentation.
• Obligations have expanded without additional capacity — more systems, more assurance requests, shorter windows, same funded headcount.

Obligation-to-capacity, not who is slow

Green by Heroics asks the systems question underneath: where has the obligation-to-capacity ratio changed without a corresponding change in resources? Signals fall into families — work expansion, effort expansion, quality compression, human-capacity deterioration — and they support a capacity hypothesis at team or function level. The intended output is a support response, not an individual score.

Read that fence carefully. These signals produce a hypothesis about the control and the operating model. They must not produce a ranking of individuals, a loyalty score, or a covert performance case. The moment capacity sensing becomes personal surveillance, you have abandoned the pro-human purpose.

Support responses as first-class outputs

A signal without a support path is incomplete product design. Default responses include: provide help; clarify authority; add resources; redesign the process; remove repeated friction; reduce obligation; promote reusable knowledge. Chapter 7 elevates that list into the legitimacy contract. Chapter 11 will attach concrete responses to a full weekly specimen.

For now, internalise the design test. If a vendor demo ranks individuals by “stuckness,” walk away. If HR asks for individual AI-usage percentiles for performance season, the answer is no — and the architecture must make that answer easy by never emitting the field into the management plane.

Chapter 8 explains why this fence matters more, not less, when exhaust becomes stronger. Stronger signals raise governance stakes in proportion. Design for the stakes.

How signals compose into a capacity hypothesis

Single signals mislead. One late-night session cluster might be a launch. One specialist mention might be appropriate expertise. The compile looks for recurrence, co-occurrence and trend: the same specialist across unrelated incidents; effort expansion alongside quality compression; preventative work disappearing while obligations expand. That composition is how you earn a team-level capacity hypothesis rather than a gossip narrative.

Organisational behavioural telemetry supplies the vocabulary for path quality: whether objections were answered or merely stopped; whether evidence changed before status changed; whether independent groups formed independent views. Born-structured exhaust makes some of those questions easier because alternatives and blockers are stated more often in the clear. It does not make them automatic truths. Humans still interpret. The product still nominates.

When you present a friction signal to leadership, pair it with the support response and the unit of analysis in the same breath. “Four workstreams re-explain policy X; propose canon entry and interpretation authority; team-level only.” That sentence is harder to weaponise than “people are stuck on policy X.” Language is part of the fence.

Signals and the seven shapes (cameo only)

Institutional Failure Radar develops seven shapes institutional failure takes in soft exhaust before metrics move. This book does not re-derive them. It notes that born-structured exhaust can make some shapes easier to spot earlier — reopened decisions, compressed challenge, abnormal effort — because the causal language is more explicit. Use the radar’s nomination discipline: evidence-backed questions about pathways, receipt-bearing, human disposition. Do not upgrade a shape label into an automated risk traffic light about a person.

If your assurance function already runs a radar programme, the weekly AI-exhaust compile is a new sensor feed into the same moral geometry, not a parallel performance system. Integrate at the nomination layer. Keep the fence shared.

Keep a living list of illegitimate transforms your organisation has rejected: stuckness ranks, adoption scorecards, prompt-count productivity, after-hours individual heatmaps used in performance talks. Publish the reject list next to the signal catalogue. Culture needs both the allowed instruments and the banned ones in writing.

What a sceptical works council actually asks

Friction catalogues sound benign until someone asks how the feed could be turned on a person. Answer the hard questions in product language, not slogans.

“If you can see abnormal cognitive effort, can you rank who is struggling?” Only if you emit person grain. The product must not. Aggregate to workstream and team. Count how many workstreams show abnormal effort on work labelled ordinary — not how many sessions one person opened. If leadership asks for a personal stuckness list, that is a product change request to refuse, not a filter toggle to enable.

“Outside-ordinary-hours signals look like time tracking.” They do if you keep individual clocks. Keep only team-aggregate shifts: the portfolio moved toward evenings during a control window; preventative work disappeared while exceptions rose. Never attach a name to a late session in the management compile. Bronze may still hold timestamps for security under local rules; the management plane must not promote them into a person score.

“Same specialist everywhere is just blaming the expert.” The illegitimate reading is character. The legitimate reading is knowledge concentration and missing canon: four workstreams lack a durable interpretation artefact and a decision right. Support is documentation, authority design and capacity — not a performance conversation about the expert. If your sticky row cannot be rewritten without a person as the problem, rewrite the row.

“How do we know the signal is real without measuring rates?” You do not invent rates. Recurrence, co-occurrence and human disposition are the method: the same blocker appears across workstreams; the same support class is implied; a manager can act without reopening raw prompts. Chapter 3 already forbids fabricating density multipliers for intent language. The same honesty applies here: describe patterns and support responses; do not ship fake precision about how often friction appears per thousand words.

Protocol: dispose a friction signal in one sitting

When a weekly compile surfaces a signal, run this short protocol before the meeting ends:

1. State the signal in system language (control, workstream, decision class).
2. State the evidence type (recurring briefs, reopened decision, specialist re-entry) without naming individuals as the defect.
3. Attach one support response with a management owner and a due window.
4. Record what must not be concluded (no person rank, no AI-usage morality play).
5. Decide whether the pattern is a canon nomination, a capacity hypothesis, an authority defect, or noise to drop.

If step three is empty, the signal is not ready for leadership. Sensing without help is how friction products become guilt products. Chapter 7 hardens that contract; this chapter only insists the catalogue is useless without the dispose protocol.

The legitimacy question is not optional polish after a clever compile. It is the load-bearing constraint. If being observed does not produce help for the people who generate the exhaust, you have built surveillance with better language models.

Being seen is only legitimate when being helped.

Reframe the management question

The management question should not be “who is slow?” It should be “where are capable people repeatedly encountering an operating-system problem?” That reframe is not soft. It is a unit-of-analysis rule. Capable people hit bad process, missing authority, scarce knowledge and unfunded obligation. The compile should make those hits visible so management can act on the system.

Default response catalogue

Used positively, the default response to a signal becomes:

• Provide help
• Clarify authority
• Add resources
• Redesign the process
• Remove repeated friction
• Reduce obligation
• Promote reusable knowledge

That list is the legitimacy payload of Chapter 5’s third part. If your weekly view has a “sticking” section with no corresponding help section, you have half a product. If your help section is empty while individual annotations flourish, you have the wrong product.

