{"path":"research/what-a-record-linkage-project-teaches-deliberus.md","content":"# What a record-linkage project teaches Deliberus\n\n**Status (2026-10-06)**: Claude's reading, made at the founder's request, of what a friend's project\nsays to this one. **Proposals only.** They are in the ruling ledger as Arc Z\n([session26-open-rulings.md](session26-open-rulings.md)). Nothing here is built.\n\n**The request**: to read a friend's mail and message about what this project can learn from his, and\nto think it through. The friend and his project are not named here, and nothing of his system's\ninsides is quoted: it is unreleased and the correspondence is private. The named material is kept in\nthe private notes.\n\n**What was read, and how far**\n\n| Source | Read |\n|---|---|\n| A write-up of 18 September 2026 by the friend's coding assistant, working inside his repository. By its own list it had read twenty pages of our live site in full (eighteen documents, the landing page and the about page) and nine parts of his project's rule book | In full (the mail and its attached copy, 298 lines, thirteen lessons) |\n| A second analysis of 5 to 6 October from the same kind of session, which says it read three of our documents in full: five lessons for our sameness check, seven things his template could take from us, seven things we could take from it | In full |\n| A message of 6 October: what Deliberus should build first | In full |\n| A conversation between the friend and a different assistant about what his database is, read from the database alone | In full. Its praise is discounted (§ 2) |\n| His general template, as a database dump of 23 September | As text, never run. An early version: it states its laws in comments and enforces only keys. **The later version the October analysis describes was not seen** |\n| His project itself | **Not read.** Everything said about it here is his assistant's report |\n| Our side | [machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md) §§ 3b and 6, [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) §§ 3 and 4, [sameness-merge-and-split-across-fields.md](sameness-merge-and-split-across-fields.md) § 3b, [what-confirmed-means.md](what-confirmed-means.md), [pass-quality-and-the-choice-of-model.md](pass-quality-and-the-choice-of-model.md) § 4c; the function all model calls pass through; the live sameness batch |\n| Outside | Three papers on record linkage where the fields are not independent: one abstract, two as search extracts (§ 4a) |\n| Measured | One read-only count on the live graph: which sameness candidates have evidence besides their wording (§ 4a) |\n\n---\n\n## In plain language\n\nA friend has built a database that works out which records belong to the same company: patents,\ntrademarks, court cases, registers. That job is called record linkage, and it is our sameness\nproblem with firms in place of claims. He also lifted the design into a general template for any\nproject a language model will work in.\n\nThree things stand out.\n\n**First, write down every machine judgment as it is made.** His machinery is ordinary arithmetic,\nso anything lost can be computed again. Ours is a model's reading. Ask again and the answer can\ndiffer, sometimes from a different model. So a judgment written today without a record of what made\nit has lost its origin for good. We had already designed this record. We had placed it second in\nline. It belongs first.\n\n**Second, for sameness, count kinds of evidence and never count readers.** Three readers of the\nsame two sentences are one witness. A second kind of evidence is something else about the claim:\nwhere it sits in its own source's argument, who asserts it, what it is attacked by, what its author\nsays. On the live graph that other evidence is present for 145 of 147 candidate pairs, and nothing\nreads it.\n\n**Third, a state belongs in its own place, never in a flag.** A flag asks every later reader to\nremember a filter. We have paid for that four separate times.\n\nThe rest follows from those three, or is smaller.\n\n---\n\n## 1. Why this project bears on ours\n\nIt turns many rows into one actor. We turn many phrasings into one claim. Both compose separate\npieces of evidence into an identity, keep every judgment open to revision, and want the whole thing\nreadable by a machine that arrives knowing nothing.\n\nRecord linkage is about fifty-five years old, and it is one of the six fields our sameness document\nalready draws on ([§ 3b](sameness-merge-and-split-across-fields.md)). So the craft is not new to this\ncorpus. What is new is a long, recent run of mistakes made and recorded by someone practising it\nwith the same tools we use.\n\nOne difference in kind decides what transfers. His matching rests on registers that never knew they\nwould be combined, and on deterministic checks. Ours rests on a model's reading of contested\nmeaning. **So what transfers is the discipline around a judge, and what does not is the hope of\nreplacing the judge.