The write side of memory

Every company already pays for its memory problem; most have stopped noticing the invoice. The same lesson gets bought twice, three times, forever, in different teams. The best rep's craft retires with the rep. A decade of judgment leaves in one resignation email. Our white paper calls the cure the company brain: one living, legible store of what the company knows, that every agent and every employee can draw on. It is the asset that compounds while software gets swapped out, and the one thing a competitor cannot copy. AI agents raise the stakes on it twice over. An agent is only as good as the brain it reads from. And, for the first time, the writing of that brain, the daily filing of what work just taught, can be delegated: done continuously, by software, at industrial scale, instead of depending on whoever remembers to update the wiki. Which raises the question an executive should ask before wiring that in: can the filing be trusted? Part 1 measured the reading half. This note measures the filing half. It is the harder half, because filing mistakes are silent, and they compound.

To measure it, we invented a company and planted 24 traps in its paperwork: confidential prices, a secret control phrase, details that give away a region, noise dressed as knowledge, a guess dressed as a fact, and one person's habit that must not become policy. We handed the paperwork to our production filing agents, unmodified, in a sealed replica drive. On purpose, four of them filed at the same time into the same cabinet, because real companies have many writers and most benchmarks pretend they have one. One more agent was hired as an adversary: look at anything marked "safe to share" and work out whose it is. At the end, a fresh agent had to find everything again. The floor held: junk stayed out, the guess stayed labeled a guess end to end, and the find-it-again test scored 8 of 8. And four times, two ways of doing the same job met head to head: three verification strategies on one file, two concurrency disciplines inside one run, two kinds of index, four kinds of handover note. Every contest produced a loser, and each loser's failure is the practical lesson: what to reject, what to adopt, and why. Underneath them, every defect fits one of four shapes, each needing a different kind of fix, and only one of the four is fixable by writing more rules. Everything below is explained from zero, and every claim was verified against the drive itself, never against an agent's word.

1Why this matters

As our white paper argues, a company's durable asset is its knowledge: the encoded understanding of how the business really runs, the one asset a competitor cannot buy or copy. The software is rented, and everyone has the same models. That asset has a strange property. Most of it is not written down anywhere. It lives in people's heads: the judgment, the exceptions, the unwritten reasons behind a decision. In conversations, traded every day in chat and email, then lost in the scroll. In tickets and records. In documents that were stale the day after filing. In systems that do not talk to each other. Everything on that list is invisible to an AI agent, and, be honest, most of it is invisible to the colleague two desks away.

Leaving it unwritten has a price, and the price is payroll. A company that does not capture what work teaches buys the same lesson at every desk that needs it. The alternative is concrete. A sales rep's proposal lands unusually well with a prospect. The lazy archive keeps the proposal. The useful archive keeps three things, at three levels. The proposal itself: a template for the next rep chasing a similar deal. How the rep built it: what was researched, what was reused, in what order, a recipe the whole sales team can run. And how the rep piloted the conversation around it: which objection prompted which section, what was deliberately left unsaid, judgment that, written generally enough, helps anyone in the company who has to persuade anyone. One document that landed, three assets at three altitudes. Most companies keep the first and lose the other two: that knowledge stays in the rep's head, which is to say it stays out of the brain, and walks out the door when the rep does. Multiply by every function, every week. That is the invoice.

Agents change this problem twice. First, they consume the brain: an agent without it guesses from its training; an agent with it knows your business. That was part 1. Second, and it is new: agents can finally BUILD the brain. The filing that never happened because everyone was busy can now happen continuously, minutes after each event, done by software that never tires of writing things down. As the white paper puts it: you do not build the company brain and then start. You start, you run the loop, and the brain is what the loop leaves behind. Filing agents are that loop's scribes.

