A brief is a budget.

The brief an agent receives at the start of a task has a fixed size, and the agent has no way to tell which of its sentences to discount. A memory distilled from every revision of a file restates the same fact many times, and without a rule against it the budget can be spent saying one thing. The rule has to be symmetric, and it has to know the difference between a repeat and a disagreement.

§1Distillation says one thing many times

The memory these notes describe is filled by distillation: each file an organisation writes down — a design note, a runbook, a decision record — is read, and the durable facts in it are stored as short sentences with a citation back to the file. A file is revised, and distillation runs again. A fact that survived the revision is re-derived, in slightly different words, and stored again. An identical sentence is stored once; a reworded one is stored as a new memory.

None of this is wrong as storage. It becomes wrong at the moment of reading. A brief is assembled by taking the memories nearest to the task, in order, until a character budget runs out — and the memories nearest to a task about the store can easily be the many ways of saying what the store is. Assembled that way, a brief can be mostly one fact in many phrasings from one file. In the test described in §3, six paraphrases of one process, all distilled from one file, are together longer than the 700-character brief they compete for. Every sentence in such a brief is true and well cited, and together they tell the agent one thing.

§2Overlap has to be measured against the larger of the two

The test is deliberately plain, so that it can be checked by hand. A sentence is reduced to its content tokens: lower-cased words of three or more characters, a fixed list of common function words dropped (through and see are not on it), a trailing plural folded to the singular. Two sentences overlap by the number of tokens they share. A candidate block is skipped when its overlap with any block already in the brief reaches 0.6, with two exemptions discussed below.

The question is what the shared count is divided by, and the first answer was the wrong one. Divide by the smaller of the two token sets and the measure asks "is the shorter sentence covered by the longer?" — which sounds like the definition of a repeat and is not.

carried   The stack is Postgres, Redis and the API, started with docker
          compose; see the deploy notes for the box.
          → stack · postgre · redis · api · started · docker · compose ·
            see · deploy · note · box   11 tokens
candidate Deploys go through docker compose on the box
          → deploy · through · docker · compose · box   5 tokens

shared 4
  4 ÷ smaller (5)   = 0.800   ≥ 0.6 → skipped: the precise line is lost
  4 ÷ larger (11)   = 0.364   < 0.6 → carried

Dividing by the larger set makes the test symmetric: each sentence must cover most of the other before either counts as a repeat. A vague line can no longer silence a specific one merely by containing its nouns. The cost is that some true repeats of unequal length survive, and that is the right direction to err in — a brief with one sentence too many is slightly less efficient, while a brief missing the one precise sentence is wrong.

§3Disagreement is not repetition, and a file is one voice

The first exemption is polarity. Two sentences that share every content token and differ by a negation are a pair a brief must not collapse, because they are a disagreement — one team does something, another has forbidden it, or a rule was reversed. A block whose negation differs from a carried block's is never a repeat, however high the overlap. The second exemption is length: a block of fewer than three content tokens is too short for the ratio to mean anything and is never skipped on it.

One more rule is not about wording at all. Paraphrases that escape the overlap test still share a source: they cite the same file. A brief therefore carries at most two blocks citing the same first source — in practice, one file; blocks with no source are not capped — and the rest of the budget goes to other files, other repositories and the organisation's own layer. Two, rather than one, because a long document can hold two genuinely different facts a task needs. The overlap test removes restatements that are close in wording; the cap is what stops the loose ones.

The rules have their own tests. In one, three paraphrases of the store's technology are carried once, and both "Run migrations by hand" and "Never run migrations by hand" are carried. In another, six long paraphrases of one process from one file meet a 700-character brief: at most two are carried, and the room left goes to other sources — here including the organisation's own fact, seated by the mechanism in Paper 002.

§4A filter at reading time cannot fix what was stored

Every rule above runs while the brief is assembled, on sentences already stored. That is where it is cheap to be careful, and it is also where some errors can no longer be undone.

Loose paraphrases survive. "The memory store is Postgres with the pgvector extension" and "Embeddings live in a Postgres table indexed by pgvector" say the same thing and share two content tokens of six: an overlap of 0.333, and both are carried. Token overlap measures shared words, not shared meaning. The real cure is upstream — distillation that recognises it is re-deriving a fact it already holds, and updates it rather than adding a new one — and that is not what this paper describes.

Polarity is read from a list of words. "never", "not", "avoid", "without", a contraction ending in n't and a few others mark a sentence as negated. A reversal phrased without any of them is not seen as one:

carried    Run migrations on API boot.              run · migration · api · boot
candidate  Stop running migrations on API boot.     stop · running · migration · api · boot

shared 3 ÷ larger 5 = 0.600   neither negated   → skipped as a repeat

The reversal is dropped, and the brief carries the rule it reversed. Extending the list narrows this without closing it, because the ways English reverses an instruction are open-ended. Where a reversal matters, it should be stated with a negation — or, better, the superseded sentence should be retired from the store, so that there is nothing for a brief to choose between.

The cap does not read polarity. If one file already has two blocks in the brief, a third that contradicts them is dropped like any other third block. A reversal is safest stated where it is not the file's third voice, or with the superseded sentence retired.

Order decides which phrasing survives. Of two repeats, the one ranked higher is carried, and the other is skipped. The rules decide how much of the budget one fact may take, not which of its phrasings is best. And when several repositories state the same practice independently, that agreement is evidence in its own right: it is counted, not deduplicated, by the stage described in The Inherited Default.