The word "current" in a policy file
An underwriter's working vocabulary is thick with deixis and rarely notices it. "The current accumulation in this zone." "Latest model version." "As of now, the treaty is unbroken." Each phrase points outward from the moment of utterance to something that must be looked up, not stated. Strip the moment and the sentence keeps its grammar and loses its truth value — Bar-Hillel's old problem, wearing a loss ratio.
This matters because underwriting is not really about hazard. It is about pricing a promise that has to hold across an interval nobody controls. The hazard curve is background. The deictic centre — where the underwriter is standing when they say "now" — is what turns a curve into a number on a slip.
What arrives
Four streams run continuously, at different rates, from different origos.
Claims flow arrives fast and messy: first notice of loss, reserve changes, litigation updates, subrogation recoveries, each timestamped to the reporting system's clock rather than the loss event's clock. A hurricane claim opened in September may still be moving in reserve eighteen months later.
Catastrophe models update slower and more consequentially. A vendor cat model revision — a new view of Gulf Coast storm surge, a re-estimated secondary uncertainty term for wildfire — can shift modelled loss for an entire book by double digits overnight, with no change to a single physical risk.
Exposure registries update continuously as policies bind, endorse, lapse and renew. A schedule of locations for a commercial property account is a moving document; a new warehouse added mid-term changes concentration in a CRESTA zone the model has already scored.
Reinsurance terms sit above all of this, resetting at treaty renewal on their own calendar, often annual, sometimes mid-term for reinstatement provisions. The retention, the reinstatement premium, the hours clause — all fixed at a moment, all silently assumed to still hold.
Four clocks. Four origos. None of them is "the market" in any single sense; each is a partial, dated view of it.
What is held
A Large Language Model, trained on underwriting manuals, historical bordereaux extracts and cat-model documentation frozen at some cutoff, holds none of this as a stream. It holds sentences that once pointed at something. "The current view of hurricane frequency" appears in its training data as a string, detached from the model version, the vendor, the year, the file that produced it. Asked to reason about "the current hazard curve for this territory", the system has no choice but to infer a plausible origo from the statistics of its corpus — typically an averaged, dated view that happens to sit near the training cutoff, presented with the same confidence as a live figure.
A Large World Model, if one were built for a single renewal — ingesting the submission, the current schedule, the current cat model run, the current treaty terms for one bound scene — recovers a working origo. Within that scene, "current retention" and "this location" resolve correctly. But the scene closes at bind. Six months later, when a wildfire redraws the wildland-urban interface the schedule was priced against, there is no mechanism inside that bounded model for the anchor to move. The words in the file still say "current"; the world that made them true has gone.
A Large Universe Model, as argued elsewhere on this axis, keeps the origo running: claims flow, cat model versions, exposure registry, treaty terms, each stream tagged with its own arrival time, each claim in the belief base carrying provenance — which stream, which instant, which vendor version fixed this number as "current". Not a single now. A frontier of nows, ordered and bounded, which is the honest version of what "now" means to a book with four moving parts.
What triggers revision
The characteristic failure sits exactly here. A book is priced on a hazard curve. The wording on the slip says "based on current modelled view of hurricane frequency in the Gulf." Two seasons pass. The vendor issues a model update reflecting revised secondary uncertainty for storm surge and a materially higher frequency assumption for major hurricanes making landfall in warm-anomaly years — a change grounded in the last two seasons' actual behaviour, not in the two seasons the original curve was fit against. The word "current" in the file has not changed. Its referent has moved twice: once when the model updated, once when two seasons of loss experience made the old model wrong before the update even landed.
Nothing in a frozen file can detect that its own deictic anchor has been pulled out from under it. The sentence still parses. The number still looks like a number. Revision has to be triggered from outside — by someone noticing that a stream has moved.
A running system with provenance can trigger on the stream itself: a new cat model version arriving is an event, timestamped, that invalidates every belief downstream of "current modelled view" for that peril and territory, and flags every bound policy whose price depended on the superseded version. The revision is not a periodic re-underwriting exercise. It is a standing consequence of the streams being live.
What the operator sees
The underwriter, in either regime, sees a screen. The difference is what the screen is willing to assert.
In the frozen-corpus regime, the screen shows a hazard curve with a date buried in a footnote, if anywhere, and the underwriter must independently remember to check whether that date is stale relative to the vendor's release calendar and relative to the last two loss seasons. This is exactly the failure mode: the check is manual, occasional, and easy to skip under renewal-season volume, when forty accounts are due by Friday.
In a system built on a running, provenanced origo, the screen shows something closer to: "modelled frequency, v14.2, released March, superseded — v14.3 available, reflects 2023–24 landfall experience, applied here." The underwriter still makes the pricing decision. What changes is that the deictic content of "current" is made explicit and datable rather than assumed, and disagreement between the treaty wording's "current" and the model's "current" becomes visible before bind rather than discoverable at claim.
Give the underwriter a system prompt with today's date and a search tool pointed at the vendor's model repository, and the anchor problem disappears. That is cheaper than maintaining live belief state over four streams, and it is auditable — you can see exactly what string was injected.
This is worth taking seriously, because prompt injection of a stated origo genuinely works for a large share of cases, and underwriting workflows already lean on exactly this pattern — a system prompt naming the as-of date, a retrieval call to the latest model documentation. But someone still has to decide what counts as "current" for the purpose of the query: current as of bind date, as of the last full model run, as of the reinsurance treaty's own effective date, which may not match any of those. Injecting a date answers when is now; it does not answer which stream's now applies to which clause, and a book with four asynchronous streams has exactly that problem at its centre. Relocating the decision to whoever writes the prompt is not the same as resolving it, and it puts the decision one step further from the record than provenance would.
What it costs
Maintaining a running, provenanced origo across claims flow, cat models, exposure registries and treaty terms is not free, and the honest accounting has to admit the second objection has force: most of what an underwriter relies on is not deictic at all. The mathematics of a copula, the structure of a quota share, the case law on "occurrence" versus "claims made" — these are stable, well served by static reference material, and do not need a running clock.
But the moment stable knowledge is applied to a book — this territory, this schedule, as of this renewal — the application is deictic even where the underlying knowledge is not. The failure that sinks books is almost never a misunderstanding of copula mathematics. It is a correct method applied to a stale anchor: the right formula, fed the wrong "current".
The frontier, not the moment
There is a version of the objection that survives every answer above: a book spanning four streams with four clocks does not have one now, so speaking of "the" origo is itself a fiction. That is correct, and it is the right place to end. A distributed underwriting system does not get a single instant called now. It gets a frontier — claims flow arriving at one rate, cat model versions at another, exposure changes continuously, treaty terms annually — each stream's most recent update forming a bound on how current any downstream belief can be. Provenance does not manufacture a single moment. It records the frontier explicitly, so that "current hazard curve" can be read as "as fixed by v14.2, released March, not yet superseded" rather than as an unmarked claim to timelessness. That is not less deictic than the corpus. It is deixis done honestly, with the pointing finger still attached to the hand that moved it.