Prohibited outputs

The management product must not emit:

• Individual AI-usage percentiles or “adoption scores” used for performance
• Personal “stuckness” ranks or continuous contribution scores from session volume
• Automated discipline triggers from session content
• Leaderboards of who prompted most or least
• Keystroke-style productivity overlays sold as AI insight

Bronze identity may still exist for security and audit under local requirements. That is a different plane. The management compile must not promote identity into a score. Chapter 1’s prohibition is product architecture, not a values poster.

Why the fence is empirical, not only moral

Conventional workplace monitoring already carries a wellbeing cost. The American Psychological Association’s Work in America work found employees in monitored environments more likely to feel burnt out, emotionally exhausted and less motivated than those not monitored.³ Experimental work published in Harvard Business Review found employees more likely to break rules when they knew they were monitored than when they were not — a mechanism that includes reduced felt responsibility.⁴

Gartner estimated that technology-based worker monitoring in large businesses rose from about 30% before the pandemic to roughly 70% afterward — a shift reported in secondary synthesis of workplace surveillance studies.⁵ Those findings are not a reason to abandon organisational sensing of systems. They are a reason to refuse the individual-scoring path that produces the harm.

A one-page policy sketch (not legal advice)

If your vendor cannot support these constraints, you do not have a fair implementation. You have a monitoring product wearing insight language. Chapter 8 shows why stronger signals make that distinction more urgent, not less.

Response obligation as a measurable practice

Make the response obligation operational. For each sticky row in the weekly view, require a named support action, an owner on the management side, and a due window. Track whether actions closed. If signals accumulate while help never ships, legitimacy is failing in practice even if the policy PDF is beautiful. Workers will notice. They will stop externalising blockers. The sensing layer will rot.

Transparency of purpose also matters. People should know the compile exists, what it is for, and what it is not for. Secret AI-session analytics sold as “insight” recreate the trust failures already documented in monitoring research. Purpose limitation is not only a legal concept; it is how you keep the write path honest. Workers who believe sessions feed performance scores will write different sessions — performative, thin, less useful for help and for learning.

Finally, separate the right to improve systems from the right to judge persons. The first is the product. The second is prohibited here. Organisations that need performance management already have processes for that; they must not silently ingest AI deliberation as a new evidence channel without explicit, lawful, bargained design — which this book does not attempt and does not endorse as a default.

What “help” looks like in the first ninety days

Early help actions should be boring and visible: named owners, unblocked access, short canon pages, deadline renegotiation, specialist time protected by documentation rather than heroics. Publish a monthly “help shipped” list alongside friction trends. Workers should be able to point to something that got better because the compile existed. If they cannot, you are extracting signal without returning value — the definition of illegitimate monitoring in this doctrine.

Resist help theatre: leadership messages that say “we care” while the only concrete outputs are more dashboards. The contract is operational. Resources, authority, process redesign, obligation reduction, knowledge promotion. Count those.

If you need a single sentence for the board: we will read AI work exhaust to find where capable teams hit operating-system problems, and we will measure success by help shipped and learning retained — not by scoring individuals on how they use the tools.

Works-council objections, answered honestly

The prohibited-outputs list earlier in this chapter is the product fence. Sceptics still ask how the fence holds when power is asymmetric. Meet the questions without inventing legal regimes this book does not own.

“Why should we trust a help default if management can change the dashboard later?” Trust the schema more than the slogan. If individual rank fields do not exist in the management API, quiet conversion is harder. Publish the reject list. Give employee representatives a standing review of the compile template. Require a visible product change — with consultation appropriate to your jurisdiction — before any person-grain field is added. Soft policy alone is weak; missing fields are stronger.

“Bronze still holds first-person sessions. Isn’t that the real risk?” Yes, bronze is sensitive. That is why managers do not browse it by default, why access is exceptional and logged, and why the weekly product is a compile at team grain. Confusing bronze retention with a performance archive is exactly the drift this contract forbids. Security and audit planes may need constrained access under local law; they must not quietly become the management product.

“What stops a manager from opening briefs and writing a private scorecard anyway?” Culture and process still matter. Product can make misuse harder: no rank endpoints, purpose notices, help-action logs, and a rule that performance processes do not ingest AI deliberation by default. Product cannot make misuse impossible. Owning that limit is more honest than promising total technical prevention. The doctrine still forbids designing the easy path to person scoring.

“If help never ships, is the whole programme illegitimate?” In practice, yes — even if the policy PDF is perfect. Response obligation is not poetry. Track help actions closed. If friction rows accumulate while support stays empty, pause expansion until the loop works. Workers will stop stating blockers if stating them only feeds a dashboard. The sensing layer dies when legitimacy dies.

“Can we use deliberation in promotion packets?” Only through human-mediated recognition with local consent and process rules you design with counsel — not through a continuous algorithmic contribution score. Chapter 9 develops credit without ranking. Here the contract answer is: narrative credit may be legitimate; automated person scores derived from sessions are not this product.

A one-page legitimacy test before go-live

1. Can a team member state the purpose of the compile in one sentence that matches the product?
2. Does every sticky row template require a support response field?
3. Does the management API expose any individual rank or stuckness endpoint? If yes, remove it before pilot.
4. Is raw session browse off by default for managers?
5. Is there a published help-shipped channel back to the teams that generate exhaust?
6. Is the legal/HR path named as a separate workstream (not claimed complete by this doctrine)?

If any of the first five fail, you are not ready. If the sixth fails, you still need counsel — but you must not pretend product doctrine replaced jurisdiction. Name the boundary; stop; staff the legal work properly.

Institutional Failure Radar fuses weak behavioural signals into nominations for human review. Discussion volume, compressed dissent, abnormal effort, what disappears between thread and pack — none of those alone proves failure. Together they can justify examining a pathway. The moral geometry is correct: nominations about systems, not accusations about people.

Born-structured exhaust changes the detection problem. Intent, alternatives and blockers arrive closer to declared fields. Detection moves from inference over weak residue toward reading explicit structure. That is a stronger sensing layer. It is also a more dangerous one if misused.

When the signal stops being weak, the governance stakes rise in exact proportion.

What “stronger” means in practice

Weak signal: three teams send more messages about a “closed” decision. You infer reopening. You may be wrong. Stronger signal: three session briefs explicitly reopen the same decision, state the same missing authority, and list the same blocker. You are reading, not only inferring. The nomination is still a nomination — humans still dispose — but the evidence density is different.