**\n\n## 2. How far to trust the source\n\n- **What it says about us held.** Each claim about this corpus that was checked against the\n  repository was right: three of twenty-five call sites record which model answered; 89 *overlaps*\n  answers, about 64 of which mean *one is a reason for the other*; the finding that a model asked\n  directly does about ten points worse; the four kinds of similarity our April document asked for;\n  five meanings of one flag.\n- **What it says about his system cannot be checked from here.**\n- **One family of model read both projects, wrote both analyses, and is the reader here.** The\n  write-ups are careful and say what they read. The other assistant's conversation is flattering\n  throughout, and is used only for facts about the database.\n- **The write-up offers the two projects' agreement as evidence.** By this corpus's own test\n  ([contingency-and-convergence-across-levels.md](contingency-and-convergence-across-levels.md))\n  agreement is evidence only when the lines are independent, and these are not: the same assistant\n  family helps write both rule books, the builders are friends, and his assistant had read our\n  documents. The independent ground is the older literature, and § 4a leans on that.\n\n## 3. The record of machine judgments comes first\n\n**Their argument.** A derived row in a deterministic system can be lost and computed again. A\nmodel's judgment cannot: it is testimony. So the record of who judged, on what, under which\ninstructions, is the one thing that cannot be rebuilt later, and every day without it loses some for\ngood.\n\n**What we already had.** [Machine work § 3b](machine-work-as-a-standing-property.md) designs this\nrecord: one row per examined subject and pass, with the model, the instructions, the inputs and\ntheir revisions, and one of four outcomes (found something, examined and found nothing, could not\ndecide, failed). § 6 of that document orders the building, and puts this record second, after the\nmachinery for embeddings. It also places all of that building after the open reviews, and nothing\nhere changes that.\n\n*One thing, three names so far*: that document calls it the *coverage watermark* (§ 3b) and the\n*judgment ledger* (§ 6); this one says *the record of machine judgments*, to keep it apart from the\nruling ledger. They are the same record.\n\n**Built since (2026-10-06, the same day).** On the founder's word (*\"Yes build it.\"*) the record's\nfirst step exists: every call through the shared function is written as a row, and the database\nrefuses a change or a deletion. What it does and does not yet do:\n[storing-the-record-of-machine-judgments.md](storing-the-record-of-machine-judgments.md) § 7a.\n\n**What their reading adds.**\n\n1. **The order.** The embedding machinery was put first because embeddings are safe to practise on.\n   That reason stands, and it is a reason about machinery. The record is bookkeeping with no model\n   call in it. It can go first without contradicting the order of the rest.\n2. **The pipe writes it, never the pass.** One function receives 25 of the places that ask a model\n   for a judgment (three others go around it, below). If that function writes the row and hands\n   back its number, no pass can forget to. Today three of the 25 keep any record of which model\n   answered.\n3. **The model is read from the reply, never from the request.** Our fallback chain changes which\n   model answers without saying so, and both splits written on the live site in October were cut by\n   a model the splitter was never measured on.\n4. **A version is a hash, never a typed label.** A label typed by hand stays the same while the\n   thing it names changes. A hash of the instructions as sent cannot.\n5. **Every machine write carries the number of the judgment behind it**, and one test fails when a\n   path writes a machine-made link or claim without one.\n\n**Three things their account leaves out.**\n\n- **Three model calls do not pass through that function**: reading a PDF, extracting a PDF's text,\n  and transcribing audio. A transcription is a judgment about what was said, so they belong in the\n  record too.\n- **A community's compartment has its own database**, and judgments about its material must stay\n  there. The record holds instructions, inputs and replies, so whether it is ever shown publicly is\n  a separate decision.\n- **A person's label is about a judgment.** When someone marks a machine's answer right or wrong,\n  the label should carry the judgment's number. Otherwise a label on a link says nothing about\n  which model and which instructions it grades, which is what\n  [the evals reading](the-evals-loop-and-our-machine-passes.md) needs it to say.\n\n**When it is owed.** Few machine judgments are written on the live site in a day now, so each day\nwithout it loses little. It is owed before the next run of a pass over the whole graph, before the\npaid engine is switched on, and before any machine merging. Each of those writes judgments by the\nthousand.