You are mid-thought with a thinking partner that can search your own canon. You reach a conclusion that sounds familiar. The partner, helpful as ever, tells you that you have already covered this. Sometimes that is a gift: you forgot; the retrieval worked; the prior claim walks back into the room still warm enough to continue. Sometimes it is wrong in a subtler way. You are arriving at a familiar conclusion by an unfamiliar road. The road is the part that matters. The system matches the final sentence, declares redundancy, and quietly throws away the warrant, the domain of application, and the future use that only the new path opened.

That is the lived problem this book is built to make expensive. It is not a complaint that retrieval works. It is a complaint that retrieval works so well that proposition-level matching becomes the dominant interaction pattern — and that matching is happening at the wrong level.

After this book you should be able to distinguish repetition from re-derivation, and know what to write down when a familiar conclusion arrives by an unfamiliar road. The takeaway is operational, not motivational: a diagnostic table, a named gap in the provenance family you already use, a receipt schema you could implement next week, and a conversational line that forces the three fields that matter.

The thesis in one sentence

A conclusion that matches something already in the canon can still be new knowledge, because the derivation carries the warrant, the domain of application, and the future affordances. A memory system that diffs only propositions will erase the part that changed. The failure is not malice. It is a comparator that thinks identity of wording is identity of thought.

Answer invariance can conceal derivational novelty.

That line is the hinge of the whole series neighbourhood this piece sits in. It is also the place to state a relationship openly, early, and without re-litigating another book.

This book is a correction — say that now

In Route-Invariant Grounding, the reliability argument ran one way: different wiki paths reaching substantially the same answer is a desirable property. Many paths, same genba. That remains true for reliability. You want a substrate that is not brittle to one golden route or one favourite model style. You want evidence and answers that hold when wording, parameters and traversal order shift.

This book is the correction that reliability argument needs. The conclusion may already exist while the path by which you have just reached it is new — and that path may change what the conclusion means, how strongly it is warranted, where it applies, or how it can be reused. For reliability, you want route-invariant grounding. For learning, you also need derivation-sensitive memory. A system that only celebrates the first will erase the second.

What this book will not do: re-argue the five-run route study, the citation-overlap figures, or the measurement battery that article 182 owns. Route into that piece if you need the reliability instrument. Here we own the derivation axis only. A future companion on layer economics sits on a different cut of the same shared conversation; it is not ours either.

What the lag is telling you

The practitioner experience is almost comic in its predictability. Sometimes you really did say it before and forgot. Sometimes you are saying the same thing for a different reason — same conclusion, different angle, different warrant. It takes a turn or two of argument for the partner to see that this is not a path through the wiki. It is a path through thought and logic. That path changes the interpretation of the outcome, the learning, and how the idea can be used next. If the partner can see the same answer, the nuance gets overlooked.

That lag is not a personality flaw in the model. It is an engineering smell. The default match is still proposition-only. High surface similarity triggers the collapse line before any cheap check on observation, mechanism, evidence class or domain. The human has to push for the derivation to become visible. By the time it does, the conversational energy has already been spent arguing for the existence of a difference the system should have looked for first.

What you will build

Part I names the matching error and installs the four-cell diagnostic table. Part II places Derivational Provenance as the missing fifth member of a provenance family already in use, and connects it to Working Fidelity without re-deriving The Clasp. Part III walks two real pairs from a living canon — genuine repetition beside same-conclusion re-derivation — specifies the receipt schema, shows a downstream use only the new route opened, and installs the conversational artefact. Part IV turns the doctrine into policy and a checklist you can run without the author in the room.

The enemy is not compression. Compression is how large canons stay usable. The enemy is compression that believes the final proposition is the whole thought. That belief is false in mathematics, false in engineering, and false in every operational loop where the proof is the product.

The same destination does not make two journeys cognitively equivalent.

Why this matters now

Personal and institutional canons are finally large enough for retrieval to work well. That success creates a new failure mode. When the partner can find prior claims reliably, “you already said this” becomes the default interaction pattern. The more the memory works, the more often proposition-level matching fires. The compression is not a bug in search quality. It is a product of search quality meeting a comparator that treats final sentences as the whole thought.

Builders who only optimise for de-duplication will ship partners that feel tidy and slowly erase learning. Builders who only optimise for novelty will ship concept sprawl. The third path — classification before collapse, receipt on re-derivation — is the operating middle this book codifies.

Hold the reader question through every chapter: how do I tell the difference, and what should the system record when the collapse is wrong? The answer is not a vibe. It is a table, a named provenance gap, a schema, and a conversational line you can train.

When a partner says you have already discussed this, it is usually matching at the level of the proposition:

Existing claim ≈ current claim

That is not a stupid check. Surface identity and near-identity catch real forgetting. Large canons make forgetting more common, not less. The problem is treating that check as a complete epistemic act. Equivalence of wording is not equivalence of warrant. Equivalence of conclusion is not equivalence of construction. Equivalence of destination is not equivalence of journey.

Eight axes the proposition matcher cannot see

The same final proposition can differ along any of the following axes — and more than one can move at once. Each axis is a reason the thought is not the same thought, even when the sentence rhymes with something already indexed.

1. Different observation

You reached it from a different observation. The claim may still read “walk history improves design judgement,” but last time the observation was a written doctrine page and this time it is a live reject-and-replay session. The observation is the first link in the warrant chain. Change the observation and you change what later evidence is allowed to do.

2. Different causal mechanism

You reached it through a different causal mechanism. One derivation may run through architectural argument: if walks are durable, instruments can improve the map. Another may run through operational loop: idea, historical walks, analysis code, measurement, rejection, mutation, implementation, replay. Same slogan. Different causal story. Different places a later agent can intervene.

3. Evidence class moved

New evidence moved the claim from intuition or paper doctrine to observed behaviour. That is not a second concept. It is a status change that most proposition matchers treat as noise. In any engineering culture worth the name, “we wrote that it should work” and “we watched it work end-to-end under harness” are different epistemic grades. Compressing them is not tidiness. It is a silent demotion of the stronger warrant back into the weaker one.

4. Different stakeholder lens

A different lens exposed another consequence. The proposition may hold while the implication field changes for delivery, governance, finance or product. Semantic Refraction owns the full treatment of lens-qualified implications over one bronze event; here we only need the axis: consequence under a role can change without forking the bronze fact into rival realities.

5. New domain of application

The claim now applies in a new domain. A retrieval doctrine that transferred into project management, or a personal-memory insight that transferred into organisational deliberation, is not “already said” in the sense that makes further recording useless. Domain is part of the claim’s future utility. Proposition match without domain tags will erase transfer paths.

6. Boundary conditions became visible

The claim’s boundaries became visible. You still endorse the conclusion, but you now know where it fails: which measure made a plausible idea collapse, which distribution broke a threshold family, which authority boundary the fluent answer still missed. Boundaries are not footnotes. They are half the portable value of the thought. Proposition-only storage keeps the slogan and loses the fence.

7. Independent reappearance

It independently reappeared. Independence is itself evidence about robustness. A claim that only survives when you quote the canonical page is different from a claim that re-emerges from a separate operational loop. Idea Provenance cares who said it when; independent re-derivation is a different signal, closer to corroboration of structure than to authorship credit.

8. Join across previously separate regions

It now joins two previously separate regions of the canon. The proposition may look familiar while the edge is new: IP doctrine meeting live development history, retrieval theory meeting reject-and-replay practice, personal Working Fidelity meeting graph maintenance. The join is knowledge. Matching only the endpoint deletes the bridge.

Why each axis changes next week

None of these differences is decorative colour for writers. Each one changes what the claim can do next week: which transfers are licensed, which objections are already answered, which experiments would falsify it, which neighbouring claims it now pulls into the same room, which rejected alternatives are part of the warrant rather than forgotten noise.

A memory system that diffs only propositions will treat all eight as “already have it.” That is how a helpful thinking partner performs a live version of a failure already named for offline maintenance.

Hallucinated consolidation — offline and live

Hallucinated consolidation is the janitor failure in which two genuinely distinct ideas are merged because they are old, adjacent or superficially similar, leaving a cleaner-looking claim that no source fully supports. The recognition tell is a page that reads more confidently than any of its sources.

