AQFP

INTERIM MODEL. Installed 2026-10-12 although it did NOT pass its own pre-set quality bar. Across the wide design range it serves, 4 of 200 functional calls on fresh designs failed the circuit test (2.0%, 90% CI 0.9 to 4.4%; the limit was 3% with an upper bound of 5%, and this part was met), against 10 of 205 (4.9%, 90% CI 2.9 to 8.0%) for the previous model on the same designs (it was not served there); it called functional 196 of 239 of the wide designs that pass (recall 0.82). On a second, independent fresh sample (801 new wide draws at new noise seeds, 2026-10-15) it got 6 of 204 functional calls wrong (2.9%, 90% CI 1.5 to 5.6%); the two samples together give 10 of 404 (2.5%, 90% CI 1.5 to 4.1%). What it missed: inside the previous, narrower range it was to do no worse than the previous model, and it missed that by one design on 1000 fresh in-range designs: 4 of 133 functional calls failed (3.0%, 90% CI 1.4 to 6.5%) against 4 of 134 for the previous model (3.0%, 90% CI 1.3 to 6.5%), recall 0.956 against 0.963; in that range the two are statistically tied. On the second sample (1000 new in-range draws) 2 of 123 failed (1.6%, 90% CI 0.5 to 4.8%); together 6 of 256 (2.3%, 90% CI 1.2 to 4.5%). The circuit test: the design's gate, fed by other gates and followed by four buffers built from the same cell, must carry a 30-symbol pattern covering all eight input rows, every bit with the right sign and at least 0.15 Ic, at nominal excitation AND with both supplies scaled to 90%, and it must do so at three noise seeds (3-seed label, used for training, calibration and test). The reference Blackburn cell is called functional (p 0.97). Scope: a circuit built only from this cell, a 5 GHz clock, 4.2 K, simulation not silicon, both supplies scaled together; this head serves Lx from 3.7 to 14.8 pH, drive from 0.6 to 1.6 x Ic, components inside the wide corpus sampling ranges and arm beta_L below 0.95, refuses anything outside them by name, and refuses designs whose chain gate latches below 0.35 Ic (outside its envelope). Above the operating point is untested (the Blackburn cell itself fails at +5%), so 'functional' does not imply upward margin. In realistic circuits built from this cell, changed one thing at a time (an 8-stage chain, a 2-way splitter built from the design's own cell, a 64-symbol pseudo-random pattern, five more noise seeds, a 0.5 GHz clock, both supplies at 90%) and with every change in one circuit, 8 of 100 of this model's functional calls on fresh wide-range designs failed at least one of them or the circuit test itself (90% CI 5 to 14%) and 11 of 100 on fresh designs in the previous range (7 to 17%); the circuit with every change at once failed 0 of 100 and 2 of 100. Same read as the circuit test (every bit right and at least 0.15 Ic). None of the 100 of these 200 calls with p at least 0.99 failed any of them. A 16-stage chain and a 3-way splitter, run separately, failed 0 and 1 of 100 wide (0 and 1 in range). Mixed neighbours are a known weakness: placing the design's gate between cells of a different (standard Blackburn) design failed 11 of 100 wide and 8 of 100 in range; the model was not trained on mixed neighbours. Functional answers now carry an interface flag (interface_to_standard_cells) from a physics margin: the static current the design's output delivers into a standard Blackburn cell, and that cell's onward link, relative to the thermal input requirement; 'weak' below 1.92. Among the functional calls it clears, the mixed circuit failed 2 of 256 (0.8%, 90% CI 0.3 to 2.3%); it flagged 27 of 29 mixed failures and 17 of 271 designs that worked. The threshold was calibrated retrospectively on 200 mixed-circuit outcomes and then tested on 300 fresh designs under a pre-registration. Simulation only; one mixed circuit and one noise seed per design. A 'weak' answer stays functional: the call is for a circuit built only from this cell. Measured in the aqfp_phase2e, aqfp_phase2f, aqfp_realistic_v2e and aqfp_mix_stage1 bench runs.

Confirmed in simulation by a pre-registered test (2026-10-18). On 735 fresh designs across the served range, 0 of 200 functional calls failed the circuit test (90% CI 0.0 to 1.3%; the pre-set limit was 3% with an upper bound of 5%), and 0 of 90 on fresh in-range designs (90% upper 2.9%). It called 200 of 245 working designs functional (recall 0.82, 90% CI 0.77 to 0.85). Of 259 fresh designs the served majority head calls functional, it called 0 non-working buffers (90% upper 1.0%). The reference Blackburn cell is called functional (p 0.999). The circuit test: the design's cell, driven by a gate of the same cell and followed by four more, must carry a 20-symbol pattern covering every 4-bit history, every bit with the right sign and at least 0.15 Ic, at nominal excitation and with both supplies at 90%, at three noise seeds. Scope: chains built only from this buffer cell, a 5 GHz clock, 4.2 K, simulation not silicon; Lx from 3.7 to 14.8 pH, drive from 0.6 to 1.6 x Ic, components inside the sampled ranges, arm beta_L below 0.95, kout held at -0.3, and a chain gate latching at or above 0.35 Ic; anything outside is refused by name. Known limits: above the operating point is untested (the Blackburn cell itself fails at +5%), so 'functional' does not imply upward margin. Placed between standard Blackburn cells (a standard cell driving it and a standard chain after it), 8 of 100 of its functional calls failed (4.6 to 13.6%, 90% CI). The interface flag (interface_to_standard_cells, 'weak' below 1.92) marked all 8 of those failures and 7 of 92 designs that worked; 0 of the 85 it cleared failed (90% upper 3.1%). One sample, one noise seed; the flag's threshold was set on the majority cell. Measured in the aqfp_cell_buffer bench run.