\n\n**The smallest version**: one table (pass, the model that answered, a hash of the instructions, the\ninputs with their revisions, the reply as received, the outcome, the time), made write-once by the\ndatabase itself; one change to the shared function; a judgment number on every machine write; one\ntest.\n\n**How it is stored, proposed the same day** on the founder's question (*\"how can we store it best\nto not swamp the graph/DB with this metadata?\"*): on disk and never in the graph, every text once,\none narrow row per call, one number in the graph. Measured and costed in\n[storing-the-record-of-machine-judgments.md](storing-the-record-of-machine-judgments.md): an estimated 10 to 15 MB for everything to date, against a graph of 84 MB in memory.\n\n## 4. Six lessons for the sameness check\n\n*The sameness check (the keeper) reads two claims and gives one of seven answers: same, contains\neither way, opposes, overlaps, unrelated, or cannot tell.*\n\n### 4a. Count kinds of evidence, never readers\n\nThe founding model of record linkage adds up agreement across fields, and it assumes the fields are\nindependent. Where they are not, agreement is counted twice. One study measured the weight given to\ntwo correlated fields agreeing at 16.87 when the dependence was ignored and 13.55 when it was\nmodelled, and concluded that ignoring it overstates how much the fields identify (Tromp and\ncolleagues 2008, abstract read; the assumption itself is stated in two later papers read as\nextracts).\n\nFor us: the similarity search, the sameness check and a second family of model all read the same\ntwo sentences. That is one field read three times. Our plan for making checkers independent adds\nreaders ([the two-family jury](fragile-checkers-and-the-verification-bottleneck.md)). For sameness\nit needs a second axis: **other kinds of evidence about the claim.**\n\n| Kind of evidence | Where it lives today |\n|---|---|\n| The two sentences | read by everything |\n| The claim's place in its own source's argument | extraction's links within a source |\n| Who asserts it, and in which source | `attributed_to`, the source link |\n| Which sense of a shared term it uses | the concept layer, on about a quarter of claims |\n| What attacks it, and what it is split into | the graph |\n| How people vote on it | votes |\n| What its author says it means | nowhere yet (the consent channel) |\n\n**An example from the live graph.** Two opposing articles on the minimum wage both carry the\nproposition that a wage floor harms the workers it is meant to help. One reports it as what\nopponents of the law argue. The other reports it as what a majority of economists believe. Read as\ntwo sentences they are one claim. The other kinds of evidence say what the sentences do not: both\nare reports, the two reporters disagree about it, and neither asserts it in their own voice\n([self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md)).\nA check that reads only the two sentences cannot see any of that.\n\n**One caution of our own.** A link the automatic linker wrote from text similarity is not a second\nkind of evidence. It is the first kind again, stored. So structure counts as a witness only when\nanother process wrote it: extraction reading the source, or a person.\n\nThis is the same finding as [the evidence research](what-counts-as-evidence.md), which found that\nshared origin is not represented here at all (ledger Q4). It reaches sameness from a second side.\n\n**Counted on the live graph, 2026-10-06** (read-only, no model call;\n`uv run python scripts/sameness_evidence_kinds.py`). There are 147 resemblance links between\nsubstantive claims, every one across two sources. The sameness check read 135 of them and the\nlinker wrote the other 12 on similarity alone; either way the judge was shown the two claim texts\nand nothing else. How many have a further kind of evidence at both ends?\n\n| Further kind, present at both ends | Pairs of 147 |\n|---|---|\n| A place in its own source's argument (a link to another substantive claim of that source, not written by the linker) | 105 |\n| A named asserter | 135 |\n| A link to a concept | 13 |\n| None of the three | 2 |\n\nSo the material is there for nearly every pair, and nothing reads it for sameness. Two cautions\nkeep this from being good news yet. **Present is not the same as telling**: the asserter is the\nsame name at both ends in 2 of the 135, and a different asserter says nothing against sameness.\nAnd **nobody has defined how two places in two arguments are compared**, which is the kind most\npairs have. The kind that would tell most, the sense of a shared term, is there for 13.\n\n### 4b. Only ruling out produces a no\n\nAgreement can pile up without limit and still be wrong, as two identical sentences that mean\ndifferent things show. A no comes from something that rules sameness out. Ours: an attack link\nbetween the two, different senses of a shared term, the check's own *opposes*, and the strongest,\nthe author refusing the rendering.