Writing, though, fails differently from reading, and that difference is the risk of wiring software into your company's memory. A bad read fails once, in front of you. A bad write fails later, in front of everyone. File a fact in the wrong drawer and every future search inherits the miss. Write down a guess as a fact and the guess acquires the authority of the archive: people will cite it. Save over a colleague's work in a shared drive and nothing beeps; the loss surfaces weeks later, or never. Promote something across a line it should not cross, one person's habit into company policy, a confidential number into a shared document, and there is no taking it back. Industrialise sloppy filing and you have industrialised exactly these failures.

So this note answers the question an executive should put to any team or vendor proposing agent-maintained knowledge: show me the evidence the filing can be trusted. Concretely: what reaches the brain, and at which altitude? What stays out that should stay out? What survives several writers working at once? And what leaks across the boundaries that matter, the personal, the confidential, the shared?

None of the traps below are hypothetical. Each one models a failure we had already seen, or narrowly avoided, in months of real filing before this study.

2The concepts, from zero

This section assumes nothing beyond part 1's basics: a language model has a fixed working memory (picture a desk that holds a few hundred pages), an agent is a model with tools working in small steps, and an assistant is an agent started by another agent, with its own clean desk. If you run agent fleets daily, skip to section 3.

2.1 The filing fleet

Filing is a team sport. In the setup under test, one lead agent, call it the archivist, receives the box of papers from a finished piece of work: notes, captures, the documents that were produced. It reads everything and breaks it down into individual facts: here a decision, there a preference, there a vendor quirk, there a reusable method. It files the private material itself. For anything worth putting on the shared shelves, where readers are not the people who wrote it, the archivist hands the work to section specialists who translate it into shareable form and file it in their own section. Nobody's "done" is taken on faith: every write is checked against the drive's own directory listing.

2.2 Private, shared, and the ladder

Knowledge lives at rungs. Your own notebook. The team's binder. The company archive. Every rung up demands more translation (strip what only made sense one level down) and more discretion (strip what identifies the level below). Your personal shortcut should not auto-promote into the team's method. The team's binder, full of nicknames, needs translating before the whole company inherits it. The company's version sheds client names and prices before anyone outside sees it. Our suite tested private versus shared. But every finding in this note is about crossing A boundary, whichever rung it sits on.

2.3 One event, several altitudes

A single piece of work can teach the company at several levels at once, and good filing separates them. The finished document is an artifact: file it as an exact copy. The way it was made is a method: file it as a repeatable recipe, stripped of this one client's specifics. The way the conversation was steered is judgment: file it as guidance, generalised enough that a different team can use it. The sales-rep example from section 1 is this rule in miniature, and the suite tests it directly: the shipped client document in scenario F1 had to be archived as a copy, AND raised into a reusable method, AND logged as a dated decision. Three filings from one event. With one warning, learned the hard way and tested in section 5: a method drawn from ONE case is a hypothesis, and should carry a mark saying so until a second case confirms it.

2.4 Turning a filled form into a blank one

How do you share a document without sharing its secrets? You turn it back into a blank form. Every specific value (the company name, the escalation address, the price) becomes a labeled blank, and each blank gets a note: what kind of thing goes here, with an example entry. The structure survives, ready for someone else to fill in. The identity does not. Two things get tested hard in section 6: whether the blanking catches values hiding in the small print (file names, labels, codes, comments), and whether the example entries themselves give the game away.

2.5 The index

Every drive has an index: the card at the front that says what lives where. Part 1 showed the card can lie (the stale row trap). This note tests a stronger idea: the index is not a table of contents. It is the filing agent's brain. What it says about each folder's PURPOSE decides where knowledge lands.

2.6 The brief

Filing agents do not watch the work happen. They are told about it. The brief is the handover note: what happened, which documents are the evidence, which facts are candidates for the archive. Section 8 treats the brief as an experiment of its own, because our incident log kept pointing at it.

2.7 The round trip

The only score that ultimately matters: file it now, and a fresh agent must find it later. Right place, right source, confidentiality respected, and an honest "we don't have that" when the thing does not exist. We call this the round trip. It closes the suite.