That difference is the book’s big claim about governance. Stronger exhaust does not license looser ethics. It demands tighter unit-of-analysis rules, tighter access and a default response path that cannot silently become individual discipline. Chapter 7’s contract is not a soft introduction; it is the proportional response to stronger sensing.

Declared state and felt state

Formal systems record the declared state: status codes, completed checklists, authorised roles. Soft exhaust samples the felt state: uncertainty, ownership avoidance, compressed dissent, abnormal cognitive effort. Green means the process completed, not that the work is healthy. Organisational behavioural telemetry reads the path to green, not only the light.

AI work exhaust does not abolish felt state. People still perform. People still hide. What it does is make more of the felt state legible when workers need help from a model badly enough to say what is hard. That legibility is the asset. It is also the hazard. A performance culture that punishes visible struggle will teach people to stop externalising blockers — and you will lose both the help and the sensing.

Complementary instruments

Keep the instruments distinct:

• Institutional Failure Radar — weak-signal fusion over soft pathways; nominations about deformation before metrics move.
• Green by Heroics / obligation-to-capacity — team-level capacity hypotheses; no individual scoring.
• Institutional Linter — static analysis of codified procedures and controls; complementary, not a person score.
• This book’s weekly compile — born-structured exhaust read for friction, help, recognition and learning under the help-not-judgment contract.

Do not collapse them into one “AI monitors staff” product. Collapse is how the fence fails. The radar’s geometry — systems not people — remains mandatory even when signals are strong. Stronger evidence about a pathway is still not a licence to score a person.

Design for proportional stakes

Practically: raise the bar on aggregation as evidence gets more explicit; require support responses before friction tiles ship; separate bronze access from management compile; human-gate anything that becomes durable canon; never ship individual rank fields even as “optional.” Optional fields become mandatory culture. If the signal is strong enough to help a team quickly, it is strong enough to harm a person quickly. Build for the first; refuse the second.

Chapter 9 turns from sensing harm-avoidance to the positive claim: invisible cognitive work made creditable. Recognition is new ground relative to capacity replenishment. Hold the stakes claim as you move: credit is still not a continuous algorithmic score of persons.

Proportional controls checklist

When signals strengthen, tighten controls in the same release train — not in a later ethics retrofit:

• Raise default aggregation grain as explicitness rises.
• Require support-response fields before friction tiles go live.
• Split bronze access from management compile access in the permission model.
• Human-gate durable canon; never auto-promote person judgments.
• Delete individual rank fields from the schema rather than hiding them in the UI.
• Log management access to raw sessions as exceptional, not routine.

These are engineering tasks. They are also the proportional governance answer to the book’s strongest claim. If you only ship stronger sensing, you have shipped half the system — the half that can harm. Pair every sensing increment with a fence increment. Chapter 7 gave the contract language; this chapter insists the stakes scale with signal strength so nobody can claim surprise later.

A short dialogue that shows the stakes

Platform lead: “We can now see explicit blockers in most AI sessions. Should we build a personal stuckness index?”

Doctrine answer: “No. Build a recurring-blocker index at workstream grain with mandatory support responses. The stronger the blocker field, the more valuable the system index and the more abusive the person index. Proportional stakes: ship the fence in the same sprint as the field.”

That dialogue should be written into platform RFCs. When the field lands without the fence, you have already chosen the high-stakes path without paying for governance.

Strong signals also change false-positive economics. Weak-signal fusion needed many weak cues before nomination. Explicit blocker fields can nominate sooner. That speed is valuable for help and dangerous for over-reaction. Keep human disposition. Keep pathway grain. Speed is not a reason to skip the fence.

Composite week: when the signal is strong enough to read

Here is what “reading rather than inferring” looks like in one week’s four-part compile when blockers and alternatives arrive in the clear.

What moved

Two client packs advanced; a control interpretation note was revised after an assumption was withdrawn; a dashboard proposal moved from “full join” ambition to a narrow pilot recommendation. Movement is visible as decisions and overturned assumptions — not as hours-in-tool.

Where sticking (strong signals)

Three workstreams independently stated the same blocker in session briefs: no named owner for a cross-team data definition. Two workstreams explicitly reopened a decision marked closed last month, each listing the same missing authority. Review briefs on one high-stakes pack showed open questions retained while status language hardened. These are not weak vibes about “busy” or “tense.” They are near-declared fields: blocker, reopened decision, unresolved questions under green status.

Where management can help

Name the data-definition owner by Friday, or admit the decision right does not exist. Put the reopened decision back on a real forum with the missing authority present. Protect review time or cut scope on the high-stakes pack. Each help line maps to a sticky signal; none maps to a person score.

What learned

Force-merge of three datasets rejected again with rationale; policy ambiguity re-explained; canon nominations raised for both. Learning is portfolio property.

What a manager must NOT conclude

• Not: “these three people cannot decide.” Conclude: decision rights and owners are missing at system level.
• Not: “this reviewer is soft.” Conclude: schedule is compressing challenge on a consequential pathway.
• Not: “AI adoption is uneven, rank the laggards.” Conclude: nothing in this compile licenses individual AI-usage ranking.
• Not: invent a rate of how much denser these sessions are than email. Chapter 3 still governs: shape and recurrence, not fabricated multipliers.

Stronger exhaust made the sticky rows faster to draft and harder to dismiss as anecdote. That is exactly why governance stakes rose: the same clarity that enables help enables abuse if person grain sneaks in. Ship the fence with the feed.

Much valuable knowledge work produces little visible output. A worker might spend two days establishing that a proposal cannot work; reconciling conflicting interpretations; locating an undocumented dependency; discovering that three datasets cannot be safely joined; persuading several groups not to take the obvious path; narrowing a problem until the eventual document becomes simple.

The final artefact may be one paragraph: “Option B is recommended.” The AI conversation may contain the actual contribution — the alternatives walked, the constraints found, the people unstuck.

The final artefact may be one paragraph. The conversation may contain the actual value.

What recognition can show

Capturing deliberation makes invisible work legible. It can show that someone:

• Investigated the difficult alternatives
• Identified the hidden constraint
• Saved the team from a poor decision
• Repeatedly supplied scarce domain knowledge
• Helped several other people get unstuck

That improves institutional learning. It can also improve recognition — credit for cognitive work that thin artefacts erase. Organisations already know, vaguely, that some of their best people “make hard things look easy.” They rarely have a receipt. Promotion packets fill with visible deliverables. Investigations that prevented a bad path leave almost nothing to point at.