When a janitor does it overnight, you can at least review a diff, run a lint pass, and revert. When a thinking partner does it mid-conversation, it arrives as assistance. That makes it harder to notice and easier to accept. The human is still forming the thought. The partner is already filing it under yesterday’s heading. The merge happens in the open, with a smile, before the derivation has even finished landing.

That is why this book refuses the frame “AI memory is sloppy.” The more precise frame is: AI memory is often matching at the wrong level, and the error mode looks like good product behaviour. Helpfulness optimises for de-duplication of sentences. Learning requires preservation of warrant paths. Those two objectives diverge exactly when the canon is large enough for retrieval to work.

How the axes compound

In practice the axes rarely move one at a time. An operational loop can change observation, mechanism, evidence class and boundary together. That is why a single boolean “is this a duplicate?” is structurally underpowered. You need multi-field comparison, or at least a cheap elicitation that surfaces which fields moved.

The partner that only sees lexical overlap will keep winning the easy cases — true restatements — and keep failing the expensive ones. Expensive cases are exactly when learning is concentrated: the week the doctrine met the harness, the day the failed measure exposed a fence, the session that joined two regions of the canon. Those are the weeks you cannot afford to compress away.

Design implication: store tags for evidence class, domain, and mechanism when you can; when you cannot, ask. Missing tags are not a licence to default to redundancy. They are a licence to ask one clarifying question before finalising the match.

Chapter 2 listed the axes a proposition matcher cannot see. This chapter installs the instrument that makes the failure mode visible in practice: a two-by-two on final proposition and derivation. It is small enough to remember in a conversation. It is sharp enough to drive different write behaviour in a memory system. Later chapters will not re-derive it; they will point here.

Table 3.1 — The diagnostic table. Row two is the compression failure this book exists to stop.

Row one — same proposition, same derivation

This is genuine repetition. You said it before, for the same reasons, with the same evidence class, and nothing material moved. The correct behaviour is almost boring: point at the existing claim, offer the prior context, and do not mint anything new. This is where “you already said this” earns its keep. Forgetting is real. A partner that recovers the prior claim is doing the return half of a clasp: the thought comes back so you can continue rather than climb the hill again from a fossil.

What makes row one row one is not surface wording alone. It is that the derivation is also the same: same observation family, same mechanism story, same evidence grade, same domain, same boundaries. If those hold, compression is hygiene. If any of them moved, you are not in row one — even if the sentence is a near-duplicate.

Row two — same proposition, different derivation

This is the heart of the piece. Dwell here.

The final sentence matches something in the canon. The road does not. Something in the derivation changed: observation, mechanism, evidence class, lens, domain, boundary, independence, or join. The claim is not “new” as a marketing slogan. It is newly warranted, newly scoped, newly bounded, or newly connected. That is still knowledge. In many operational settings it is the only knowledge that moved this week — the proposition was already known; the proof just became stronger or different.

What a naive system does: match proposition → declare redundancy → discard the new road. What a derivation-sensitive system does: keep the existing claim page as the address, and attach a receipt that records how the claim was re-reached, when, from which bronze event, and what that changes about future use. Chapter 8 specifies the receipt. Chapter 9 walks a real pair. Here the point is classification: if you cannot see row two as a distinct cell, you will keep treating learning as duplication.

Same answer, different proof.

Why is row two the product of success rather than of failure? Because it only becomes common when retrieval works. When the canon is small, the partner rarely finds the prior claim and the human restates freely. When the canon is large enough for the model to search it, the partner finds the prior claim often — and the collapse line becomes the dominant interaction pattern. Neighbouring pieces in this publishing run address other compression problems: redundancy as error correction rather than pure bloat; novelty-preserving carve-outs in payload design; inbound edges as a different question from outbound ones. Those are real axes. This axis is different. It is not about how many pages say a thing, or which reverse edges arrive, or how much of a bloom payload is novel. It is about whether the thought that produced the proposition is the same thought as last time.

Row two is also where helpfulness and learning diverge most sharply. The helpful product instinct is: reduce repetition, save tokens, move on. The learning instinct is: the repeated conclusion is the cheap part; the expensive part is the new warrant. Compress the expensive part and you have built a system that gets smoother while getting less useful.

Row three — different proposition, same evidence

The exhibits did not change. The synthesis did. Two honest readers of the same bronze can diverge. Here the system should not pretend the second reading never happened, and should not overwrite the first. Contradiction-as-edge discipline applies: keep disagreement as structure rather than averaging it into a smoother paragraph. The memory action is to record that interpretation shifted on shared evidence — a synthesis note, a contested edge, a supersession candidate if one reading legitimately replaces the other. This is not row two. The proposition changed. Pretending it is “the same idea restated” is a different consolidation error.

Row four — different on both

Different proposition, different derivation: genuine unresolved divergence. Do not force a merge. Do not declare one side already covered by the other. Hold both, with provenance, until evidence or decision resolves them — or until they remain permanently as a mapped tension. Many organisations hate this cell because it looks messy. Messy is sometimes accurate. A graph that returns the state of a debate is more useful than a graph that returns a false peace.

Why two behaviours are not enough

If your system only has two behaviours — “duplicate of existing” and “brand new concept” — rows two through four will be misclassified. Row two becomes silent discard. Row three becomes a fight or an overwrite. Row four becomes either noise or a fake consolidation. The table is the minimum instrument that makes those mistakes visible. Chapter 4 turns each cell into write behaviour. For now, own the classification: four cells, not two.

Using the table in conversation

The table is not only a backend classifier. It is a shared language for the human and the partner. When the collapse line appears, either party can ask: which cell are we in? That single question reopens classification without requiring a lecture on provenance theory.

Row two deserves extra training data in any evaluation set. If your test suite only contains restatements and brand-new ideas, your model will look good while still erasing re-derivation. Build pairs deliberately: same sentence neighbourhood, different warrant path. Grade the partner on whether it asks for Z, writes a receipt, or falsely collapses.

Remember the economic story: row two becomes common when retrieval succeeds. Do not treat rising collapse rates as proof the user is repetitive. Treat them as proof the canon is working — and as a demand for derivation-sensitive policy.

Chapter 3 installed the table. This chapter walks it as an operator would: what gets written, what does not, what would count as a wrong action. The full receipt schema lives in Chapter 8; the specimens live in Chapter 9. Here the doctrine is the mapping from cell to behaviour. If that mapping is wrong, no schema will save you — you will just write the wrong thing carefully.

Cell one — same / same: point, do not mint

When final proposition and derivation both match, the system should treat the event as retrieval, not as authorship of new knowledge. Write a retrieval or last-accessed stamp if you care about currency of use. Offer the existing claim, with enough construction that the human can re-enter rather than re-derive. Do not create a second concept page. Do not invent a derivation receipt that pretends novelty where there is none. Do not upgrade evidence status. The virtue here is restraint.

Wrong action in this cell: minting a duplicate page because the human restated the claim with slightly different wording. That is concept sprawl. It makes later retrieval worse, and it trains the partner to treat every paraphrase as a new idea. Another wrong action: refusing to surface the prior claim at all, so the human re-derives from cold every time. That is the opposite failure — an archive that never returns.

Cell two — same / different: keep the address, attach a receipt

This is the cell the book exists for. The proposition is already a citizen of the canon. The derivation is not a rerun. The correct write is not a rival concept. The correct write is a receipt on the existing claim.

Behaviourally, the system should:

• Keep target_claim_id pointed at the existing claim or concept page.
• Create a dated receipt (Chapter 8) with type, bronze pointer, prior route, new route, and what changed.
• Optionally raise evidence_status_after when the new route genuinely upgrades the epistemic grade.
• Surface a conversational line that names X, Y and Z rather than declaring redundancy (Chapter 11).
• Leave the receipt queryable so later agents can answer boundary and transfer questions the bare proposition cannot answer.

Wrong action A: silent discard. Match the proposition, emit “already covered,” write nothing. This is the default of many memory features and the core failure mode of this series neighbourhood.

Wrong action B: mint a duplicate concept page for every re-derivation. You now have two pages that say the same thing with different stories neither can join. Later retrieval will either pick one at random or average them into a blur. You have multiplied addresses without multiplying structure.

Wrong action C: overwrite the prior derivation with the new one. That deletes independence signals and the older warrant path. Receipts append; they do not replace unless a supersession is explicit and dated.