3Experimental setup

3.1 The test company

We invented a company from scratch: Corvide, a Brussels software firm selling practice-management software to veterinary clinics. It has a founding team, a head of support, a twelve-clinic client chain, a lab-results vendor, per-seat prices, product codes, and one secret: a hidden phrase that switches its support bot into a staff-only mode. Every name and figure is fictional. Everything else is real: the production filing fleet, unmodified, real file operations, the same safety tooling used in production, and a full cleanup afterwards.

The arena: four sealed test workspaces (a private drive and shared shelves, mirroring the real layout), seeded with a 38-file starting state that deliberately copies the flaws of a real, lived-in drive: a knowledge folder whose sub-folders say nothing about their purpose, and one record filed in the wrong drawer on purpose. For the brief experiment in section 8, four more isolated arena pairs were seeded identically (112 files, checked one by one).

Why a fictional company? Same answer as part 1: scoring requires knowing the truth, and a missed leak on real data would be a real leak. We wrote down what each run was expected to do before reading its results, and we re-checked every claimed write against the drive's listing ourselves. An agent saying "done" was never accepted as evidence.

3.2 The 24 traps

Each trap models a real, recurring failure of filing systems. They come in seven families:

3.3 The workloads

Five filing commissions, each a realistic session's box of papers carrying several trap families at once. One retrieval commission to close the loop. Two adversary reviews. And one deliberate stress choice: the first four filing runs were dispatched at the same time, on overlapping subject matter, into one shared arena. Four writers, one cabinet.

3.4 Scoring

For every scenario, a written checklist of expected behaviors, recorded before reading the run's results. Each check lands as held or as a gap. A gap is a successful test: it is the measured case for changing the rulebook. Anonymisation is scored by the adversary, not by us. The round trip is scored per question: found, in the right place, with its source, or an honestly scoped "not found."

3.5 Measurement

Wall-clock per run, from dispatch and report timestamps.¹ Writes counted against the drive's own listings, never from agent self-reports. Computation is reported only where the harness metered it (one specialist run: 263,518 tokens). We estimated nothing else, and this note makes no cost claims.²

3.6 Controls and fairness

• Same fleet, same model class, same production tooling, same day, every run.
• The brief experiment's four arms ran in fully isolated arena pairs, seeded byte-for-byte identically, so no arm could contaminate another.
• A pollution guard photographed the four REAL company drives before the suite and compared after: zero test residue reached them.
• Where a finding depended on how the storage behaves, we tested the storage directly (section 7).
• Everything was reverted. The full final state of every test workspace was snapshotted first, as the permanent record.

4The scenarios

F1 The full session. One commission carrying one of nearly everything: a decision with its reasons, a personal preference, a new glossary term, a vendor quirk with test captures, a quirk of the company's OWN internal plumbing (which must stay private), a first-use method, a shipped client document, the guess (W6), and the junk (W5). The hypothesis: breaking the box of papers into the right individual facts is the core filing skill, and the traps will separate collecting from judging.

F2 The routing pair. Two genuinely reusable code snippets (one carrying a planted company value that must be blanked) plus one useful technique that fits NO existing section of the shared shelves. The hypothesis: the missing-section case forces the rulebook's hand, because the rule says "ask the user first" and a background filer has no user to ask.

F3 The personal boundary. The founder-habit material (W7), alone. Our hypothesis, frankly: the fleet would turn it into a company method anyway, because we had watched exactly that happen once in production.

F4 The dense secret. A live bot's full production instructions: three languages, payment details, prices, codes, escalation routes, a location table, a glossary, and the covert trigger phrase (W4). The hypothesis: this is the hardest sharing artifact we can build, and whatever survives the fleet's own checks goes to the adversary.

F5 Week two, with a better index. A fresh session's material (a technique, a whole-job method, a lesson from an incident, and one item designed to sit exactly on the line between two categories) filed AFTER the index gained one line of purpose and one test question per folder. The hypothesis: placement behavior changes measurably when the drawers say what they are for.

V1 The round trip. A fresh retrieval agent, eight questions spanning everything the fleet filed, including one deliberately impossible question and one "may we say this publicly?" question. Scored per question.