Differentiation: capacity vs credit

Recognition without algorithmic person scores

Credit a contribution in a human process: a manager cites the investigation in a narrative; a team records a “saved us from Option A” entry; a promotion case includes a deliberation-backed example with consent where required. Do not emit a continuous “contribution score” derived from session volume. The fence is the same fence: system learning and human credit, not algorithmic ranking of persons.

Under a recognition path, the investigation — not the four sentences — is what the organisation values and retains. Under a performance-score path, the thin artefact looks like low output. That contrast is the whole point. Thin artefacts under a scoring regime punish the best investigations. People learn to ship noisy slides instead of simple conclusions earned the hard way.

Invisible work and the weekly view

Part four of the weekly view — what the organisation learned — should explicitly surface investigations that prevented bad paths, not only deliverables that shipped. Part one — what moved — can include “assumption overturned” and “option rejected with rationale” as first-class movement. Those rows are how recognition becomes operational rather than sentimental.

Chapter 12 walks one composite investigation end to end. Hold the doctrine here: credit is positive, system-aligned and human-mediated. It is not a back door to individual scoring with friendlier vocabulary.

Why thin artefacts are often a success

In knowledge work, simplicity in the final note can mean the hard work already happened. Narrowing a problem until the answer is short is a craft skill. Scoring systems that reward bulk output will systematically punish that craft. Born-structured exhaust is one of the few scalable ways to keep a receipt for narrowing work without forcing people to produce theatrical intermediate decks.

Recognition culture still needs human judgment. Not every long session is valuable. Not every short artefact hides brilliance. Managers must still know the work. What changes is the evidence available when they look: alternatives walked, constraints found, other teams unstuck. That evidence supports fair recognition. It does not replace management with an algorithm.

Keep credit processes narrative and intermittent rather than continuous and scored. Continuous scores recreate the monitoring harms Chapter 7 cited, even if you rename them “contribution.” Narrative credit — in 1:1s, promotion cases, team forums — can use deliberation-backed examples without turning the organisation into a live ranking feed.

Invisible work types to teach managers to see

• Negative results: proving an approach cannot work.
• Interpretation labour: reconciling conflicting definitions across systems.
• Dependency discovery: finding the undocumented lag or join key collision.
• Coordination labour: persuading groups off the obvious path.
• Problem narrowing: deleting complexity until the artefact can be short.

Each type leaves thin artefacts and thick deliberation. Each type is easy to under-credit. Each type is exactly what born-structured exhaust can make legible if the recognition path exists. Teach the typology in management onboarding for teams adopting AI surfaces. Otherwise managers will keep reading only the final paragraph.

Finally, invisible work includes teaching and unsticking others. When session briefs show the same person repeatedly supplying scarce domain knowledge to many workstreams, the system read is knowledge concentration and missing canon — with recognition for the teaching labour and capacity relief for the teacher. The wrong read is either hero worship without system change or quiet resentment that one person is “blocking.” Credit and redesign together.

Composite week: recognition inside the four-part view

Walk the same instrument as Chapter 5, this time with recognition as a first-class outcome rather than a soft afterthought.

What moved

Assumption overturned: a full three-system join for an exception dashboard is unsafe under real volumes. Option B (narrow segment join) recommended with residual risk listed. A second workstream withdrew a board-pack recommendation operations cannot support. Thin artefacts; thick movement in judgment.

Friction signal that surfaced

Invisible investigation made legible: multi-day session briefs hold rejected alternatives, colliding keys, conflicting status definitions, and persuasion notes that kept two other groups off a force-merge. The published note is a handful of sentences. Without deliberation, the week looks light. With deliberation, the week prevented a bad operational path.

Support response implied

• Credit the investigation in the team forum and, where process allows, in narrative promotion evidence — not via a contribution score.
• Retain the rejection rationale as a canon nomination so the next workstream does not re-pay the learning cost.
• Name residual-risk owner and lag-dependency decision so help is system change, not applause alone.
• If the same scarce verifier role appears across unsticks, fund documentation and authority design rather than endless heroics.

What a manager must NOT conclude

• Not: “low output week” from artefact length alone.
• Not: “too many model turns means inefficiency” — turn counts are not a performance metric in this doctrine.
• Not: a continuous personal brilliance score derived from session volume.
• Not: public shaming of long exploration; that teaches people to hide uncertainty and kills born-structured exhaust.

Recognition and friction share a fence. Both use system grain and human mediation. Both refuse algorithmic person ranks. The positive claim of this chapter is only that credit is available for work thin artefacts erase — and that the weekly view can carry that credit without becoming a scoreboard.

Most sessions stay project history. That is correct. If every chat becomes institutional canon, the organisation drowns in first-person noise and the legitimacy contract fails. The durable write path must be selective, reviewed and typed.

What deserves promotion

Some patterns deserve a nomination pass:

• A recurring client objection
• A repeated operational constraint
• A common spreadsheet mistake
• A policy ambiguity repeatedly explained
• A dependency that surprises every new team
• An approach tried and rejected several times
• A workaround quietly becoming standard practice

The system can review accumulated AI conversations and nominate reusable rules. Humans dispose before anything becomes team or organisational canon. Nomination is not publication. Publication is not automatic.

Second-order learning as the ritual underneath

Software teams already have a lightweight extraction ritual: after work, ask where you stuck, what surprised you, and what you should never do again. Born-structured exhaust makes those questions answerable from the work itself rather than from a heroic retro that never gets scheduled. The nomination pass is that ritual at estate scale — still human-gated before anything becomes “how we work here.”

Staging buffer and governed record

Institutional Memory describes the composition: episodic staging buffer, consolidated governed record, and a promotion path between them. Raw sessions are bronze. Briefs are pointers. Weekly views are management compiles with shorter default life. Canon is durable after review. Collapse those layers and you either lose learning or over-retain raw first-person material under a management gaze it never earned.

Anti-repetition is the institutional payoff

Without promotion, three teams re-reject the same approach for the same reason and never know the others paid the learning cost. With promotion, a deprecated-but-visible claim stops the 2027 proposal from re-living the 2019 failure. Individual memory prevents an agent from re-trying yesterday’s dead end. Institutional memory prevents the company from re-living last decade’s dead end after the people who remember have left.

That is why Part four of the weekly view matters. It is not a nice learning slide. It is the feed into the only durable write path that turns born-structured exhaust into organisational advantage without turning it into permanent individual surveillance.

Part III now pays the proof burden: one compiled week, one invisible investigation, owned optimism, and a Monday install path.