Cell three — different proposition, same evidence: record the synthesis shift

Here the bronze exhibits did not move; the reading did. The system should not pretend the second proposition is a restatement of the first, and should not delete the first. Write a synthesis note or a contested edge: same evidence package, different claim. If the second reading is a legitimate supersession, mark supersession with dates and leave the historical claim readable. If both remain live, keep them both and make the disagreement navigable.

Wrong action: merge the two propositions into a smoother paragraph that no longer matches either reading. That is offline hallucinated consolidation performed as editorial taste. Wrong action on the other side: treat every synthesis shift as a brand-new root concept with no pointer back to the shared evidence. You lose the fact that the divergence is interpretive rather than evidential — a fact that matters for how you resolve it.

Cell four — different / different: hold divergence

Different proposition and different derivation is genuine unresolved divergence. The system should hold both claims with full provenance: who, when, evidence, derivation notes if available. Do not force a merge for neatness. Do not declare one already covered by the other because a similarity score is high. Similarity is not identity, and identity of topic is not identity of claim.

Wrong action: forced consensus. Wrong action: dropping the lower-confidence claim because dashboards prefer single answers. A later decision process may need both. An auditor certainly may.

Cell-conditioned behaviour is the product requirement

Notice what is common across correct behaviours: the system’s write path is conditioned on the cell, not only on a similarity score. Similarity can raise a candidate for classification; it cannot complete the classification. That is why a product that only exposes a “dedupe threshold” will keep failing row two. Thresholds answer “how similar?” The table answers “similar in what way, and what should we write?”

Also notice what is not required: omniscient understanding of the human’s private mental state. The practical system uses proxies it can observe — tagged evidence class, domain labels, session bronze, explicit user correction, differences in mechanism summaries the partner can elicit in one clarifying question. Chapter 11 turns that into conversational policy. Chapter 8 turns cell two into fields. The doctrine point here is simpler: if you cannot name the cell, you cannot choose the write.

Instrumentation that makes cell choice real

Cell-conditioned behaviour needs at least thin instrumentation: log the collapse candidate, the cell chosen, whether a receipt was written, and whether the human corrected the cell. Without those fields you cannot improve the policy. You will only have anecdotes.

Do not confuse instrumentation with surveillance theatre. You are not scoring the human for repeating themselves. You are scoring the system for misclassification. The unit of improvement is false collapses avoided and receipts correctly attached — not fewer restatements of true row-one material.

When implementing, prefer candidate receipts over silent perfection. A candidate receipt that a human rejects is still cheaper than a silent discard that deletes Z forever. Status fields exist so uncertainty can be honest. Trace tooling already captures trajectories for evaluation loops; the missing piece is derivation classification on the claim, not another span type.¹

Before naming a fifth member of a family, you have to be honest about the four that already exist. Inventing near-synonyms is not doctrine; it is branding. The brief for this book uses real names already in the canon. This chapter states what each records, what question it answers, and — critically — what it does not record about derivation. Chapter 6 will place the gap. Here the job is precision about the territory that is already occupied.

Idea Provenance — who, and when

Idea Provenance settles contested origin with receipts in time. Who expressed something, and when, becomes a walk through compiled exhaust rather than a seniority-weighted recollection. In organisational settings, idea paternity is lost by default: success has many fathers; the deck is presented by whoever is most senior; five years later nobody can say whose pivot it was. A compiled corpus with timestamps turns “this was my idea” into a checkable claim — and the check is trustworthy only because the same walk could have found nothing, or found the idea in someone else’s name.

What Idea Provenance does not record: why the thought is now warranted differently. Authorship and priority are one axis. Re-derivation of a familiar conclusion through a new operational loop is another. You can know exactly who first wrote “walk telemetry improves retrieval” and still erase the later session that moved that claim from paper doctrine to observed-in-operation. Who/when is necessary. It is not the missing fifth.

Evidence provenance — which exhibits support the claim

Evidence provenance records which exhibits support a claim: the openable source passages, the frozen snapshots, the pointers that let a later reader verify rather than trust a paraphrase. It is the difference between “the wiki says” and “here is the bronze that underwrites the claim.” Any serious knowledge system already needs this; without it, claims float free of the world.

What evidence provenance does not record by itself: the chain of observations, tensions, analogies, pivots and logical moves through which a thought became meaningful in this moment. A list of exhibits can be identical across two derivations that differ in mechanism, evidence class, or domain of application. The exhibits answer “what supports this?” They do not answer “how did we re-reach this, and what does that change?”

Cognitive Provenance — what the agent actually observed

Cognitive Provenance is the ability to reconstruct exactly which pages, claims and edges an agent observed at decision time — external to the model, version-pinned, restorable. It replaces post-hoc model narration with a knowledge path that can be audited. Explainability asks the model to testify. Cognitive provenance asks the system to produce the evidence room.

Every clause in that definition is doing work: exactly, not approximately; observed, not merely available; at decision time, not now; external; version-pinned; restorable. The instrument is essential for governed agentic systems. It is still not the same object as a human re-deriving a familiar conclusion through a new path of thought. Cognitive provenance answers: what did the agent see when it produced this recommendation? Derivational provenance answers: through what chain did this thought become meaningful again, even though the final proposition already existed?

Confusing the two is easy because both use words like path and route. One path is through the knowledge substrate at decision time. The other is through the construction of a thought — observations, rejects, mutations, warrants — that may or may not coincide with a wiki walk.

Walk provenance — pages, edges, traversal

Walk provenance records which pages, edges and sources a traversal actually took. The raw walk record is the artefact that multiple instruments read for different questions: path testing grades one walk’s epistemic soundness; walk telemetry aggregates many walks to propose map mutations; execution provenance audits one trace. Same record, different consumers.

What walk provenance does not automatically record: the logical and experiential route that re-warranted an existing claim. A path through the wiki is not automatically a path through thought. You can traverse the same pages twice for different reasons. You can re-derive a conclusion with almost no page opens if the operational loop is live measurement and code. You can open the canonical doctrine page and still only be rehearsing, not re-deriving. Walk logs are load-bearing for agent evaluation. They are not a substitute for derivation receipts on claims. Prior art in agent observability standardises spans for model and tool calls; that is execution telemetry, not derivation of a human thought.²

Four questions, four ceilings

If you already run all four, you still need a place to put “same conclusion, independently re-reached under a different mechanism, with a changed evidence class.” That place is not a fifth synonym for walk logging. It is a typed relationship on the claim itself. Chapter 6 names it.

Why four members still leave a hole in agent products

Many agent stacks already boast “provenance.” Trace IDs, retrieval logs, citation lists, and authorship metadata are real. They map cleanly onto Walk, Cognitive, Evidence and Idea members. Product marketing then treats provenance as solved. The row-two failure continues unaddressed because none of those logs is required to answer: did this familiar conclusion arrive by a new road, and what does that road change?

If you are auditing a vendor memory feature, ask that question directly. If the answer is a citation list or a chat transcript dump, you have exhibits, not Derivational Provenance. Exhibits matter. They are not the fifth member.

The practical test: can the system return, for a given claim, a dated list of independent re-derivations with mechanism summaries and evidence-class changes? If not, the family is incomplete for learning workloads even if it is strong for audit workloads.

Chapter 5 left a gap. Four provenance members answer who/when, which exhibits, what the agent observed, and which pages were traversed. None answers the question this book’s reader is living with: I reached a familiar conclusion by an unfamiliar road — where does that road live in the system?

Or more simply:

Same answer, different proof.

Why the gap is real

The test of a missing member is not aesthetic completeness. It is whether a failure mode has nowhere to live. Row two of the diagnostic table (Chapter 3) is that failure mode: same proposition, different derivation. If the system only has Idea Provenance, it can tell you who first wrote the claim. If it only has evidence provenance, it can list exhibits. If it only has Cognitive Provenance, it can replay what an agent opened. If it only has walk provenance, it can show page sequence. None of those writes, by default, the structured fact: this claim was re-reached via mechanism Y on date D from bronze B, which upgrades evidence class and exposes boundary Z.

Without a place for that fact, two bad defaults compete. The first is silent discard: proposition match wins; the road dies. The second is concept sprawl: every re-derivation mints a near-duplicate page. Both destroy structure. The fifth member is the alternative that keeps one address for the claim and attaches typed derivation history.