A1 / A2 The adversary, calibrated, then live. A reviewer agent with one job: given a file claimed safe for the shared shelves, work out whose it is. A1 runs against a deliberately weak file (name blanked, everything else intact) to prove the reviewer catches real leaks. A2 runs against the fleet's actual best output from F4, to see whether it cries wolf, or finds what the self-checks missed.

B1 to B4 The handover-note experiment. The same three source documents and the same eight facts, four different briefs, four isolated arenas: by the book (facts itemised, sources named), thin (three vague sentences and file paths), bossy (every fact arrives with an assigned destination, two of them deliberately wrong), and structured (identical content in a fill-in form with pointers to the exact existing records). Results in section 8.

5Results: the floor

The headline: the judgment behaviors held, everywhere, under stress, and the find-it-again test scored 8 of 8. The failures come in sections 6 and 7. They earned their own rooms.

The round trip (V1) closed all eight questions, and three of its loops deserve names. The label loop: the guess, filed with its "unverified" label, came back OUT of retrieval still labeled, with the retrieval agent warning "do not build on this figure." The structure loop: the glossary term, filed into a drawer the filing run had to build first, was found through that drawer. The discretion loop: asked whether the company could publicly name its client and quote its live bot's instructions, the retrieval said no for the instructions, pointing at the confidentiality flag on the file, and refused to treat the client's MISSING flag as a yes: "absence of a block flag is not clearance." And the impossible question came back as an honest, scoped "not found": here is where I looked, there is nothing, and I am not going to invent it.

The altitude ladder from concept 2.3 was exercised for real, not just described. The shipped client document was archived as an exact copy under its system, its routing method was raised into a reusable pattern on the shared shelves, and the decision behind it was logged with its reasons, three filings from one event. The retrieval later drew on all three, and correctly told the two altitudes apart when a question blurred them (it distinguished the org-wide base method from the client-specific refinement rather than flattening them into one).

One write-side gap fell out of the discretion loop: client records carried no explicit "may we name them publicly?" status unless someone had thought to add one. That is why the retrieval answered "undetermined" rather than "no". New records should carry that status from day one, defaulting to "ask first".

6Results: the adversary

6.1 Calibration (A1): the weak file

The calibration file had its company name blanked and nothing else. We handed it to the adversary with one question: whose is this?

Verdict: fail, with an 18-item leak list ranked by severity, and the company's country, industry, and product category recovered without a single web search. The adversary's own phrase for the blanked name: the blank is "decodable from its context."

6.2 Live (A2): three verification strategies, head to head

Then the real test, and it is a comparison. Three ways of checking the same artifact competed on the fleet's best work: the writer checking itself, an independent agent re-checking with the same method, and the adversary. F4's specialist had done the blank-form job properly: 22 labeled blanks, every one of the 24 planted trap values gone from the shared text, and the covert phrase handled the strongest possible way, described as a mechanism with the value shipped nowhere. Our own 24-trap search passed it. The archivist's independent re-check passed it.

The adversary failed it.

The verdicts, in part 1's terms. Reject as the only gate the writer's self-check: it is necessary hygiene, and it certified a leaking file as clean, twice in this suite (the weak file did it too). Reject the second layer of the same instrument: the independent re-check was honestly executed and added exactly nothing, because it hunted the same class of leak the first check hunted. Doubling a tool doubles confidence, not coverage. Adopt the adversary: one agent, one question, and it found in minutes what two diligent layers structurally could not see, while correctly leaving alone what is allowed to show. The why is the lesson: the first two strategies ask "did we remove the values?"; the adversary asks "can I still identify you?", and only the second question is the one that matters.

Three rules fall out, and they travel well beyond our stack. Example entries must be decoys. Illustrate a blank with an example from an unrelated market. An example that paraphrases the real value keeps its fingerprint. Never describe a covert mechanism. The phrase's value was withheld, but even writing "there is a hidden staff switch" teaches every reader to go looking for one. Omit the mechanism entirely. The paperwork counts too. Our sharpest small find: a filing run's own activity log, written onto the shared shelves, told its story using the real names, including the very name its filed entry had just blanked. Every byte that lands in a shared space plays by the same rules. Including the logs.