Nomination quality over nomination volume

An eager system will over-nominate. Every repeated phrase will look like canon. Humans must dispose with a bias toward precision: prefer patterns with clear rationale, clear recurrence, and clear system benefit. Prefer anti-patterns and constraints over slogans. Prefer claims that future workers can apply without re-reading ten sessions.

Contested facts should remain contested. If two parts of the organisation disagree, do not flatten the disagreement into a false consensus claim. Preserve the debate as structured disagreement where needed. Promotion is not a way to launder politics into timeless truth. It is a way to stop re-paying settled learning costs and to surface constraints that otherwise live only in heads.

Dream-time review — offline passes over accumulated briefs — is the right compute pattern. Do not interrupt workers mid-session with canon bureaucracy. Do not make every user a wiki editor by force. Extract, nominate, dispose, publish. That pipeline keeps born-structured exhaust from becoming either a dark lake or a spam wiki.

Worked promotion example (composite)

A composite pattern: four workstreams independently rediscover that a status code means different things in two systems. Nominations fire. A human editor accepts a canon entry: definitions, examples, anti-pattern (do not join on status alone), owner of the definition. Next month’s weekly view shows fewer re-explanations. That is promotion ROI without any person score.

If the same pattern were “promoted” as a note that certain teams are confused, you would have converted learning into stigma. The dispose step exists to prevent that conversion. Train disposers to reject person-shaped nominations even when the underlying sessions are real.

Version canon entries. When a constraint changes, supersede rather than delete, so agents and people can see why the rule moved. Born-structured exhaust will keep nominating outdated constraints if the governed layer cannot show what replaced them.

Worked instance: promotion across one composite week

Monday — weekly compile nominates, does not publish

Part four of the weekly view lists three candidates:

1. Status code “closed” means different things in System A and System B (surfaced in four workstream briefs).
2. Force-merge of three exception feeds rejected for false operational action (surfaced in two workstreams independently).
3. A workaround using a manual spreadsheet bridge is becoming standard for a monthly control.

None of these is canon yet. None names a person as the defect. The compile only feeds the nomination queue.

Wednesday — disposer session (human)

A named human disposer — not an automated merge — opens the queue with a thirty-minute pass:

• Accept (1) as a short canon card: dual definitions, examples, anti-pattern “do not join on status alone,” definition owner role, link back to bronze briefs as provenance pointers.
• Accept (2) as a deprecated-but-visible approach: when force-merge was tried, why it failed, what narrow join remains acceptable with residual risk.
• Park (3): the spreadsheet bridge may be a capacity smell, not a method to bless. Request a capacity hypothesis (obligation vs funded process) before any “standard practice” claim. Do not promote a workaround that hides understaffing into eternal process truth.

What the disposer explicitly rejects

A noisy auto-nomination that read “Workstream Gamma struggles with data quality.” That is person- or team-stigma shaped. Rewrite or drop. Canon is for constraints, methods and anti-patterns — never for judgments of who is confused.

Friday — publish and close the loop

Accepted cards land in the governed layer with version dates. Next week’s sticky section should show fewer re-explanations of the status-code trap if the card is findable from the AI surface and the project world. Help shipped includes linking the card into the company assistant’s retrieval set — otherwise promotion is a wiki page nobody reaches.

What this instance proves

• Promotion is a pipeline with human dispose, not a pile of chats labelled “knowledge.”
• Weekly part four is the feed; canon is the durable store; retention and access differ.
• Parking is a success mode: not every recurrence deserves institutional blessing.
• System grain held: no individual score, no stigma entry, no fabricated frequency statistic for how often the ambiguity appears per thousand words.

Run this loop every week on a pilot portfolio before you claim an institutional learning system. If nominations never get human time, you have built a graveyard of good intentions. If dispose turns into person commentary, you have left the doctrine.

Minimum proof burden: one compiled weekly view produced from real session patterns, with the friction signals it surfaced and the support response each implies. What follows is that specimen.

Four-part compile — Portfolio Ops (composite week)

What moved

• Client exception-queue analysis advanced from problem framing to a constrained recommendation.
• Policy interpretation pack for the payments control materially revised after legal assumption withdrawn.
• Two workstreams overturned the assumption that three legacy datasets could be joined for a dashboard.
• Steering pack for Thursday approached review with three open questions still unresolved in deliberation.

Where the organisation is sticking

• Same policy ambiguity re-explained across four workstreams; same human expert named as only verifier.
• Three sessions deferred action pending a cross-team owner who has not been named.
• Two teams independently rejected the same integration approach for the same latency reason.
• Review challenge language shortened on a high-stakes deliverable while status language hardened.

Where management can help

• Fund a one-page canon entry for the policy ambiguity; decide who may interpret it without re-taxing the expert.
• Name the cross-team owner by Friday — or admit the decision right does not exist.
• Promote the rejected integration approach to shared dead-end log before a third team re-pays it.
• Protect review time or re-scope; do not treat green status as evidence that challenge happened.

What the organisation learned

• Rejected join strategies for the three datasets with residual risk on the narrow join that worked.
• Operations cannot support the recommendation that was removed from the policy pack.
• Canon nomination: policy ambiguity that surprises every new team.
• Related prior work on exception routing resurrected from an earlier engagement world.

Friction table — read every row

Walk the rows

Scarce specialist. The exhaust does not say the expert is a bottleneck personality. It says four workstreams lack a durable interpretation artefact and a decision right. Help is canon plus authority design. The illegitimate read — “expert is unhelpful” — is exactly what Chapter 6’s fence forbids.

Waiting on authority. Analysis progressed. Action did not. The open field is ownership, not effort. If management cannot name an owner, the control environment is incomplete. Calling the teams passive is a category error.

Independent re-solve. Two teams paid the same learning cost. That is institutional memory failure, not individual inefficiency. Promotion to canon is the support response. Shame is not.

Review depth compressing. Status language hardened while challenge shortened. Declared state is greening; felt state is thinning. Protect time or cut scope. Do not celebrate the green light.

Invisible investigation. Thin artefact, thick deliberation. Recognition path and dead-end retention. Chapter 12 expands this row into a full case.

Abnormal effort on routine work / preventative displacement. Obligation-to-capacity territory. Team-level capacity hypothesis. Support is redesign, tooling or funded capacity — never a person score for “struggling with AI.”

What this specimen deliberately omits

Names ranked by stuckness. AI adoption scores per person. Time-on-tool heatmaps as performance evidence. Fabricated percentages of “explicit intent density.” The semantic-quality study from Chapter 3 is still a study to run, not a number inserted into this table. Shape statements beat fake precision.