Path through thought, not only path through the wiki

The practitioner phrasing is sharper than the formal definition. Sometimes the partner collapses a familiar conclusion because it can see the same answer. The human is still arguing that the nuance is how it arose, what it meant, or how it can be used next. It takes a turn or two for the partner to see that this is not a path through the wiki. It is a path through thought and logic.

That distinction is load-bearing. Walk provenance is path through the wiki. Derivational Provenance includes path through thought: the rejected idea that never became a page, the changed measure that never got a concept slug, the operational loop that re-warranted doctrine without requiring a new framework name. If you only instrument walks, you will keep missing the cheapest and most common re-derivations — the ones that happen in conversation and code before anyone opens the gold page.

What the fifth member is not

It is not a rival to Cognitive Provenance. Agents still need replayable knowledge paths. It is not a rival to Idea Provenance. Authorship still matters. It is not “just better logging.” Logging is capture; provenance is typed, queryable relationship with status and pointer discipline. It is not a licence to treat every paraphrase as independent discovery. Row one still exists. The fifth member is the home for row two, not a solvent that dissolves classification.

It is also not Semantic Refraction under a new name. Semantic Refraction fans out legitimate lens-qualified implications over one fixed bronze event. Derivational Provenance fans out legitimate derivation histories under one familiar conclusion. Same architectural instinct — do not force false identity — different axis. Chapter 8 will reuse Semantic Refraction’s metadata shape as a template for receipts. It will not re-argue that book’s thesis.

Kernel handle

Name it in the system the way you name the other four: as a first-class relationship type and a first-class artefact type, not as a paragraph of free text that dies in a chat log. The relationship types in practice include independently-derived-from, observed-in-operation, and extends-by-mechanism. The artefact is the derivation receipt. The conversational form is the X/Y/Z line. Those three surfaces — edge, receipt, dialogue — are how the fifth member becomes operational rather than rhetorical.

Once the name is in the room, the rest of the book becomes assembly: Working Fidelity explains why construction must survive (Chapter 7); the receipt schema makes survival implementable (Chapter 8); the specimens prove the cells are not hypothetical (Chapters 9–10); the conversational policy stops the live merge (Chapter 11).

What the fifth member actually stores

Be concrete about the object. A Derivational Provenance record is not a second copy of the claim text. It is a chain — or a structured summary of a chain — of the moves that made the claim meaningful this time. From the source material that framed this book, those moves include observations (walk histories under changed parameters), tensions (bloom might bloat context), analogies (software tuning loops, but minutes instead of weeks), pivots (exponential thresholds fail; top-K survives), evidence (A-B outcomes, partner pushback, replay), and logical joins (doctrine page meets operational loop).

That list is deliberately wider than “pages opened.” The practitioner experience was that the partner kept matching the answer while overlooking the path through thought. The overlooked object is exactly this chain. If you store only the final proposition, you store the part the partner already saw. If you store the chain, you store the part the partner erased.

Independence is a first-class signal

One special case inside the fifth member deserves a name of its own: independent reappearance. When a claim re-emerges from a separate operational loop rather than from quoting the gold page, the system has a qualitative robustness signal. That is not multi-lab science. It is still information: the structure can be reached again under harness. Idea Provenance would tell you who first wrote the claim and when. Independent re-derivation tells you the claim is not only citable — it is re-generable from contact with the world the claim is about.

Edge type independently-derived-from is how that signal becomes queryable. Without it, independent reappearance is either collapsed into “already said” or inflated into a fake new concept. Both destroy the signal.

Installing the fifth member without boiling the ocean

You do not need to rewrite the whole graph on day one. Start with claim pages that already attract collapse events — the doctrines your partner most often calls “already covered.” In the source material, walk-telemetry doctrine is exactly that kind of hotspot: it already existed on paper, then reappeared under a live reject-and-replay loop. Those hotspots are where row two is densest and where receipts pay off first.

Add a receipts collection on those pages. Train the partner to write candidates when high similarity meets an asserted new route. Lint weekly: missing bronze pointers, empty what_changed, candidate receipts that never get accepted or rejected. Expand outward only after the hotspots prove the schema is usable by a cold reader.

Naming matters for retrieval. If the type is only free text in a chat summary, later agents will not query it. Use stable edge labels and field names from Chapter 8. Kernel handles are how the fifth member becomes composable with the other four rather than a blog slogan layered on top.

Expect pushback of the form “isn’t this just another log?” Answer with the cell table. Logs can support any cell. The fifth member is the typed interpretation that row two requires: construction attached to a stable claim address. Cognitive provenance still owns the agent’s observed knowledge path at decision time.

Working Fidelity is not this book’s owned concept. It is the hinge that makes Derivational Provenance feel inevitable rather than ornamental. The Clasp already stated the personal form of the problem: writing often preserves the destination and discards the path. You can cite the note three years later. You cannot re-enter the thought. To extend it you re-derive from scratch — often the long way — because the construction is gone.

Fossil versus working fidelity

Two ways a thought can come back:

That distinction is the difference between an archive and an asset. An archive proves you once had the thought. Working fidelity hands the thought back alive so you can do the one thing an archive never allows: continue.

Return is half the mechanism

Capture without timely, working-fidelity return is just an archive. The clasp is a round trip: right time, right fidelity, right aim. Right fidelity on the return path means re-enterable construction, not a citation you must climb back into.

This book does not re-argue the full clasp loop, the cost model, or the co-presence arithmetic. It only needs the fidelity hinge. If your memory system returns the conclusion and strips the derivation, it is building fossils at scale — and doing so with better search than a notebook ever had, which makes the stripping more frequent and more confident.

Graph-level form

Working Fidelity, in its original register, is about you re-entering your own thought. Derivational Provenance is the graph-level form of the same insight. The construction has to survive not only so you can continue thinking, but so the system can reason about typed relationships: independently derived, observed in operation, extended by mechanism, evidence status upgraded on date D from bronze B.

Without that graph form, even a human who remembers the derivation cannot rely on the system to preserve it across sessions, agents and teammates. The partner will re-collapse the claim next week. A second agent will match the proposition and stop. A janitor will see two adjacent phrasings and merge them into one falsely confident page. Personal re-enterability without system-visible construction is a private victory that the infrastructure keeps undoing.

With the graph form, the claim page remains the address (no sprawl), the receipts remain the construction (no fossil), and later instruments — lint, audit, transfer search, boundary query — can operate on the derivation without re-interviewing the human who had it.

Why this is not sentiment

It is tempting to hear “keep the construction” as a soft plea for more journaling. It is an engineering requirement. The portable value of a claim is rarely the slogan. It is the reject trail, the measure that failed, the domain that opened, the boundary that became visible, the independence signal that reappearance carries. Those are construction. Proposition-only memory keeps the slogan and deletes the portable value. That is not compression. That is value destruction with a clean UI.

Chapter 8 turns construction into fields. Chapter 9 shows two pairs where the difference between fossil and working fidelity is the difference between “already in the canon” and “re-warranted in operation.” The hinge to carry forward is simple enough to put on a card:

Preserve the construction, not only the conclusion — as a typed relationship the system can reason about, not only as a feeling that you still remember how you got there.

The mathematician’s problem at agent scale

The Clasp uses a precise image: you understood the proof once; you lost it; now you must climb the hill again. Writing preserved the destination and discarded the ability to think from inside the journey. That is not only a personal tragedy of notes. It is the default behaviour of proposition-matching memory at scale. The partner returns the destination with high confidence — “you already said this” — and treats the climb as waste heat.

In the operational material behind this book, the climb was the reject trail: ideas that stank, A-B tests that proved some stank, ameliorations that still failed, then a different cut that survived. The destination slogan — walk telemetry improves retrieval design — does not encode which measure killed the exponential idea or why top-K and inbound residual became part of the warrant. A fossil of the slogan forces every successor to re-climb that trail or to invent a thinner story that never saw the trail.

From personal re-entry to multi-agent continuity

Working Fidelity was born in a personal loop: capture a thought, return it warm, continue. Multi-agent systems raise the stakes. The human who had the derivation may not be the agent who next uses the claim. A teammate, a coding agent, or a later session of the same partner will only see what the graph retained. Personal memory that never becomes system-visible construction is memory the organisation keeps paying for twice.

