Coldstart is in final cutover and has not been released. The behavior on this page runs in the development build; there is no release you can install today.
A reusable method — how to audit a container image, how to structure a piece of outbound writing — does not belong to any one project, and it should not be permanently loaded in every project either. It sits in a library, and something has to connect a request to the right entry without reading the whole shelf.
This page is about that connection, and mostly about what it does when there is nothing to connect to.
[Visual 4.2 — The query arrow from the section opener, enlarged into a left-to-right flow. On the left, a box labelled "one request, one tier per call" with three stacked entry points: "exact trigger phrase", "closed facet values", "free text". Each entry point feeds a single middle column drawn as two separate panels stacked: an upper panel labelled "what a query matches against" and a lower one labelled "what a query returns", joined by a bracket reading "both generated from the cards; a disagreement between them is a refusal, not a degraded answer". Out of the middle column run two exits. The upper exit passes through a narrow gate labelled "eligibility — rows this machine cannot use are removed" and ends in a short list of at most five rows, tagged "returned". The lower exit ends in a bordered receipt showing four lines: the query as asked, the near misses if any, "NO-MATCH", and a closing instruction reading "say no card covers this and proceed on general skill; do not invent one". A caption reads "the two exits are the whole contract".]
What is on the shelf
The unit is a card: one folder holding a single method, plus any references or scripts it needs.
What makes it findable is the contract at the top of the card, and every part of it is declared rather than inferred. A card carries a stable name, a description that must include an explicit statement of what it is not for, a list of trigger phrases, and three values drawn from three closed vocabularies: what it produces, the domain it belongs to, and the shape of the work it suits. It also carries an eligibility block and a provenance record naming where the card came from and under what licence.
Each of those vocabularies is a fixed list. A value outside it is not a new category; it is a defect, and the card is refused rather than filed. That strictness exists because the alternative is well documented: an open vocabulary drifts, the same idea arrives under three spellings, and the index quietly stops being able to group anything.
What a query actually reads
Nothing queries the cards directly. Two projections are generated from them: one that a query is matched against, and a separate one holding the short rows a query returns.
Splitting those is deliberate. The thing you search is not the thing you read, so each can be shaped for its job without compromising the other. The important consequence is what happens when the pair is missing, unreadable, or disagrees with itself — the query refuses outright. It does not fall back to a partial index or a stale one. A search surface that quietly degrades is worse than one that stops, because a thin result set looks exactly like a small library.
Three tiers, one per call
A request enters through exactly one of three routes, and they are ordered by how much interpretation they involve.
An exact trigger phrase is a lookup, not a search. It reads the index and returns what is declared under that phrase, with no ranking involved. Case and word-order variants of a declared phrase still resolve, because the comparison is on the set of meaningful words rather than the literal string.
Facet values intersect the closed vocabularies: everything that produces this kind of thing, in this domain, of this shape. Here a value outside the vocabulary is treated as a mistake in the question rather than an answer of zero — you asked in a language the index does not speak, and being told that is more useful than being told nothing matched. A legal value that no card happens to carry is a genuine empty result and is reported as one.
Free text is the only ranked tier. Cards are ordered by the product's own scoring, weighted toward a declared trigger phrase appearing whole in the request, then toward facet values, then plain word overlap. Anything scoring below a threshold is not a match at all.
That threshold was measured rather than chosen: it was set from a small authored set of known question-and-card pairs, low enough to keep the hardest paraphrases and high enough to drop rows that share a single incidental word. At most five rows come back from any tier, which is a deliberate ceiling on how much a query can pull into the session.
The eligibility filter is not a ranking
Before results are returned, rows the current machine cannot use are removed — a card requiring a part of a stack the profile does not have, or a service tier it is not on.
This is a filter, not a preference. It never reorders results and it never promotes anything; it only removes. And when there is no profile to read, nothing is filtered at all, with the result stating plainly that the filter was not applied. An absent profile constrains nothing, and silently behaving as if it did would be the same class of lie as a partial index.
The empty result, as it actually behaves
The worked example is one that was run against the development build rather than imagined.
The request: extend an existing CSV export without accidentally exposing a new field. It is a plausible, specific, well-formed question, and it is the kind of thing a capability library sounds like it should answer.
What came back was NO-MATCH, with near miss: none — nothing in the library shared even a word
with the request — and the closing instruction to say that no card covers this and to proceed on
general skill without inventing a card name. There is no CSV-export card, so there was nothing to
find.
That result is the mechanism working. The failure mode being avoided is the one where a nearly-related entry is returned with confidence, gets applied to a job it was never written for, and looks like expertise the whole way through. An empty answer that names what it looked for leaves the session exactly where it was, which is the correct place to be when the shelf is empty.
Limits
Reachability is not quality. Everything here establishes that declared cards can be found by the phrasings they declare. None of it establishes that a card teaches a method well, that the returned card was the best one, or that routing makes any piece of work faster.
The ranking is not proven better than the naive alternative. Against the authored test set, the structured route ties a simple word-matching baseline rather than beating it. It is a defensible design with an honest evidence boundary, not a demonstrated improvement.
An honest miss can also be a wrong miss. A relevant card that scored below the threshold, or that was removed by the eligibility filter, produces the same empty result as a library that genuinely has nothing. The receipt tells you the search found nothing. It cannot tell you that nothing was there.
Next
Cards are one shape of guidance. The other is much smaller and much more numerous: depth on demand.
How current this page is
Checked on 10 August 2026 against the product's current card contract, its query tiers and its recorded release state. The worked example is a query run against the development build on that date; a later library could answer it differently. To ask whether a claim here still holds, or to report one that does not, write to [email protected].