A works council reading this artefact should see the unit of analysis without a lawyer translating. Legal regimes still sit outside this book. The product doctrine is already readable as system-level help.

How a manager should run the meeting

Do not present the weekly view as a tribunal. Present it as an operating review of the system. Open with what moved so the room sees progress before friction. Walk sticky rows with support responses already drafted. Assign management owners for help actions in the room. Close with canon nominations and recognition call-outs for investigations that left thin artefacts.

If someone asks for the “who” behind a sticky row, answer with role and workstream grain, then redirect to the system defect. If someone asks for individual AI metrics, refuse and point to the product rule. If someone wants raw prompts projected, decline unless a specific reconstruction is required under a constrained access path. The meeting is where culture either enforces the fence or dissolves it.

After the meeting, publish the help actions back to the teams that generated the exhaust. Closing the loop is how being seen becomes being helped. Without the loop, the specimen is only a clever table.

Mapping rows to chapter doctrine

Each row in the specimen is an instance of earlier doctrine. Scarce specialist and independent re-solve are Chapter 6 signals and Chapter 10 promotion candidates. Waiting on authority is a help-contract test from Chapter 7. Review depth compression is felt-state sensing from Chapter 8. Invisible investigation is Chapter 9 recognition. Abnormal effort and preventative displacement are obligation-to-capacity territory shared with Green by Heroics, still without individual scoring. The specimen is not a new framework. It is the frameworks running on one week of composite exhaust.

Minimum proof burden: one documented case of invisible cognitive work made legible. Prefer consent-backed real cases when they exist in the source pack. Here the source material gives the pattern; we present a clearly labelled composite rather than invent a person or employer.

The thin artefact

What eventually published was roughly four sentences: Option A is unsafe under real volumes; a narrow join is possible with residual risk; Option B is recommended for the pilot; residual risk must be owned by the data steward. To a milestone dashboard, that looks like a light week. To a scoring culture, it looks like low output.

What the sessions held

Day 1 — Intent and first failures

Intent stated: determine whether the three datasets can support a unified exception dashboard without false merges. First approach: join on customer key as used in the operational system. Failure: key collisions under real volumes; model helped enumerate collision classes; worker rejected “just dedupe” as silent data loss.

Day 2 — Alternatives

Three alternatives compared: (1) force-merge with confidence scores; (2) restrict to a narrow product segment; (3) abandon unified dashboard for dual views. Force-merge rejected: false positives would drive wrong operational action. Dual views rejected for the pilot goal. Narrow segment retained as candidate with explicit residual risk.

Day 3 — Hidden constraints

Conflicting field definitions discovered across systems (status codes overloaded; “closed” means different things). Undocumented dependency: one feed lags by a business day. Sessions record persuasion work: two other groups had wanted the obvious full join for a slide. Deliberation holds the argument that convinced them not to.

Day 4 — Outcome and open questions

Outcome: drafted recommendation for Option B (narrow join) with residual risk list. Open: who owns residual risk; whether lag can be funded; whether canon should record the rejected force-merge. Final note remains short because the problem was narrowed until the answer could be simple.

Recognition path vs scoring failure mode

Chapter 9’s differentiation still holds. Green by Heroics would ask whether obligation outran capacity on this team. That question remains valid. This chapter asks whether the investigation is creditable. Capacity replenishment and recognition are related; they are not the same claim.

How the weekly view should carry this

• What moved: assumption overturned — full three-dataset join unsafe; Option B recommended with residual risk.
• Sticking: residual risk owner unnamed; lag dependency unfunded.
• Help: name residual-risk owner; decide lag funding; stop other teams proposing force-merge.
• Learned: rejected force-merge with rationale; field-definition conflicts; lag constraint; canon nomination for join anti-pattern.

Recognition checklist (human-mediated)

1. Session briefs retain alternatives rejected and constraints found.
2. Manager narrative can cite the investigation without opening raw prompts in a performance meeting.
3. Dead-end enters promotion queue so the organisation does not re-litigate Option A.
4. No continuous contribution score is derived from session volume or turn count.
5. Consent and access rules for using deliberation in promotion packets follow local requirements — legal treatment out of scope here; product must not require illegal use.

Credit the investigation or the institution will keep punishing depth. People will learn to optimise for artefact thickness rather than problem narrowing. Born-structured exhaust makes the better equilibrium available — only if recognition is designed as help for the meaning-making process, not as a new scoreboard.

What consent-shaped recognition looks like in product terms

Even without writing jurisdictional law, product design can avoid forcing illegal or abusive use. Prefer team-level celebration of prevented bad paths. Prefer opt-in inclusion of deliberation excerpts in promotion packets. Prefer manager access to briefs over broadcast of raw sessions. Prefer redaction of unrelated personal content. Prefer purpose notices that state recognition and learning uses, not performance scoring.

The composite case shows why those choices matter. The analyst’s value was in rejection and narrowing. Any system that only counts shipped pages will mis-rank that week. Any system that publicly shames long exploration will teach people to hide uncertainty from the model — and then the born-structured advantage collapses. Recognition design protects the write path that makes sensing possible.

Connecting the case to the three proof burdens

This chapter satisfies the invisible-work proof. Chapter 11 satisfied the compiled weekly view. Chapter 3 specified the semantic-quality study without faking its result. Together they meet the minimum proof burden without inventing numbers or naming real people. If a later edition runs the study and gathers consented cases, replace composites with measured results and consented narratives — still under the same fences.

Managers reading this case should ask on their own portfolios: which thin artefacts last month hid thick investigations? If they cannot answer, the recognition path is not yet real — regardless of how polished the friction dashboard looks.

From case to operating habit

A single composite proves the shape. An operating habit needs a weekly slot. Add three standing questions to the portfolio review that already uses the four-part view:

1. Which thin artefacts this week hid thick investigations?
2. Which rejected approaches must enter the nomination queue so we do not re-litigate them?
3. Where did recognition and capacity relief both need to fire (credit the work and fix the system that made heroics necessary)?

Answer in system language. If the only answers you can form require naming individuals as defective, you are sliding toward guilt. Rewrite until the defect is a missing owner, missing canon, missing capacity or missing authority. Role descriptors may appear in recognition narratives (“the policy analyst role that owned the join investigation”) without becoming a stuckness rank.