\n\n**When strong evidence for sameness meets one of these, the result is a contradiction, and a\ncontradiction is its own outcome.** It is neither a tie nor *cannot tell*, and no machine settles\nit. It is also the most useful thing to put in front of a person: either a link is wrong, or one\nwording is carrying two claims.\n\nOn the live graph, 9 of the 147 resemblance pairs also carry an attack link between the two\nclaims. Resemblance is not a sameness verdict, and no pair is marked for merging today, so no\ncontradiction of this kind stands. These nine are where the first one would appear.\n\n### 4c. Compute the answer from narrow questions\n\n[Microtasks § 4](microtasks-in-broad-daylight.md) already proposes deriving the answer from four\nyes-or-no questions (ledger T5, on which the founder said on 2026-09-30 that he was still absorbing\nit and not ready to decide). Their reading supplies the fixed table:\n\n| If | Then |\n|---|---|\n| each entails the other | same |\n| only one entails the other | contains |\n| they cannot both be true | opposes |\n| they share content, and neither entails the other | overlaps |\n| none of these | unrelated |\n\nTwo things we would change in it.\n\n- **The fourth question has to be *is there a statement that both assert*.** Asked as *do they share\n  content*, it would be answered yes for the about 64 pairs where one is a reason for the other, and the\n  key would keep both meanings it has today.\n- **The last row should read *no sameness relation*, because relatedness should be asked once.**\n  Today the sameness check can answer *unrelated*, which ends the matter, and the relation\n  classifier then asks its own *related* question of every pair that got through. That is one\n  question with two askers, the shape that let the less careful asker win on *same* in September.\n  Whether one claim bears on the other would be the classifier's question alone.\n\n### 4d. One question, one asker, one writer\n\n*Same* and *related* are different decisions. A relation never turns into a merge. Two of our\nmeasured faults were these two decisions tangled: the classifier answering *same* over the sameness\ncheck, and *overlaps* carrying both *a reason for* and *a shared piece*. A third is still open: two\npasses write the resemblance link (ledger W4).\n\n### 4e. A group must not vouch for its own growth\n\nOne of their lessons, as reported: a group picked up wrong members, a summary of the\ngroup was cached, and new records were then compared with the summary, so the wrong members\nrecruited more of their kind.\n\nFor us this is a rule to hold before any group of same claims exists: **a candidate is compared with\nthe members' own words and their own links, never with a summary, an average or a stored profile of\nthe group, and sameness is never chained from pair to pair.** The second half is already in our\nsameness document, from the Semantic Web's experience. The first half is new here, and it is the\nsame hazard our concept research named when it found a word's listed senses reflecting the pass's\ntemplate, and a model's averaged sense standing in for anybody's\n([core-periphery § 4 and § 6](core-periphery-and-the-concept-layer.md)).\n\n### 4f. The test set needs look-alikes that are different\n\nA test set that holds only pairs that should be joined cannot show a check that joins too much. The\nlive sameness batch has 20 planted wrong pairs among 120, and all 20 are of one kind: one claim\nattacks the other. None is the other trap, **the same words carrying different claims**. We have\nreal ones: the fourteen support links between look-alike claims about different things, retired on\n2026-09-13. They belong in the answer key for the machine.\n\n**Where our job is harder**, by their own account: we have no registers, meaning is contested where\na firm's identity is only hard to find, and people will try to game us. **Where we are ahead**: more\nanswers per pair than same or not, asking each pair in both orders, and testing the person who\nreviews.\n\n## 5. Beyond the sameness check\n\n### 5a. A state is a place, never a flag\n\nTheir template refuses a mistake by making it impossible to write, in place of forbidding it. A\nproposal sits in its own table, and the code that decides cannot read that table.\n\nWe have learned this four times without naming it:\n\n- a strength of 0.5 meant *nobody looked* and was compared as if it meant *middling*;\n- refused links kept moving strength, because three of four queries lacked the filter;\n- community material was given its own database, because a hidden flag holds only while every\n  reader remembers it;\n- one flag came to have five meanings ([what-confirmed-means.md](what-confirmed-means.md)).