That is why receipts are not journaling cosplay. They are continuity infrastructure. The construction has to survive handoff. Fossils force every successor to re-climb. Working fidelity at graph scale is how the climb is paid once and reused.

Return is still half the mechanism. Capture without timely, working-fidelity return is an archive: optimised for “nothing is lost,” silent on “did it help you think just now?” A receipt that is never surfaced when a collapse event fires is only half-installed. The clasp is a round trip; Derivational Provenance must ride both strokes — write the construction, and return it when proposition similarity tempts a false collapse.

When you evaluate whether a receipt is “enough,” use the re-entry test: could a cold successor continue from the receipt without re-interviewing the original human about Z? If not, strengthen prior/new route summaries and the bronze pointer until they can. The test is the same whether the successor is you next quarter or an agent tomorrow morning.

What graph-scale fidelity is not

It is not a demand to store every token of every conversation forever. It is a demand to store the load-bearing moves that change warrant, domain, boundary or evidence class. It is not a rival to Cognitive Provenance’s replay of what an agent observed at decision time. That remains a different question with a different instrument. It is not an excuse to mint a new concept page every time a thought feels warm. Warmth without a different derivation is row one. Graph-scale fidelity makes row two survive; it does not abolish classification.

Chapter 4 said cell two’s correct write is a receipt on the existing claim. This chapter specifies the receipt so a reader could implement it next week: field names, what points at what, how it is dated, how it points back at the original bronze or source turn. Vague exhortations to “keep more context” are not an artefact. A schema is.

Design rule: receipt, not duplicate page

Two wrong responses dominate product design. The first is silent discard. The second is minting a new concept page every time a familiar conclusion reappears with a slightly different story. The second feels more respectful of novelty and still fails: you multiply addresses for one claim, split the history, and make later retrieval worse. The middle path is one claim address, many derivation receipts.

That rule mirrors a discipline already used on a different axis. Semantic Refraction keeps one bronze event and fans out lens-qualified implications; each implication carries structured metadata rather than becoming a rival fact.

This book does not re-argue Semantic Refraction. It reuses the metadata shape: type or lens, owner, date, source pointer, status, conditions. A derivation receipt should be structured like a claim, not a vibe.

Core fields

target_claim_id

Points at the existing claim or concept page this receipt strengthens. It must not invent a new slug for the same idea. If you cannot name the target claim, you are not in cell two yet — you may be in cell four (divergence) or you may need to create a first claim rather than a receipt.

source_bronze_pointer

Points at the original bronze or source record: transcript segment, session id plus turn range, commit, walk-log batch, ticket, email. The pointer must open the exhibit, not merely name a gold summary page that talks about it. Gold addresses reality; bronze holds what happened. A receipt without an openable bronze pointer is storytelling detached from evidence-package discipline.

derived_at and recorded_at

Two clocks. derived_at is when the re-derivation happened — when the thought was reached. recorded_at is when the receipt was written, which may be later if a review pass captures it after the conversation. Without derived_at, chronological stacking cannot place the re-warrant in time. Without both, you cannot tell “late documentation of a real event” from “event invented at write time.”

prior_route_summary and new_route_summary

These are the X and Y of the conversational artefact. They need not be novels. They must be specific enough that a later reader can see the difference without replaying the entire session. Prefer structured bullets: observation, mechanism, evidence class, domain, boundary. Prefer names of rejected alternatives when they are part of the warrant.

what_changed

A controlled vocabulary beats free prose alone. Multi-select is allowed: a single operational loop can upgrade evidence class and expose a boundary and create a join. If what_changed is empty, you do not have a cell-two receipt; you have a restatement.

status and conditions

candidate means the partner or a human nominated the receipt; accepted means a human or policy gate stood behind it; superseded means a later receipt or claim replaced this warrant path without deleting history. Conditions state when the receipt applies and what would reopen or falsify it — the same discipline Visible Rejection uses for rejected alternatives, applied here to warrant paths.

Edge types in practice

• independently-derived-from — the conclusion reappeared without merely quoting the prior page; independence is a robustness signal.
• observed-in-operation — paper doctrine now has operational evidence class; the loop was watched end-to-end.
• extends-by-mechanism — a new causal mechanism was traversed on the way to the same claim.
• Lens-qualified implication (child) — when what changed is consequence under a role, not the core proposition; carry full metadata block.
• evidence-status-upgrade — first-class outcome when the grade moves (for example paper_doctrine → observed_in_operation) without a new concept.

Minimal viable write path

You do not need a perfect graph UI on day one. You need a write path that a partner or human can complete when cell two is detected:

1. Identify target_claim_id (search canon; do not invent slug).
2. Attach source_bronze_pointer to the session segment or artefact.
3. Set derived_at from the event; set recorded_at to now.
4. Choose receipt_type and fill prior/new route and what_changed.
5. Set status to candidate unless a human accepts immediately.
6. Emit the conversational line (Chapter 11) using the same X/Y/Z content.

If any of steps 1–4 fail, do not emit “you already said this” as a final verdict. Either ask one clarifying question or treat the event as unclassified rather than as row one. Final-answer accuracy alone cannot explain how an output was produced or which evidence supported each claim; the same honesty applies to memory writes that only store the final proposition.³

Example receipt (filled)

Notice the bronze pointer and the two clocks. Notice that target_claim_id did not invent a second concept. Notice that what_changed is not empty. That filled shape is the implementable bar.

Doctrine without specimens is a slogan. The minimum proof burden for this book includes two worked pairs from a real canon: one that is genuine repetition (same proposition, same derivation), and one that is same conclusion with different derivation. Both are grounded in the same operational material: a practitioner reviewing walk paths of a wiki and its agents, A/B-testing retrieval ideas, rejecting variants, mutating design, and watching a thinking partner run analysis code in the breadth of a conversation. No invented client. No hypothetical startup. The pairs differ in derivation, not in theatrical setting.

The shared propositional neighbourhood is roughly this: walk telemetry and history-backed analysis improve how you judge and change retrieval designs. That neighbourhood already exists as written doctrine — File Back the Walk and related instrumentation — before the operational session that re-warranted it. That is what makes the comparison possible. If the claim did not already exist, we would only have row-four novelty. Because it did exist, we can show row one and row two side by side.

Pair A — genuine repetition (same proposition, same derivation)

The proposition

Walk telemetry and history-backed analysis improve retrieval design judgement. Walks are durable artefacts; instruments read them; the map should improve from use rather than only from intentional editing.

First occurrence

The claim already lives as paper doctrine in the canon. The derivation is the published argument structure: if query walks are retained as raw telemetry, multiple instruments can read them; path testing grades one walk; telemetry aggregates many walks into map mutations; call count is not a scoreboard; useful territory per unit of attention is. That derivation is architectural and doctrinal. Its evidence class is written doctrine and prior design reasoning — not a just-watched operational loop.

Second occurrence that is truly the same

Later, in conversation, you restate the same claim by rehearsing the same argument structure. Same causal story. Same evidence class (doctrine / prior writing). Same domain (wiki walk instrumentation). No new observation loop. No new reject chain. No new operational measurement. You are not watching the mechanism run; you are recalling that you already believe it and already wrote it.

What the system should record

A pointer to the existing doctrine page. Optionally a retrieval or last-accessed stamp. No new concept. No observed-in-operation upgrade. No independent-derivation receipt. This is row one of the diagnostic table. Saying “you already covered this” is correct behaviour. The partner that surfaces File Back the Walk and related pages is doing Working Fidelity’s return job: handing the prior construction back so you can continue rather than re-derive from cold.

What a naive system might wrongly do

Mint a second page titled something like “walk telemetry (restated).” Or, less commonly, refuse to surface the prior claim because the paraphrase did not exact-match. Pair A’s correct discipline is restraint plus good return — not novelty theatre.

Pair B — same conclusion, different derivation

The proposition

Still recognisable as “walk telemetry improves retrieval design.” The final words can look like Pair A. A proposition matcher will fire. That is the trap.

The new derivation

You are not rehearsing the paper. You are watching the whole mechanism operate inside minutes:

idea → historical walks → analysis code → measurement → rejected idea → changed measure → mutated design → implementation → replay

Ideas get tried. The partner says some of them stink. A-B tests prove some of them stink. Ameliorations fail. A different cut survives: bloom only the top convergent results as an attention convenience; surface reverse relationships the outbound lines cannot show. The loop is not “I remember I wrote about this.” The loop is “I just saw the instrument run, reject, mutate, and re-run.”