Failure modes unique to invisible-work programmes

• Trophy theatre: celebrating investigations in all-hands while never funding residual-risk owners or canon entries.
• Score creep: starting with narrative credit, then asking engineering for a “simple” contribution index from session volume.
• Privacy theatre: locking bronze so tightly that even consented, redacted recognition examples cannot be prepared — then declaring recognition impossible.
• Thickness bias: continuing to reward long decks in promotion culture while the compile quietly proves short notes can encode deep work.

Counter each failure with a product or process rule already in this book: help-shipped logs, no rank endpoints, purpose notices, human-mediated credit, and weekly part-four learning rows that treat rejection as movement. The invisible-work case is not a sentimental interlude. It is the proof that born-structured exhaust can reprice depth upward — if the institution refuses to reprice it as a person score.

Hand the case to Chapter 14’s install checklist: recognition call-out slot, canon queue, and schema test against individual ranks must all be live before you claim the proof is operational rather than literary.

One honesty requirement remains, and it will be checked for specifically. The pro-worker framing in this book was not a neutral finding that emerged at the end of analysis. It was a deliberate starting constraint.

We wrote this piece to the brief: show the fair implementation and the upside. That optimism was a constraint we accepted at the start, not a finding we discovered at the end.

What the author asked for

In the source turn that originates the corporate application, the author is explicit: think about the upside and the solution and the good parts; do not spend the page only nitpicking problems. The framing is constrained toward fair implementation — mine for friction, not guilt; make monitoring positive for the people observed. Presenting that optimism as if it were an emergent, value-free conclusion of the analysis would be dishonest.

Owning the optimism does not mean pretending the failure mode is impossible. Privacy, misuse and power imbalance are real. Conventional monitoring already correlates with worse wellbeing outcomes, as Chapter 7 cited. The design answer is to build the fair path first-class: born-structured exhaust, four-part weekly view, invisible-work credit, canon promotion, and a hard ban on individual performance measurement. That is the product. The pessimist’s path — score people with AI chat logs — is the failure mode we refuse to ship as the default.

Why now (with real, dated sources)

McKinsey’s 2025 State of AI survey finds that nearly two-thirds of organisations have not yet begun scaling AI across the enterprise, while roughly one-third report that scaling has begun.² Adoption is wide enough that privacy and gateway decisions are being made; scale is incomplete enough that schemas are still soft. Capture fields can still be kept — or dropped as noise.

Nielsen Norman Group’s synthesis of three studies found generative AI tools increased business users’ throughput by 66% on realistic tasks — one reason knowledge-work AI will keep expanding where it is allowed.⁶ Use that figure as adoption pressure, not as a licence for individual performance management of AI use.

Gartner’s survey of 5,141 employees (April–July 2024) found only 23% of digital workers completely satisfied with work applications in 2024, down from 30% in 2022.⁷ Corporate AI surfaces will keep being redesigned. The exhaust those surfaces emit will either be governed for help or lost to retail tools and dark logs.

Name the legal boundary and stop

Series position without self-link

Article 189 established deliberation as source. Article 190 established the join. This volume established the management read and the legitimacy contract. Live neighbours in the series already own route-invariant grounding, wiki redundancy, novelty-preserving design, inbound edges, derivational provenance, gold that addresses reality, and the AI partner as challenger. Link them when you need them; do not invent URLs for unpublished work. Do not self-link this piece inside its own body.

Chapter 14 turns doctrine into a Monday install path. The optimism remains a constraint on design: ship the fair path as the product, not as a footnote after the scoring feature ships.

How to talk about optimism without naivety

You can hold two thoughts at once. First: AI work exhaust can be a pro-worker sensing and recognition layer if designed that way. Second: the same exhaust can be a potent individual surveillance layer if designed that way. This book chooses the first as the product to specify, because the author constrained the brief that way, and because the second path is already crowded with monitoring vendors.

Choosing the fair path is not a prediction that organisations will behave well. It is a design commitment: make the fair path the path of least resistance in the schema, the UI, the access model and the weekly ritual. Make the scoring path require an explicit, visible, contested product change — not a quiet dashboard toggle. That is what it means to own optimism as engineering, not as rhetoric.

A note for sceptical readers

If you came for a pure risk catalogue of AI workplace surveillance, this book will feel incomplete on purpose. The author constrained the assignment toward fair implementation and upside. Risks are acknowledged with real sources; mechanisms are fenced; legal depth is deferred. What is not deferred is the positive design. Read it as a specification for the product that should exist if organisations are going to retain AI work exhaust at all.

Document the constraint in your own programme notes when you implement: “We are building the fair path by design choice, not claiming neutrality.” Future readers of your architecture decision records deserve the same honesty this chapter requires of the book.

Why owning the constraint makes the piece more usable

Readers who build platforms live inside contested politics: privacy officers, works councils, security, ambitious vendors, executives who want both insight and control. A book that pretends its pro-worker design fell out of value-free analysis is harder to use in those rooms, not easier. Opponents will smell the framing anyway. Allies will not know how to defend it.

Owning the constraint does three practical jobs.

First, it sets the evaluation criteria. Success is help shipped, learning retained, friction reduced at system grain, and recognition that does not become a score. Failure is a polished individual AI-productivity dashboard. If you hide the constraint, people will evaluate you on vendor default metrics and call the fair path “incomplete analytics.”

Second, it licenses refusal. Product managers need a sentence that is not only personal preference: the programme was scoped to fair implementation; person ranks are out of scope by design. That is cleaner than endless case-by-case ethics debates after a demo has already shown a leaderboard.

Third, it improves trust with sceptics. Counter-intuitively, stating “we chose the upside path on purpose” is more trustworthy than smuggling optimism as neutral science. Sceptics can then argue with the design choice, demand fences, demand legal process — all of which this book invites — without also accusing the text of fake objectivity. Honesty about framing is part of the legitimacy contract applied to the book itself.

None of this weakens the empirical anchors that remain: monitoring wellbeing research, monitoring backfire experiments, unstructured-estate scale, partial enterprise AI scaling, and the unmeasured status of intent-density comparisons. Constraints choose which product to specify. They do not invent statistics. Chapter 3’s refusal to fabricate a density multiplier still governs every expansion in this volume.

When you brief a steering committee, lead with both sentences: the causal layer is increasingly available at birth when people work with AI; and we are specifying the fair management product under an explicit pro-worker design constraint, with system-level only rules and a hard ban on individual performance fields. That pairing is usable. False neutrality is not.