\n\nRetiring that flag is ruled (2026-09-28, ledger Q32) and its build is pending. The ruling gives\neach of the flag's jobs a home, and for a machine proposal's standing the home it names is *its own\nfield*. Their lesson bears on that one job: a field is still a flag. **A proposal that must not\ncount gets its own kind of link, or its own table**, so a query for the links that count cannot\nreturn it. We already do this in one spot: the sameness check's split proposals are kept on a\nresemblance link and never written as part-of links, so no query for parts can return one. Since\nthe build has not started, this costs nothing to take in now.\n\n### 5b. The door\n\nTheir template has one function a newly arrived model calls first. It says what the system is, what\nits rules are and what enforces each, and what state each check is in, all computed at the moment\nof asking. Three properties are worth taking for our agent surface.\n\n- **Computed, never kept by hand.** Our index for agents once promised links that did not exist\n  ([dogfood run 4](dogfood-run-4-agent-surface.md)).\n- **It says what it scans.** A register can be clear about everything it shows and silent about\n  whole classes it holds. An index that states its own rule can be checked against it.\n- **Every reading carries when it was made and by what.** A number with no age reads the same fresh\n  or stale.\n\nA door would also force two things we know are held only by accident: the public totals do not say\nwhich claims are asserted and which were minted, and the function that would say so has no caller.\n\n### 5c. When the system knew\n\nThey keep three times: when the world changed, when it was said, and **when the system worked the\nanswer out, with which version of its rules.** We have the first two in part and not the third. An\nanswer handed to an agent with that third stamp can be worked out again later and compared. It is\nwhat makes a machine-readable answer a source and not an oracle. The version in the stamp is a hash\n(§ 3, point 4).\n\n### 5d. Witnessed change for the rules\n\nThey sort everything by one question: if this were destroyed, could it be restated exactly from\nwhat survives? Ground cannot, so it is never touched. A judgment can be re-derived and its origin\ncannot, so it is corrected in the open and kept. A cache can, so it is disposable. The rules that\ndo the deriving are a fourth thing: change them as often as the work needs, and never silently.\n\nFor us the rules are the strength code, the instructions to each pass, and the list of argument\nschemes. Our rulings are recorded in the founder's words and corrections have their own kind of\ncommit, so the habit exists. What is missing is the join: nothing ties a change in an instruction's\nhash to the ruling it rests on.\n\nTheir sharpest question for us: *who may change the scoring rules?* Today the same person, from the\nsame seat, who can edit any claim. That is the gap three of our documents call the missing\nconstitution. **A record of every change to the rules, with what changed, why, and on whose word,\nis a first article that can be built before governance is solved.**\n\n### 5e. A helper is not a blind reader\n\nTheir measurement: a helper started inside the project had read the project's instructions before\nits first step. Ours agrees: a general helper here starts with both instruction files.\n\nSeveral of our tests need a reader who does not know the corpus: the stranger test, and the planned\ncomparison of what an agent can do with raw sources, flat claims and the graph. We have measured\ntwo ways to start a reader without the files (a search helper, and a model run from an empty\nfolder). **A test that needs a blind reader should run on one of those and say which.**\n\n### 5f. Smaller things\n\n- **No count in a name or a standing line.** A count written into a name goes stale the day the\n  count changes. Ours: the always-loaded file says *five known flattening mechanisms* while a sixth\n  is proposed, and the sameness check is *the keeper of the seven keys*.\n- **A check that tries every forbidden act against the live system and undoes it.** We have one\n  such, the sweep that asks every route for a stored address. Most of our other refusals are\n  tested on test data only.\n- **Which concepts earn a node.** Their test: would this concept, put at the centre, connect two\n  things the world thinks unrelated? For us, *freedom* fails and *freedom in one stated sense* may\n  pass. It also bears on how candidate pairs are found: a shared hub concept is a poor reason to\n  compare two claims.\n- **Type the identifier.** A bare number handed across a tool boundary gets guessed at, and a wrong\n  guess that returns a plausible object is the worst failure. Our claim identifiers say what they\n  are. Whether every other identifier does was not checked here, and it should be before an agent\n  surface ships.\n\n## 6. What does not transfer\n\n- **Their doctrine of scale.** Their database is many orders of magnitude larger than ours. We hold\n  33 sources.\n- **Settlement by one person's word.** It suits one keeper digging through a finished past. It does\n  not survive a commons, and by their own account it has nothing to say to our two hardest open\n  problems.