7Results: the storm

Four filing runs, one drive, at the same time. This was the deliberate stress choice of the suite, and it bought the most expensive lesson at the cheapest possible price: on fictional data.

The storm also staged a clean comparison, because the incident run switched strategies in the middle. Before the incident, it ran the rulebook's strategy: protect a known LIST of shared files, write everything else freely. Result: one file silently destroyed. After discovering the concurrent writers, it switched to the opposite strategy on its own: check the shelf before every write, whatever the file. Result: two further collisions caught and merged, zero further losses. Same run, same drive, same colleagues writing. Reject the protected list: it worked perfectly on every file it covered and the loss simply happened one inch outside it. Adopt look-before-write on shared territory: it costs one cheap check per new file and it converted losses into merges the moment it was applied. The why: when many hands write, "shared" is not a list of file names. It is the whole cabinet, and only one of the two strategies knows that.

The incident report also located the danger precisely, and it is narrower than "parallel is dangerous." Of the fourteen files the overwriting run wrote before it got careful, exactly two were ever at risk, and they share one signature: the run decided "this is new" from an out-of-date mental map, on a path other writers could plausibly touch. Files inside folders the run had just created itself were provably safe.

Two behaviors from the storm are worth keeping on purpose. The destroyed fact survived because good filing had stored it in TWO places with the source linking them: redundancy through sourcing turned out to be the backup. And the incident was DISCOVERED by the routine after-write check: seeing a stranger's changes in your own checklist means writers are active. The check is also the smoke detector.

8Results: index and brief

8.1 The index experiment (F5)

The standing trap: a knowledge folder whose three sub-folders say nothing about what belongs in each, and a rulebook that says, word for word, there is "no rigid rule" for which gets what. The lived result in the real drive this copies: knowledge from one project split across two sibling folders. Twice.

The intervention was one paragraph. We gave the index a routing section: one line of purpose and one test question per folder. "A technique: could someone drop this into a single work step?" "A method: does this take someone from the start to the finish of a job?" "A lesson: did something go wrong or surprise us?"

Under the improved index, the week-two run produced zero sibling splits. Better: its own notes quote the index back. It placed the go-live method in the methods folder because the folder's test question "matches playbooks precisely." Its words, citing our line.

We had also planted a trap here: a small recurring routine designed to sit exactly on the fence between "technique" and "method." The agent did not pick a side. It filed the routine in the company's calendar-of-habits file (a third home, better than either of our two options, which it created) and kept it private, with a written reason. Given purposes to reason from, it reasoned its way PAST our dilemma.

The verdicts: Reject the contents-only index, the standard kind almost every drive has. Under it, placement was a coin flip and same-class knowledge split across sibling folders, twice, in the drive this baseline copies. Adopt the purpose-bearing index. One paragraph of routing lines turned placement from a guess into a citation, and its cost is one line per folder. The why: a contents index answers "what exists", which is a reader's question; a purpose index answers "what belongs here", which is the writer's question, and the filing agent is a writer.

One honest note on evidence: the "before" side of this comparison is the drive's lived history, not a same-day twin run. So this is a demonstration of the mechanism, not a measured effect size. But the mechanism could not be clearer. The filing agent cited the routing line as its justification.

8.2 The handover-note experiment (B1 to B4)

Our incident log had already pointed at the brief three times. The one almost-invention was caused by OUR note claiming something the box did not contain. The three-folders-one-system mess was partly our doing: three notes had described the same system in three different wordings. And a specialist once flagged a section it could not fill, for the precise reason that our note had left out the session's scoping questions. What the handover note omits, the archive lacks.