Doctrine without an install path becomes a thoughtful PDF. This chapter is the operational close: surfaces, compile, fence — in an order a platform team can run.

Seven steps

1. Company-owned AI surfaces that emit the same structured work events (intent, alternatives, blockers, artefacts touched). Join details per Provenance-Coupled Work.
2. Raw sessions retained as bronze; automatic distillation into session briefs; no manager dump of raw prompts by default.
3. Weekly compile into the four-part view, aggregated to team/workstream grain.
4. Friction signals with attached support responses, not person ranks.
5. Recognition path for investigations that leave thin artefacts.
6. Human-gated canon promotion queue for recurring patterns.
7. Explicit product rule: no individual performance field derived from AI exhaust.

Two-week pilot protocol

Cognitive Git as larger picture — not a substitute

The full enterprise metaphor can wait until the weekly view works. Conversations as working sessions; Knowledge Work Commits as semantic waypoints; Engagement Worlds as project branches; institutional memory as the governed main branch; promotion as reviewed merge; bronze as reconstructable history; managers receiving role-shaped checkouts of the current world. That architecture is real and larger than this book. It is not a substitute for the manager’s artefact or the help-not-judgment fence. Ship the artefact and the fence first.

The cost of inaction

Without this compile, next quarter’s board narrative is again a senior person sampling threads. The scarce specialist is again re-taxed for free. Option A is rejected again by a new team that never saw last quarter’s dead end. The person who spent two days proving a join was unsafe is again invisible because the artefact is four sentences. And somewhere a vendor is demoing an individual AI-productivity score as if that were the innovation.

The doctrine in this book is the alternative: born-structured exhaust, read for friction not guilt, with governance stakes named in proportion to how strong the signal has become — and with optimism owned as a design constraint rather than smuggled in as fake neutrality.

The deeper corporate opportunity is not capturing more documents. It is capturing the meaning-making process that produced them.

Mine for friction, not guilt. Make invisible work creditable. Promote what recurs. Fence the individual score. Help first. That is the Monday job.

Failure modes to watch in the first month

• Prompt dump demos that thrill executives and terrify staff.
• Empty help sections while friction sections grow.
• Person-shaped rows that escape review into the compile.
• Canon spam from over-eager nomination without human dispose.
• Gateway field loss where intent and blockers are dropped as noise.
• Parallel retail tools that recreate dark exhaust outside the company surface.

Assign an owner for the fence as well as an owner for the features. If only the sensing features have a product manager, the fence will lose every prioritisation meeting. Legitimacy is a feature. Ship it with a name, a test, and a veto.

When the pilot works, expand by workstream, not by individual metric. Keep grain stable. Keep support responses mandatory. Keep Chapter 13’s honesty available when someone asks whether this is “just the optimist’s story.” Yes — by design. The fair implementation was the assignment. This book is the specification of that assignment made concrete.

Checklist you can paste into a project tracker

• [ ] Gateway emits intent/alternatives/blockers/friction fields
• [ ] Bronze retained; manager compile role cannot list raw sessions by default
• [ ] Brief job running; quality sample reviewed weekly
• [ ] Four-part view template live for pilot workstream
• [ ] Support-response required field on sticky rows
• [ ] Recognition call-out slot in weekly meeting
• [ ] Canon nomination queue with human disposer named
• [ ] Schema test: no individual rank endpoint exists
• [ ] Team-first pilot readout before wider management demo
• [ ] Help-shipped log published back to pilot team

When every box is checked, you have not finished enterprise transformation. You have finished a legitimate pilot of born-structured exhaust read for friction, not guilt. Expand from there.

Closing

You now have the thesis, the hard prohibition, the economics inversion, the study shape, the capture stack, the four-part view, the friction catalogue, the legitimacy contract, the strong-signal stakes claim, the recognition gap, the promotion path, two proof specimens, owned optimism, and a Monday checklist. The work that remains is organisational: ship the fair path as the default product. Mine for friction, not guilt.

Return to the thesis when scope debates arise. AI-assisted work manufactures the organisation’s causal layer at birth. That fact raises governance stakes. The only proportional product is system friction, recognition and learning under a help-not-judgment contract. Everything else is either incomplete or unsafe.

The most actionable page: next ten working days

If you do nothing else, run this ten-day sequence on one pilot workstream. Keep composites out of production demos: use your real sessions under your real purpose notice, but keep reporting at team grain only.

Days 1–2 — Surface and purpose. Confirm the workstream uses a company-owned AI surface (or pause until it does). Publish a one-paragraph purpose notice: system friction, help, recognition, learning — not performance scoring. Name the fence owner and the compile owner as two roles, even if one human temporarily holds both.

Days 3–4 — Capture fields. Verify session retention and brief distillation for the pilot. If automation is incomplete, hand-build briefs from sessions using the Knowledge Work Commit fields (intent, alternatives, blockers, friction, learning). Do not invent density rates. Do not open a person score spreadsheet “just for the pilot.”

Day 5 — First four-part view. Draft what moved / sticking / help / learned for the week. Every sticky row gets a support response with a management owner. Show the draft to the pilot team before any wider management audience.

Day 6 — Legitimacy check. Ask the team: did being seen produce help, or only exposure? If exposure only, stop expansion and fix the help loop. This is a hard gate, not a survey nicety.

Days 7–8 — Recognition and nomination. Call out at least one thin-artefact investigation if it exists. Put at least one recurring pattern into the canon nomination queue with a human disposer named and a dispose date.

Day 9 — Schema and access test. Confirm no individual rank endpoint exists; raw browse remains exceptional; help-shipped log is visible back to the team.

Day 10 — Decision. Either schedule the next four weekly cycles with the same fences, or shut the pilot cleanly and write down why. Do not leave a half-dead sensing feed that trains people to hide blockers.

Escalation paths when politics hits

• Executive wants person ranks: refuse; offer system friction with help actions; cite Chapter 1 prohibition and Chapter 7 prohibited outputs.
• Security wants infinite bronze retention for “HR later”: separate planes; management compile is not an HR lake; take retention to counsel without converting the product purpose.
• Vendor demo shows adoption leaderboards: walk away or contractually disable; leaderboards are not a harmless extra.
• Team freezes externalisation: treat as legitimacy failure; publish help shipped; reduce raw access; re-earn the write path.

You do not need a multi-year architecture programme to start. You need one workstream, one week, one four-part view, one help action closed, one nomination disposed, and one hard no to person scores. That is the close of the book and the start of the work.