From Q&A to COA: where a banded answer does — and doesn’t — reach the verdict

A fair question came up mid-build: is the system actually parsing the banded Q&A information when it scores courses of action? I don’t see how it extracts the relevant impact from the items. It is exactly the right thing to be suspicious of. A demonstrator that claimed every intelligence answer feeds every verdict would be lying, because it doesn’t — and it shouldn’t. So we traced three real Meridian knowledge items from their banded answer all the way to a Commander’s verdict, and instrumented every step (tests/k3-trace.test.ts, ten tests). Three items, three different fates.

The connection from a knowledge item to a COA verdict is spatially mediated — not a direct input→output function. A banded answer is compiled into a channel (a threat surface, a civil-density surface) over a named region of the map. A commitment’s metric then walks a force element’s route and, at each leg, may or may not read that channel. Whether an answer reaches a verdict depends on which metric asks and where the route goes. That produces exactly three patterns.

See all three at once

Below is the taxonomy as a live three-panel embed. Each panel takes one item from its banded answer, through provenanceChip and the real channelTrace into the compiled world, to the verdict the shipped scorer produced. The band pills, provenance chips, channel trace, and the four-stop matrix are the demonstrator’s own components; the numbers are real Meridian fixtures. Drag Panel 2’s slider — each frame is a genuine re-score of P1 × C4 with K6 widened.

K3 — the honest gap

K3 asks “what is the civil population in the Port Halcyon district?” Its answer — an 18-month-old NGO census — is [35,000–55,000] persons, and it compiles cleanly: the civil_density channel carries K3’s band as a region override, present at exactly the cells FE-FALCON passes through. And yet C3 — “no fires into the populated port district” — never reads it. C3’s fires metric is a purely geometric test: it counts how many of FE-FALCON’s route legs fall inside the district. Score P2 against C3 with K3 and without it and you get the identical answer: fires = [1, 1], violated. The population band has zero effect on the verdict.

That is not a bug — it is the discipline holding. C3 is about the decision to fire, not the scale of harm. The population number is the J-2’s evidence for why the district matters; it is not a multiplier on a binary commitment. A metric that multiplied leg-count by population would invent a quantity — weighted civil harm — that no commitment asked for, and that no source supports. ASSAY refuses to conjure numbers (ASSAY-DEC-19). The gap between “the band is present” and “the band is read” is where honesty lives, not where it fails.

K6 — a band that does propagate

K6 asks “what sortie rate can the FAC squadron sustain?” — assessed, low confidence, a deliberately wide [2–6] sorties/day from weak evidence. Here the metric does ask. C4 — “amphibious group not exposed to unsuppressed battery or FAC threat” — uses exposure, which walks FE-ANVIL’s route and reads the threat channel at every leg. In P1 the route enters fac_waters, so K6’s band propagates into the exposure sum [12, 36] band-hours, and the C4 margin is [−24, 0] — tight.

Now the payoff, and the reason the slider exists. Widen K6 from [2, 6] to [1, 8] — less certain intelligence — and the C4 margin widens with it, from [−24, 0] to [−36, 6]. The wider margin strictly contains the narrower one. That is propagation honesty (invariant G6): uncertainty in, uncertainty out — never laundered away, never invented. Drag the slider and watch the margin band breathe in step with the assessment. K3’s band has no such lever on C3, and that difference is the whole point.

K7 — compiled, but the route misses

K7 asks “what is the air-defence engagement envelope at Carrick strip?” — [8–14] km, assessed. It compiles into the same threat channel C4 reads. Unlike the fires metric, C4’s exposure would read it. But FE-ANVIL’s route in both P1 and P2 never enters the air_defence region — so K7 has zero effect on C4. This is neither the geometric gap nor a defect: it is a route-dependent miss. K7’s real role is upstream, at plan generation (generate.ts): it shapes whether FALCON is routed to stand off Carrick or to enter the envelope in the first place. K7 affects which plans exist, not how a given plan is scored. A plan whose route did cross air_defence would see K7’s band propagate into C4 exactly as K6’s does.

What a reader takes away

Three items, three fates: a band that propagates into the verdict and widens it as it widens (K6); a band that is present in the world but that a geometric metric is right not to read (K3); and a band that a metric would read but whose region the route never touches (K7). Understanding these three is understanding how banded intelligence feeds — and sometimes deliberately does not feed — COA evaluation. The system is not ignoring your Q&A; it is being precise about where each answer legitimately bears.