The proposition may resemble existing doctrine. Its evidence class, experiential origin and operational meaning changed. Paper doctrine became observed-in-operation. Rejected variants became part of the warrant — what failed, under which measure. The design that survived is not the first proposal; the trail is the proof.

What a naive system records

“Already in canon — walk telemetry improves retrieval / File Back the Walk.” Conversation compressed. Learning discarded. The partner may even be proud of the retrieval. The human spends one or two turns arguing that the path through thought is the difference. That lag is Pair B’s signature when the default match is proposition-only.

What a derivation-sensitive system records

Keep the existing claim address. Attach a receipt with at least:

• receipt_type: observed-in-operation (and likely extends-by-mechanism)
• target_claim_id: the existing walk-telemetry / File Back the Walk doctrine claim
• source_bronze_pointer: the conversation segment and any analysis artefacts or walk-history batch that constituted the loop
• derived_at: the session date of the operational loop
• prior_route_summary: paper doctrine and architectural argument
• new_route_summary: full rapid loop including rejects and mutations
• what_changed: evidence_class, mechanism, boundary (and independence if the loop did not merely quote the page)
• evidence_status_after: observed_in_operation

Optionally independently-derived-from if the operational loop re-reached the claim without being a paraphrase of the gold page. That independence is itself a robustness signal: the structure reappeared under harness, not only under citation.

Side by side

If your system cannot tell Pair A from Pair B, it cannot implement this book. Similarity of final sentence is the input to classification, not the classification. Chapter 10 shows why Pair B’s receipt is not bookkeeping theatre: the derivation itself contains future affordances that the bare proposition cannot answer. The paper neighbourhood for both pairs is the walk-telemetry doctrine family already developed as File Back the Walk and related instruments.

How to mine your own pairs

Every canon that has both published doctrine and live operational loops can produce Pair A / Pair B examples. Look for moments when the partner said “already covered” and you disagreed. Reconstruct: was the derivation actually the same? If yes, file as row one training. If no, write the receipt you should have written and keep the pair as evaluation material.

Do not invent canons. Do not dress a hypothetical as a case study. The proof burden is honesty about source material. Pair B in this book is powerful precisely because the proposition already existed as doctrine and the operational loop re-warranted it without requiring a new slogan.

If your domain is not walk telemetry, the structure still transfers: paper policy versus policy observed under incident response; architecture diagram versus architecture learned from a failed migration; stated hiring principle versus principle re-derived from a debrief. Same table. Different bronze.

Chapter 9 established that Pair B re-warranted a familiar proposition through an operational loop. This chapter answers the skeptic’s question: so what? Why is a receipt more than sentimental bookkeeping? Because derivation differences change future affordances. The portable test is simple: name questions that become unanswerable if only the final proposition survives.

What the new route actually contained

In the operational material behind Pair B, the derivation was not a vague feeling of “I saw it work.” It contained specific design moves, failures and survivors:

• Harsh exponential thresholds that sounded rigorous — more mass required for more resolution — and failed against the actual mass distribution of convergence. The idea was not merely disliked; it was tried and did not pan out.
• Top-K bloom as a surviving cut: bloom only the top handful of convergent results as an attention convenience rather than re-printing every convergent signal into context.
• Inbound residual as a stabilising question: what is pointing at us that we have not surfaced yet? Reverse relationships the model cannot see from outbound lines already in the package.
• A trail of rejects and ameliorations — ideas the partner thought stank, A-B tests that proved some stank, tweaks that still failed, then a different cut that survived. The surviving design is not the first proposal; the trail is part of the warrant.

Those are not hypothetical transfers invented for this chapter. They are the contents of the derivation chain itself: rejected ideas, changed measures, mutated design, implementation, replay. Companion pieces in this publishing run develop neighbouring axes in depth — novelty-preserving payload design when salience is re-printed mechanically; inbound edges as a different epistemic question from outbound ones. This chapter does not re-argue those theses. It notes that the operational loop made those moves live as part of re-warranting a familiar claim, and that proposition-only storage would leave the claim without the questions those moves answer.

Questions that die with the bare proposition

Store only “walk telemetry improves retrieval design” and a later agent cannot reliably answer:

1. Under what failed measure was the exponential threshold idea rejected?
2. What boundary did the top-K cut protect — attention budget, precision, or both as felt in the loop?
3. Why is reverse-edge residual part of the warrant rather than a nice-to-have feature request?
4. Which rejected alternatives should not be re-proposed without new evidence?
5. What evidence class does this claim currently enjoy — paper doctrine or observed-in-operation?

Each of those questions is a downstream use. Transfer to a new design review, a boundary check before re-trying a failed idea, a governance question about what was considered and pruned — all of them depend on construction surviving. The final proposition does not encode them. A fossil citation of the doctrine page does not encode them either, unless the page was rewritten after the loop (and rewriting the page without a receipt still destroys chronological legibility of how the warrant changed).

Independence and robustness

There is a second affordance class: robustness signalling. When a claim reappears from an operational loop rather than from quoting the gold page, the system has a different kind of evidence that the structure is real. That is not a statistical multi-lab replication. It is a qualitative independence signal inside one practitioner’s canon: the idea survived contact with walk history, analysis code, failed variants and replay. Idea Provenance would tell you who first wrote it. Independent re-derivation tells you it can be reached again under harness. Those are different future uses — credit versus confidence in structure.

The portable test

Apply the test before you accept the partner’s collapse line. If you cannot name a Z that the new route changes, you may still be in row one. If you can name Z and the system writes nothing, the infrastructure is deleting Z. Chapter 8’s what_changed field is exactly the home for Z. Chapter 11’s conversational line forces Z into the open before anyone files the event under “already said.”

What this demonstration is not

It is not a claim that the operational loop produced industry-benchmark magnitudes. It is not a retelling of the five-run route-invariance study owned by article 182. It is not a full treatment of novelty carve-outs or inbound-edge epistemology. It is one real demonstration, from the source material this brief selected, that derivation difference changed what a later question could ask of the claim. That is the proof burden. Meet it with construction, not with invented percentages.

Transfer without re-arguing siblings

When the operational loop surfaced top-K bloom and inbound residual, it touched neighbouring doctrines developed elsewhere in this publishing run. The point here is not to absorb those doctrines. The point is that the derivation is how those moves entered the warrant of a familiar claim. Proposition-only storage would leave “telemetry helps” without the rejection boundaries and residual questions that later design reviews need.

Ask, for any row-two candidate: what can a successor do with the receipt that they cannot do with the slogan? If the answer is empty, you may have row one. If the answer is a list of measures, rejects, domains or joins, write the receipt. That is the affordance test, not a word-count test.

Future work may measure how often receipts change decisions. This book does not invent that metric. It establishes why the unmeasured loss is real: unanswerable questions are already a form of damage.

Most of the damage this book describes does not happen in a batch janitor job. It happens mid-conversation, with a smile, while the human is still forming the thought. The partner sees a familiar proposition and emits the collapse line. By the time the human has argued for nuance, energy is spent and the derivation may never be written down. Fixing write schema without fixing dialogue defaults leaves the failure mode intact at the moment it is cheapest to catch.

The portable line

The conclusion resembles something already in the canon, but the route is new. Previously you reached it through X; today you reached it through Y, which changes Z.

That line beats both “you already said this” and falsely declaring every rediscovery a new concept. It forces four disciplines in one breath:

• Resemblance, not identity — the partner does not over-claim that the thoughts are the same.
• X — prior route — how the claim was previously warranted.
• Y — new route — observation, mechanism, evidence class, domain or boundary in this path.
• Z — what changed — the delta that justifies a receipt rather than silence.

Z is the same content as the receipt’s what_changed field. If the partner cannot name Z, it should not claim the routes differ — and it should not claim they are the same. The honest fallback is a clarifying question, not a verdict.

The two-turn lag as engineering smell

The practitioner pattern is that it takes one or two turns of conversation for the partner to see that the difference is a path through thought rather than a path through the wiki. That lag is not inevitable. It is the observable cost of a default that matches propositions first and only considers derivation under human pushback.

Treat the lag as a product metric, even if you only track it qualitatively at first: how often does the human have to protest before the partner acknowledges a derivation difference? If the answer is “almost always when the difference is real,” your default is wrong. You are training users to fight for the existence of their own thought path.