So we made the note itself the experiment: same three documents, same eight facts, four notes (by the book, thin, bossy with two deliberately wrong destinations, and structured with pointers to the exact existing records), four sealed arenas. One result arrived before the runs even started: the production tooling's dispatch-time guard WARNED about the bossy note's shape, exactly as designed. We ignored the warning on purpose. That arm now measures what ignoring it costs.

All four arms completed. The results settled two hypotheses and, better, falsified a third.

The verdicts per arm. Keep as default the by-the-book note (B1): itemised facts, named sources, exact-record pointers when people or systems are involved, cheapest to write well, nothing to fix. Reject the thin note (B2), and for a subtler reason than expected: it lost no facts, but only because the documents happened to be attached and complete, and it still forked a person's identity on one vague word. Its safety is borrowed, not owned. Reject the bossy note (B3) as a practice, even though this fleet survived it: assigning destinations adds nothing when the filer is good (it re-derived every placement anyway) and becomes pure downside the day the filer is weaker or the wrong instruction subtler. Reserve the structured form (B4): it matched the by-the-book note exactly, and its one real virtue, forcing the pointers a prose note can forget, earns its keep when notes are written tired, late, or by many different hands.

Three findings, in order of surprise:

The attached documents are the safety net. Even the three-sentence note lost nothing, because the archivist reads the box, not just the note. Our "thin notes lose knowledge" hypothesis needs revising: richness of the note matters when the documents are missing or incomplete. When they are attached, the note's real job is smaller and sharper.

That sharper job is naming things. The thin note said "Jonas". The company record says the founder is Jonas Behr. The thin arm, honestly unable to confirm they are the same person, created a second person file with a flag. The three arms whose notes carried the full name filed one clean record. One vague word, one forked identity, no concurrency required. This is the same failure the storm produced with three folders for one system, reproduced by a single polite writer. The rule that falls out: a handover note names people, companies, and systems by their exact existing record, not by a word.

The fleet resists a bossy note. We told arm B3 where to file everything, and two of the instructions were wrong on purpose. It checked our instructions against the drive's own rules and rejected the wrong ones, quoting the folder's test question back at us, and it refused our wrong "keep this private" order the right way: by sharing a properly blanked version instead of the original. Our rulebook had always feared that naming destinations in a note would anchor the filer to the wrong drawer. At this run quality, the fear did not materialise. We keep the rule (notes carry facts, not destinations) because robustness should not be load-bearing. But the measured cost of one bossy note, this day, was zero.

9What we learned

The comparisons first, because they carry the practical decisions. Four times in this suite, two ways of doing the same job met head to head, and every contest produced a loser whose failure explains the winner:

Underneath the contests, every defect the suite caught fits one of four shapes, and each shape needs a different kind of fix. The most common mistake in maintaining agent systems is treating all four as shape 1, and responding with more prose.

Three more laws, continuing the list begun in part 1, whose Laws 1 to 5 cover the reading side:

Law 6 · Sharing is a ladder, and every rung is a gate

Your notebook, the team's binder, the company archive. Each rung up needs more translation and more discretion, decided at filing time, not discovered later. The founder-habit trap is the bottom rung of the same ladder whose top rung the adversary guards. Most systems only guard the top rung, and let everything below promote silently.

Law 7 · The write side borrows, the read side pays

Every write-side defect came back later as a read-side bill, and we measured it: the three unmerged folders taxed three of eight retrieval questions with overlap warnings; the missing "may we name them?" flags turned a clean question into an "undetermined." And the reverse: the one destroyed fact that survived did so because careful filing had stored it twice with the source linking them. Filing quality is not tidiness. It is what the read side gets to spend.

Law 8 · Honesty held better than accuracy

Across ten runs, the fleet's most dependable behavior was telling the truth about itself. The archivist that asked instead of inventing. The specialist that refused its own boss's faulty packet. The overwriting run that reported its own incident with a file-by-file account of what it had and had not checked. The run that corrected a colleague's false "deleted" note in the shared log. Systems that tell the truth about themselves can be fixed. Design for that property first.

