The previous note explained what Apshan does: maintain the chains of evidence that reasoning operates over. Tariffs indexed by country and effective date. Trade press normalised across languages. Crop data linked to fibre type and region. Pre-built, so a model spends its compute on the conclusion, not the assembly.
This note explains why building those chains is slow work, and why we have built Apshan that way on purpose.
What reasoning-grade infrastructure takes
We use the phrase reasoning-grade infrastructure for the standard a body of evidence has to meet before a model can reason over it without producing brittle answers. Each link must be traceable to source, dated to capture, scored for confidence, and linked across the domains that determine the outcome. None of those properties can be added later without compromising the rest.
The work falls into three disciplines, none of them quick.
Sources are dated and scored at the edge.
A signal without a date and a confidence band is not infrastructure-grade. The cotton harvest report from 2024 is not the same record as the cotton harvest report from 2026. A supplier statement in a press release is not the same record as the same statement under oath. These differences are obvious, and they are slow to maintain. Every record entering Nari is verified, dated to the source's own publication, and scored before it earns a place. There is no automation for the edges of this work. The edges are most of it.
Cross-references are composed across domains.
Fashion sits downstream of weather, supply, customs, tariffs, sentiment, runway, retail, resale, and more. The relationships between these domains do not exist as a single feed anyone can subscribe to. They have to be composed, source by source, until a query about cotton sourced in West Africa returns the relevant weather signal, the relevant freight signal, the relevant tariff ruling, and the relevant retailer destocking signal in one structured response. Volume does not speed this up. The work is composing, not gathering. We call this Fashion Cross-Reasoning.
Retractions propagate as fast as signals.
Every record in Nari carries its own provenance chain. When a source revises or retracts, the change propagates downstream: every customer query that referenced that record sees the revision, with a new confidence score and a new date. Every answer carries the chain that produced it. That property only holds when the chain itself can be revised in flight. Without this discipline, the cost of a bad signal is permanent. A weak record becomes a sourcing call. A misdated entry becomes a wrong-quarter buy. A missed cross-reference becomes a missed shift in demand. Demo software's mistakes get a "we'll fix it next sprint." Infrastructure's mistakes get a wrong collection, a wrong campaign, a wrong shipment that has already left the port. Building propagation for retractions costs as much time as building propagation for signals. Not building it makes the layer rot silently.
The first wave of fashion AI gets the press cycle. Apshan will not. Operators pick infrastructure by who is still standing in five years with the same records, dated and scored the same way. Slow is the work. Reach us at hello@apshan.com.
Questions
What does reasoning-grade infrastructure mean in practice?
Reasoning-grade infrastructure means every record carries its source, its date, its confidence band, and its cross-references from the moment it enters the chain. Each of those properties must hold up under inspection by the model that queries it. Without all four, an answer built on the record will be brittle. With all four, the answer can be audited link by link.
How are chains of evidence built across unrelated domains?
One source at a time, by people who know the domain. Fashion sits downstream of weather, supply, customs, finance, and more, and each of those industries has its own data sources, formats, and cadences. The chain is built by indexing each feed, normalising its dating, scoring its confidence, and linking it to the records in adjacent domains that determine the outcome of any given question.
Why does retraction propagation matter for downstream decisions?
Because a decision made against a record that has since been corrected is a decision made against the wrong fact. Retraction propagation means every customer query that referenced a now-revised record sees the revision with a new confidence band and a new date. Without it, the chain looks intact while the answers it produces have already drifted from the truth.
What does it take to score a record's confidence?
A confidence band reflects how a record was captured, who captured it, when it was captured, and how it compares against parallel records covering the same event. A press release from a supplier is not scored the same as a court filing or a customs declaration. The scoring is manual at the edges where it matters and slow on purpose.
How is Fashion Cross-Reasoning different from a single-feed data subscription?
A single-feed subscription sells one source: sales, social signals, runway coverage. Fashion Cross-Reasoning is the discipline of composing across all the domains fashion is downstream of, with every record dated, scored, and pre-linked to the records in adjacent domains. The output is a chain that answers a question end to end, not a feed the customer has to assemble.