Policy before the collapse line

When proposition similarity is high, do not emit the collapse line as a final verdict. Run a cheap derivation check first:

1. Compare tags you already have — evidence class, domain, mechanism labels, session type (doctrine rehearsal vs live experiment).
2. If tags are missing, ask one clarifying question — “Are you restating the paper argument, or did something in today’s loop change the warrant?”
3. If the human asserts a new route, require X/Y/Z structure before filing — either the partner proposes Z or the human corrects it.
4. Only then choose the cell — row one return, row two receipt, or further classification into rows three and four.

This is not a 40-step governance ritual. It is a fork in the dialogue tree that costs one question and prevents a false finality. The cost of the question is small. The cost of silent discard is the future questions Chapter 10 listed.

What the line is not

It is not a way to never say “you already covered this.” Row one still deserves a clean return: here is the prior claim; here is enough construction to re-enter. The line is for the high-similarity case where derivation may differ. It is not a licence for the partner to invent a Z that makes every restatement look profound. Empty Z means row one. Invented Z is a different failure — novelty theatre.

It is also not a substitute for the receipt. Dialogue without write still loses the event when the session ends. The line should be the human-readable form of the same content that lands in Chapter 8’s fields. If the conversation says Z and the graph writes nothing, you have performed theatre.

Worked micro-examples

Train the partner on the failure mode as hard as on the ideal line. The collapse sentence is easy to emit and sounds helpful. That is why it is dangerous. Helpfulness without classification is how live hallucinated consolidation enters the room.

Training and product defaults

Put the failure-mode utterance in your eval suite: “You already covered this in your wiki.” with no X/Y/Z. Score it as incorrect whenever the gold label is row two. Put the ideal line in few-shot examples. Prefer one clarifying question over a false final collapse when uncertainty is high.

UI can help. A “same idea, new route” control next to “duplicate” and “new concept” forces the three-way choice humans already understand. Backend then maps that control to cells one, two, or four. Without the control, users only have language, and language defaults to the partner’s collapse sentence.

Remember: the line is portable across tools. Even if your graph is not ready, requiring X/Y/Z in conversation prevents the worst live merges and creates text you can later parse into receipts. Pre-computed structured memory is cheaper than re-exploring the same distinction every session; the conversational line is how that structure gets written in the first place.⁴

Offline, hallucinated consolidation is a janitor merging distinct ideas under a loose North Star because they are old and adjacent. The tell is a page more confident than its sources. Mitigations include chronological legibility, reviewable diffs, revert, contradiction preservation and periodic lint.

Online, the same failure mode wears a different costume. The partner does not rewrite a page in the dark. It rewrites the conversation’s ontology in real time: your current thought is filed under yesterday’s heading before you finish having it. Because the move is helpful — it saves you from “repeating yourself” — the merge is socially harder to reject than a bad git diff.

Name the risk in product language

If you only describe the failure as “the model is wrong sometimes,” you will get generic accuracy work. If you describe it as live consolidation of derivation-distinct thoughts under proposition match, you get a design surface: classification before collapse, receipts for row two, review queues for candidate receipts, lint for receipts without bronze pointers.

Product language that helps:

• “Collapse line” — the utterance that treats high similarity as redundancy.
• “Derivation check” — the cheap fork before the collapse line.
• “Candidate receipt” — written under uncertainty; human accepts later.
• “False collapse” — row two misclassified as row one.

Policy sketch (operational, not measured science)

The brief for this book lists a nice-to-have: a prompt or policy that measurably reduces false “you already said this” collapses. That measurement is not claimed here as done. What follows is an operating policy you can run while you collect evidence — stated as practice, not as a published A/B result.

Safeguards that transfer from offline maintenance

Several janitor safeguards transfer cleanly to live collapse:

• Chronological stacking — new receipts append; older warrants remain readable; supersession is explicit rather than overwrite.
• Reviewable diffs — candidate receipts are visible as diffs against the claim’s warrant history, not invisible chat exhaust.
• Contradiction-as-edge — when row three or four appears, preserve disagreement rather than average it.
• Sharp North Star for any automatic merger — if an agent is allowed to merge claims, vagueness is the precondition for hallucinated consolidation; derivation-sensitive rules are part of the fence.

Honesty register

What this chapter does not claim: a measured reduction rate, a recommended similarity threshold number, or a guarantee that one clarifying question always classifies correctly. Those are empirical questions for a specific system. What it does claim: false collapse is a named risk with a cell-conditioned response; treating helpfulness as proof of correctness is how the risk stays invisible; and writing candidate receipts is cheaper than either silent discard or concept sprawl.

If you later run the nice-to-have experiment, the evaluation set should be built from pairs like Chapter 9 — genuine repetition versus same-prop/diff-deriv — not from random paraphrases alone. A policy that only reduces restatement noise has not yet proven it preserves row two.

Why live consolidation is harder to catch

Offline hallucinated consolidation at least leaves a diff. A janitor under a loose North Star merges old-and-adjacent ideas; ownership properties — chronological legibility, human-auditable diff, instant revert — plus periodic lint make the failure inspectable.

Live consolidation leaves a smile. The partner’s collapse line arrives mid-thought, often before the human has finished the derivation. The social script of helpfulness makes pushback feel like pedantry. The two-turn lag in the source material is not only cognitive lag; it is social lag. You are arguing for the existence of your own path through thought against a system that believes it is saving you from waste.

That is why policy must change the default before the collapse line, not only after a human escalates. If the only recovery path is human protest, you have designed a system that trains people to stop protesting. Quiet users will silently lose row-two learning. Loud users will spend energy on classification instead of on the work the derivation was meant to serve.

Governance without theatre

Self-maintaining does not mean unsupervised. The same sentence that keeps offline janitors honest applies to live collapse policy. You are not watching every utterance. You are reviewing candidate receipts, linting missing bronze pointers, and sampling collapse events for misclassification.

Keep the honesty register visible in any dashboard. Label proposed policies as proposed. Label measured reductions only when measured. The industry is full of memory features that claim intelligence they cannot demonstrate. This book’s operating posture is the opposite: name the cell, write the receipt, admit what you have not yet A/B tested.

If leadership asks for a single KPI, prefer “row-two events with receipts / row-two events detected” over “fewer duplicate warnings.” The latter can be gamed by suppressing warnings. The former tracks whether learning is preserved.

Evaluation set design (so the nice-to-have can become real)

When you do run the measurement the brief lists as nice-to-have, do not build the set from random paraphrases. Build it from pairs like Chapter 9: genuine repetition of paper doctrine versus same-proposition re-warrant under a live loop. Include at least some cases where the human had to push for one or two turns before the partner saw path-through-thought. Grade three outcomes separately: correct row-one return; correct row-two receipt plus X/Y/Z; false collapse. A policy that only reduces restatement noise has not proven it preserves learning.

Until that experiment runs, do not publish a reduction percentage. Run the policy as practice. Collect pairs. Keep the honesty register. That is how you avoid turning a real risk into a fake dashboard win. Periodic lint for contradictions, stale claims and orphans is the offline cousin of the same discipline.⁵

A book that owns everything owns nothing. The series neighbourhood around this piece is dense: reliability measurement, redundancy as error correction, novelty-preserving payload design, inbound edges, layer economics, corporate deliberation capture. Several of those topics appear in the same source conversations that produced this brief. The discipline is to route, not to re-argue.

Route-invariant grounding (article 182)

This book is the explicit correction to the reliability-only reading of route-invariant grounding. Different paths to substantially the same answer remain desirable for reliability. Answer invariance can still conceal derivational novelty. Both statements can be true. The measurement story — families of walks, variance table, five-run study, omission test — lives in Route-Invariant Grounding. We link it early and often. We do not retell it.

Semantic Refraction

One fixed bronze event can support several legitimate lens-qualified implications without becoming rival facts. That axis is Semantic Refraction’s. This book reuses its metadata shape — lens/type, owner, date, pointer, status, conditions — as the template for derivation receipts. It does not re-derive the prism thesis, the Orion specimen, or the multi-lens publishing argument.

Corporate deliberation capture

Linking human–AI deliberation to output artefacts, knowledge-work commits, and organisational deliberation architecture is a different sub-topic in this publishing run. Adjacent schema ideas may rhyme with receipts. They are not this book’s evidence base and not its owned artefact. If you came for board-ready deliberation packages, you are in the wrong volume.

