The gate, drawn twice

The gate appears twice on purpose. On the way in it decides what may become a record; on the way out it re-reads the sentence the engine produced and drops any claim the memory does not support. A diagram that draws it once has drawn a retrieval pipeline instead.

The pieces

Identity. A concept is a numeric key; a word is only a label. One spelling can be two entities, and one entity carries labels in several languages — which is how the same memory answers in Turkish, English and Spanish without a translation layer.

The graph. Records — subject, predicate, value — each stamped with source, trust and time, linked by cause, then, contradicts. Closed triangles witnessed twice teach a predicate transitivity, and derived facts are written with an #inference stamp below the speaking threshold.

The evidence index. The document's own sentences, kept verbatim, indexed at several window scales. Retrieval by words is deterministic word/graph lookup — explainable, and the same answer every run.

Rows. A table's row is a unit: the cells the document put together and no others. Where a table has SHATTERED — a PDF dump that drops each cell on its own line — the row length is read off the period in the cell lengths. Where it is written with DELIMITERS (| or tabs: markdown, ASCII tables, most of what an ingested file actually contains), the row is already whole on one line, and where a rule follows the first row that row names the columns, so each cell is stored under its own name. Either way no window may span two rows: cells the document never put side by side, gathered into one window, produce a claim that is past every gate by construction.

And one column is not a value at all. A run of rows ending in a bare number that never decreases, rises across at least three rows, and has no column NAME above it is an INDEX — a table of contents, a register. Its numbers say where a thing is written, not what it is, and they sit in the same row as the thing's name: asked for a regulation's number, a memory that keeps that column answers 140 — the page — while III-52.1 stands beside it, and the answer passes every audit because the claim really is in the evidence. So the column comes off before the document is split. This is the one place the store does not keep a line as the document wrote it; the row keeps all of its words. A NAMED column is a field and keeps its values however they are sorted, which is what separates an index from an ascending price list.

The meaning channel. A word search cannot cross a paraphrase: a lease writes RESIDES and a reader asks where somebody LIVES. So a 30 MB static multilingual matrix ships in the wheel and is attached by default — one lookup per word-piece and a mean, no transformer, no warm-up, no network.

It sits at the document layer, and which layer was measured rather than assumed. Over lines it moved nothing; over document profiles it took the store's own retrieval from 92% to 99% on 103 known-item queries, because a catalogue line reads "Aidiyet ve Motivasyon." — four words carry no subject, so vectors over lines compare fragments while vectors over a profile compare subjects. The opposite was measured too: choosing the lines INSIDE a proposed document by meaning rather than by words took the answering line from 78% to 31%. Meaning says which document; inside it, the question's own words are sharper.

The two rankings are fused by reciprocal rank (RRF, K=60) — which reads order, not score, so there is no weight to tune and a line the words already found cannot be pushed out by a vector's opinion.

Late interaction. A single vector averages a line into one point, and the averaging is where a vague question loses. So every piece of the question is scored against the piece of the line it fits best and the bests are summed — ColBERT's idea without ColBERT's model, since the piece vectors are already the table. It only ever REORDERS: no line enters by it and none leaves, so nothing it decides can reach the gates. Measured, it is worth more the vaguer the question gets (six terms removed from the query: MRR 0.707 → 0.807), and end to end it takes the answering line's chance of reaching the engine from 60% to 79%.

Both are parametric: Memory(encoder=..., reranker=..., dense=False).

The entity graph. The document's own, built with no model call. An entity is a phrase whose words occur together beyond chance (pointwise mutual information over the phrase's weakest split), which keeps varied company, and whose parts do not choose freely. An edge says one thing — these two are mentioned together, here — and carries the witnessing sentence, weighted by Dunning's log-likelihood ratio so an entity that appears everywhere does not become everyone's neighbour. An edge cannot be fabricated: it is an observation. It is also the partition — a question naming two entities reads the intersection of two posting lists rather than scoring the store. On a 36,472-line novel: 417 ms to 0.01 ms.

The composer. A memory that answers only the shapes somebody wired by hand answers a finite set of questions, and shapes are endless. So the engine proposes a PLAN over verified primitives — anchor a phrase to its dated line, take the latest, gather lines, span two dates, order them, filter by a cutoff, count, tally by month — and the interpreter executes only operations it knows, on phrases the store can anchor. The sentence is built from the final step's typed value by the library's own template: the engine contributes operation names and phrases, never an output word. An unknown operation, an unanchorable phrase, a dangling reference: the plan dies and no claim is born. Plan is a proposal; the primitives are the law.

The event organs. A count is the length of a VERIFIED list (the engine lists, the store checks each item, the number is what survives); a total is the sum of verified amounts; "which came first" and "how many days between" are date arithmetic. The shape reader routes at the door and its verdict binds both ways — organs silent and plan dead, the turn refuses rather than gambling prose.

Dialogue's two columns. WHEN an event happened is read from the sentence before the envelope's stamp (the engine proposes the date, the arithmetic admits it or keeps the stamp); WHO SPOKE is kept per line, so the asker's own lines seat first. Memory.distil() writes the events a chatty passage reports into the graph — listed by the engine, admitted by the passage's own words, gated like any other fact.

A living fact (0.7). A told fact can END — "I canceled my subscription" is a record whose own value says so — and the counting walk hands its candidates, laid out by day, to the engine for a currency reading: the store still disposes, the count never exceeds what the graph gathered. Each distilled thing carries its KIND ("magazine subscription") as an index key beside the bridges — it widens what a question can FIND and no gate or spoken word reads it. The asking day itself is metadata (session.asked_at, the plan's now), so "how many days ago" is arithmetic, not prose. And when every organ dies on a store whose every line is dated and small enough to read whole, a last organ reads the telling in time order — through the chain's own two-tier gate, digits non-negotiable — so dispersed clues can connect without loosening anything.

The trust ladder.

level	source	trust
5	operator	0.75	the person running the system
4	document	0.60	an ingested file, with its citation
3	model extraction	0.50	a candidate — if the gate admits it
2	inference	0.35	derived — below the speaking threshold
1	stranger	0.30	someone nobody has vouched for

The exit gate. Candidates are generated per evidence subset, then judged: a digit the evidence does not carry is vetoed outright, a mixture sentence — evidence words wrapped around content the evidence never supplied — is dropped, and the read-back re-extracts the claims and verifies each one. What fails is dropped, not repaired: a repaired sentence has no source.

In a multi-document store

Every evidence line opens with its document's name — a dateline — so the answer can say which programme, and the verifier can confirm it. A question that names a document gets that document's sentences (weighed by how many source names share the word: a word every source carries weighs exactly zero). One document cannot monopolise the block when others hold evidence. And the census — which documents speak of this, counted — rides into the answer as one line of store-attested fact.