AQFP

Every served AQFP cell: knobs with units, defaults and served ranges, the drive, beta_arm and envelope limits, the Mondrian thresholds, model and feature versions, bar status and caveats (read from the heads, not re-typed), the margin refusals, the interface rule, the topology guard and the circuit test behind the label.

The iface-1 interface margins and flag for a design on their own: whether its output can drive a standard Blackburn cell (output margin, the rule), the input margin and the four hops, the threshold, and the head's call beside them. For a functional design the flag fields equal the ones /api/aqfp/predict serves; otherwise served_with_verdict is false and the flag is static physics with no measured record. A design the head refuses is refused with the same reason.

The exact circuit-test decks behind a design's verdict: label v2's deck for the majority, the buffer chain deck (BUF-L1) for the buffer, at nominal excitation and with both supplies at 90%. Each leg carries deck (the label's deck byte for byte), josim_deck (the same plus .print lines for every gate's LQ current) and the label's time grid; the response adds the pattern, the expected bits, the read stage, the read times and the read rule in words, so the verdict can be reproduced. The head's own call is given beside it (prediction). A design the head refuses is refused here with the same reason (200, refused true, no decks); malformed inputs are a 400 as on /api/aqfp/predict.

Functional verdict for one AQFP cell on the MIT-LL SFQ5ee AQFP process.

Returns the verdict, its conformal set, and an EXPLICIT refusal of every margin quantity with the reason attached. The served head is an INTERIM model (installed 2026-10-12, did not pass its own pre-set quality bar; model_status and functional_caveat say so with its measured record). It serves the phase 2e wide design range and refuses designs outside it, and designs below its envelope, by name. functional is null when the conformal set is not a single class: that is an escalation, not a missing value, and escalate says so. The set is class-conditional (Mondrian): functional iff p_functional >= 1 - q_pos, not functional iff p_functional <= q_neg. With the served thresholds it can never hold both classes, so an escalation is a set holding neither class (p_functional between q_neg and 1 - q_pos). Coverage is per class, measured on fresh designs; conformal.coverage_scope gives the figures. A functional answer also carries interface_to_standard_cells ('weak' or 'ok') with its margins, threshold and a sentence: whether the design's output can drive a standard Blackburn cell (iface-1, a static physics margin; the call and p are unchanged by it).

cell "buffer" (default "majority") is answered by the buffer head (fluxus.sfq.aqfp.buffer_head, installed 2026-10-18 after its pre-registered bar passed): same knobs and response shape, cell_type "aqfp_buffer", its own envelope, ranges, features and caveat; majority-only inputs are refused by name.

Predict many AQFP designs in one call.

results is positionally aligned with items, and each entry is exactly what POST /api/aqfp/predict returns for that design (its 200 body, or the {"error": ...} object its 400 carries), so a malformed value, an unknown cell or field, or a majority-only input on the buffer is an error in its own slot. A value that is not a finite number, or not a number at all, is refused for the whole request by the schema (422 naming items[i].params.knob). Over 2000 items the request is TRUNCATED with a clamped block; a deadline at the request budget stops the rest with a budget block. summary counts the calls and refusal kinds.

Whether the AQFP heads can serve, and what they were fitted on. The top level is the default (majority) head; cells lists both served cells.

Sweep one or two knobs (any AQFPCellSpec field, or iin_ua) linearly from start to stop in n points around the base design, and return the call, p_functional and refusal (kind and reason) of every point, row-major (the last axis varies fastest), with a summary. n defaults to 21 and is clamped to 2..201 for one axis and 2..41 per axis for two, with a clamped block. A base design that is wrong everywhere (an unknown field, a majority-only input on the buffer, a malformed value not on a swept axis) is one 400, not a grid of errors.