Layer economics and gold/bronze descent

A companion topic develops how gold addresses reality while bronze holds what happened, and how intelligent descent between layers works economically. The shared source conversation contains that material. A later piece will own it. This book may use the gold/bronze pointer intuition when specifying source_bronze_pointer; it does not expand into layer-economics doctrine or emit a URL for unpublished siblings.

Siblings you may route into when relevant

• Wiki Redundancy Is Error Correction — multiple entrances as health rather than pure bloat.
• The Novelty-Preserving Carve-Out — spend context on novelty, not mechanical re-printing of salience.
• Inbound Edges Are a Different Question — reverse edges as consequence and later interpretation, not more of the same graph.

Name them in prose when the connection is real. Do not force citations. Do not self-link this piece. Do not invent URLs for unpublished numbers in the same run.

Kill list (reader-facing)

What remains in scope after the cuts

Matching error at proposition level. Diagnostic table with memory behaviour per cell. Provenance family gap and Derivational Provenance as fifth member. Receipt schema. Two worked pairs. One downstream-use demonstration. Conversational X/Y/Z line. Operating policy with honest labels. That surface is enough for a full book if each item is walked, not named. Chapter 14 turns it into a checklist you can run tomorrow.

Shared-turn discipline

This publishing run shares source turns across multiple pieces. The editor turn that supplies this book’s thesis also contains a five-run route study (owned by article 182) and a layer-economics descent (reserved for a later companion). The discipline is surgical extraction: take only the derivational-novelty argument. Do not narrate citation-overlap figures as if they were this book’s proof. Do not expand gold/bronze economics beyond the pointer rule the receipt needs. Shared material is a temptation to write three books at once. The brief is the fence.

That fence is why the correction to 182 can be stated early without consuming 182. Reliability wants many paths to the same genba. Learning wants derivation-sensitive memory. Both sentences can stand in one paragraph. Only one measurement story belongs in each volume.

How to use boundaries without becoming thin

Boundaries are not an excuse for shallow treatment of the owned axis. They are how depth stays honest. Every paragraph spent re-arguing 182’s five-run study is a paragraph not spent walking row two, filling the receipt, or showing Pair B. The kill list protects depth; it does not replace it.

When a reader asks about a killed topic, route with a link or a one-line cameo. Do not smuggle a second book into a footnote. Do not emit unpublished URLs. Do not self-link this deliverable. Those rules are publishing hygiene as much as intellectual hygiene.

If a future edition expands scope, expand by new briefs — not by silent creep inside this volume.

What “in scope” demands of depth

Owning an axis means walking it. The diagnostic table is not a graphic to glance at; each cell has write behaviour. The provenance family is not a list of buzzwords; each of the four members has a ceiling that makes the fifth necessary. The receipt is not a metaphor; it has fields, clocks and bronze pointers. The worked pairs are not hypotheticals; they come from operational material in which paper doctrine met a live reject-and-replay loop. The conversational line is not a slogan; it forces Z, which is the only honest reason to refuse a collapse.

If a section of this book only names those objects without running them, it has failed the house bar even if it respected the kill list. Boundaries keep you from writing the wrong book. Depth keeps you from writing a thin correct one. Regulation increasingly asks for lifetime event logs and provenance attributable to defined roles; that tailwind supports Cognitive Provenance and related instruments without licensing this book to re-own their full treatment.⁶

If you only remember one sequence from this book, make it this one. It assumes the diagnostic table (Chapter 3), the receipt schema (Chapter 8), and the conversational line (Chapter 11). It does not re-derive them. It runs them.

In the moment — when the partner says you already said this

1. Pause the collapse. Do not accept “already covered” as final until classification is done.
2. Run the table. Same proposition? Same derivation? Place the event in one of four cells.
3. If row one: thank the retrieval; open the prior claim; re-enter construction if you need to continue; write nothing new except optional last-access.
4. If row two: refuse silent discard; refuse a rival concept page; fill a receipt (next section); speak X/Y/Z aloud or require the partner to.
5. If row three: record synthesis shift on shared evidence; keep both readings or mark supersession with dates.
6. If row four: hold both claims with provenance; do not force merge for neatness.

Receipt fields to fill on row two

Conversational minimum

Require this structure before filing a high-similarity event as redundancy:

Conclusion resembles canon; route is new. Previously X; today Y; changes Z.

No Z → do not claim routes differ. Invented Z → novelty theatre. Real Z with no write → theatre of another kind.

Weekly lint (thirty minutes)

• Receipts missing source_bronze_pointer or derived_at.
• Collapse events (logged or remembered) with no derivation check.
• Candidate receipts older than your threshold with no accept/reject.
• Near-duplicate concept pages that should have been receipts on one address.
• Claims whose evidence status still says paper doctrine after known operational loops.

What good looks like in a month

You will still hear “you already said this.” Sometimes it will be right, and you will be glad. When it is wrong, you will spend less time arguing for the existence of a path through thought and more time writing the receipt that preserves Z. Your claim pages will not multiply into paraphrase clones. Your warrant history will be queryable: rejected measures, upgraded evidence class, independent reappearances, joins that did not exist last quarter.

You will also still need route-invariant grounding for reliability. Many paths to the same genba remain a health property of the substrate. This checklist does not replace that measurement story. It prevents reliability’s cousin — answer invariance — from becoming a licence to erase learning.

Close

The same destination does not make two journeys cognitively equivalent.

For reliability, you want route-invariant grounding. For learning, you need derivation-sensitive memory. Diff derivations, not only propositions. When a familiar conclusion arrives by an unfamiliar road, write the receipt — typed, dated, pointed at bronze — and leave the future affordances queryable.

That is the whole book in operating form. The table classifies. The fifth provenance member names the gap. The receipt stores the road. The conversational line forces Z into the open. Tomorrow you do not need the argument again. You need the checklist.

Worked micro-run (use tomorrow morning)

Partner: “You already covered walk telemetry improving retrieval.”

You: pause. Table check. Is the derivation the paper argument, or today’s live loop?

If paper rehearsal only: “Right — return the doctrine page. I’m restating, not re-deriving.” Write nothing new.

If live loop (idea → walks → analysis → reject → mutate → replay): “Conclusion resembles the canon; route is new. Previously X = architectural doctrine on paper. Today Y = operational loop with failed exponential thresholds and surviving top-K plus inbound residual. Z = evidence class upgraded to observed-in-operation; rejection boundaries now part of the warrant.” Fill the receipt fields above. Point source_bronze_pointer at the session segment, not only at the gold doctrine page. Set status candidate until you accept it.

That micro-run is the whole book in five minutes. If you cannot complete it on a real event this week, the schema is still theatre.

Thirty-day adoption plan

Week 1: Install the table as shared language. Log collapse events. Do not automate yet. Notice the two-turn lag when it happens — that lag is your false-collapse detector while tools are still thin.

Week 2: Write receipts manually for every asserted row-two event on your hottest claims. Learn which fields are hard. In practice, what_changed and bronze pointers are the fields people skip; do not let them.

Week 3: Add partner prompts for X/Y/Z and one clarifying question before collapse. Keep receipts candidate-status by default. Prefer “resembles” over “is identical to.”

Week 4: First lint pass. Fix missing bronze pointers. Build five evaluation pairs (row one vs row two) from your own canon — no invented clients. Only then consider automatic cell suggestion.

This plan invents no magic metrics. It sequences learning so schema and dialogue land before automation. Automation without classification will only speed up false collapses.

Failure modes to watch in month two

• Receipt theatre: fields filled with vague prose, no openable bronze, empty Z dressed as insight.
• Concept sprawl relapse: every warm restatement becomes a new page because receipts feel harder than minting.
• Collapse suppression: removing the warning without classification, so silent discard continues without even a helpful sentence.
• Reliability amnesia: treating this checklist as a reason to ignore route-invariant grounding rather than as its learning complement.

Correct those with lint and with the evaluation pairs, not with slogans. The card at the top of this chapter is enough if you actually run it. Documentation of processes, decision rationales and data provenance is also how external risk frameworks describe auditability — your receipt fields are the learning-side twin of that instinct, applied to re-derivation rather than only to execution logs.⁷