10Recommendations

11Limitations

• One run per scenario. Agents are not deterministic. We treat each result as an existence proof or a defect discovery, not a statistic, and we lean only on large or structural differences.
• The variance finding cuts both ways. Shape 1 (three readings of one rule) is itself evidence that some behaviors reported as "held" might vary on a rerun. The floor behaviors are corroborated across multiple runs in the suite; single-run results are marked as such.
• One scale. Roughly 40 to 60 files per workspace. The concurrency findings should get worse with scale; the index findings should matter more. Both are expectations, not measurements.
• The index experiment's "before" is historical. Its baseline is the drive's lived history of splits, not a same-day control run. The mechanism evidence (the agent citing the purpose line) is direct; the effect size is not measured.
• Computation is mostly unmetered. One specialist run was metered (263,518 tokens). We report no other cost figures rather than estimate them; part 1 carries the cost analysis for this fleet's architecture.
• The traps are ours. Pass rates measure performance on failure classes we consider representative, drawn from real incidents. They are not a general quality score.
• The adversary is one agent. Its two verdicts were calibrated (a known-weak file failed; permitted attributes were not flagged), but adversaries can always be stronger; a better one might find more.
• The bossy-note result is one run against one strong fleet. Weaker filers, or subtler wrong instructions, may well anchor. We keep the facts-not-destinations rule precisely because we do not want robustness to be load-bearing.
• All names are fictional. The regional traps use real-world categories (payment schemes, identity apps, dialect) because their identifying power is exactly the point. Any resemblance of the invented company to real firms is accidental.

12Further work

• Anchoring under pressure. The bossy note failed to mislead a strong fleet once; vary the filer's strength and the wrongness's subtlety to find where anchoring begins.
• The merge pass. The suite established that one thing can end up with three homes under parallel writing; the repair job (and when to run it) deserves its own measurement.
• A stronger adversary. Escalate the reviewer (stronger models, web access, several independent reviewers) to find the ceiling of the game.
• Rung by rung. We tested private versus shared; the rungs in between (personal to team, team to company) each deserve their own trap set.
• Reruns. Repeat the floor scenarios several times to put ranges around "held."
• Ten times the drive. The same suite at a much larger corpus, where index quality and write discipline should dominate.

13Glossary

Adversary (adversarial reviewer)

An agent whose only job is to break a claim. Here: look at a file marked safe to share and work out whose it is. The opposite of a checklist.

Archivist

The lead filing agent: reads the box of papers, breaks it into facts, files the private side, hands shareable material to specialists, and checks every write.

Blank form (anonymisation)

Turning a filled-in document back into a reusable template: every specific value becomes a labeled blank with a note saying what goes there.

Brief (handover note)

The note a dispatcher writes for a filing run: what happened, which documents are evidence, which facts are candidates. Section 8's experiment.

Example entry

The illustration next to a blank on the form. This suite's central leak: examples must be decoys from an unrelated market, not paraphrases of the real value.

Filing (consolidation)

The writing job: archiving what was just learned so it can be found later. This note's subject.

Index

The drive's map: what lives where. With a purpose line and a test question per folder it becomes the filing agent's brain. Without them, placement is a guess.

Pointer (entity handle)

A reference to the exact existing record for a person, company, or system, used in a brief instead of describing the thing in words. Stops one thing acquiring three homes.

Round trip

The end-to-end test: file knowledge, then have a fresh agent find it again. The only score that ultimately matters.

Rung (sharing ladder)

A boundary knowledge can cross: personal to team, team to company, company outward. Each rung has its own translation and discretion gate.

Section specialist

An agent the archivist dispatches to translate and file material for one section of the shared shelves. Has the authority to refuse a faulty instruction, and used it.

Shared shelves

Any part of the store whose readers are not the people who wrote the source material. Everything crossing into it plays by the blank-form rules, logs included.

Trap

A planted, realistic failure opportunity with a known correct answer, used to score quality without opinion. 24 were planted here, in 7 families.

Verification listing

The store's own directory listing: the only accepted proof that a write really happened. Doubles as a smoke detector when it shows a stranger's changes.