\n- **Determinism as the aim.** Firms have registers. Meanings do not. Making sameness of claims\n  deterministic would mean choosing one definition of meaning and building it in. What transfers is\n  determinism in how answers are combined, which we already have in the strength calculus.\n- **The database as the application.** Their rules live in the database engine, as triggers and\n  permissions. Our graph store has no triggers, so our equivalent is the write path plus tests that\n  run the real query. No change of store is proposed.\n\n## 7. What we have, built or designed, that it lacks, by their account\n\nA confession value on closed lists. A test of the person who reviews. Asking each pair in both\norders. A ladder by which a pass earns the right to write and loses it (proposed here, partly\nbuilt). And one thing they call strictly better for a contested subject: publishing a commitment\ninto the graph as a claim that the graph's own machinery can attack (designed here, not built).\n\n## 8. Proposals (unruled, ledger Arc Z)\n\n| Ledger | Proposal |\n|---|---|\n| Z1 | **The record of machine judgments is built first**, before the next whole-graph run of any pass, the paid engine or machine merging: written by the shared function, the model read from the reply, versions as hashes, a number on every machine write, one test |\n| Z2 | **Sameness needs two kinds of evidence.** Every reader of the two sentences counts as one kind, and a link the linker wrote from text does not count as a second |\n| Z3 | **A contradiction is its own outcome**: strong evidence for sameness against something that rules it out. No machine settles it, and it goes to a person first |\n| Z4 | **The sameness answer is computed by a fixed table** from yes-or-no questions; *overlaps* means only that both assert a common piece; relatedness is asked once, by the relation classifier. Input to T5, which he has not ruled |\n| Z5 | **A group's membership is judged against its members**, never a summary, an average or a stored profile, and sameness is never chained |\n| Z6 | **The answer key for sameness holds look-alikes that are different claims**, the fourteen retired links first |\n| Z7 | **A state that changes what a reader may do with a row gets its own place**, a kind of link or a table, never a flag or a field. Sharpens one of the homes named in the ruled Q32 |\n| Z8 | **The agent surface opens on a computed door** that says what it scans, and every answer carries when it was worked out and under which rules |\n| Z9 | **Every change to the rules is recorded with what changed, why and on whose word**, as the first article of the constitution |\n| Z10 | **A test that needs a blind reader runs on a path measured to carry no project instructions**, and says which |\n\n## 9. What would show this reading wrong\n\n- **If a second kind of evidence cannot be made to tell.** The worry that it rarely exists was\n  counted and did not hold: it is present at both ends for 145 of 147 candidate pairs (§ 4a). What\n  is open is whether it can be read. Nobody has defined how two places in two arguments are\n  compared, a different asserter says nothing, and the sense of a shared term is recorded for 13\n  pairs. If no reading of it separates same from different, Z2 would block every merge. That is\n  worth knowing before merging is switched on.\n- **If the record costs far more than one table and one changed function.** Threading a number\n  from the shared function to every write touches each pass. If that turns out to mean rewriting\n  the passes, the order in Z1 may still be right, and \"smallest version\" is wrong.\n- **If his system's reported lessons do not hold in his system.** None was checked there.\n- **If our own counts were misread.** The 64 of 89 is one model family reading its own stored\n  reasons.\n\n## Cross-references\n\n- [machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md): the record this\n  document moves to first place.\n- [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md): the four questions, and the\n  two meanings of *overlaps*.\n- [sameness-merge-and-split-across-fields.md](sameness-merge-and-split-across-fields.md) § 3b: the\n  field these lessons come from.\n- [what-confirmed-means.md](what-confirmed-means.md): the flag § 5a would replace with places.\n- [what-counts-as-evidence.md](what-counts-as-evidence.md): independence of sources, from the\n  evidence side.\n- [the-evals-loop-and-our-machine-passes.md](the-evals-loop-and-our-machine-passes.md): why a\n  person's label needs the number of the judgment it grades.\n- [agents-as-a-consumer-class.md](agents-as-a-consumer-class.md): the agent surface the door and\n  the third clock belong to.\n- [storing-the-record-of-machine-judgments.md](storing-the-record-of-machine-judgments.md): how the\n  record of § 3 would be stored.\n- [store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md): a third mail from the same friend, on storing uses and reading meaning, word senses,\n  and a graph of real things.\n"}