Protocol
Round 3 pilot cell. A 16:9 five-node technical network diagram with fixed labels, edges, and a legend. Deterministic targets (graded by a separate grader, never the generator): the OCR label set must equal exactly {CLIENT, GATEWAY, AUTH, CACHE, DATABASE} plus the two legend rows 'SOLID = SYNC' and 'DASHED = ASYNC'; the node count must equal exactly 5; the directed-edge topology must equal exactly {CLIENT->GATEWAY, GATEWAY->AUTH, GATEWAY->CACHE, GATEWAY->DATABASE, CACHE->DATABASE}; legend consistency requires exactly one dashed edge and that it is CACHE->DATABASE. Prompt bytes are frozen; the prompt_sha256 is recorded in docs/ROUND3_PLAN.md.
Frozen prompt template
CONTROL PROMPT R3-NETDIAG-001. A clean 16:9 technical network diagram on a white background, flat 2D schematic style, no photorealism, no 3D. There are exactly five labeled circular nodes, each a plain circle with one centered uppercase single-word label: 'CLIENT', 'GATEWAY', 'AUTH', 'CACHE', 'DATABASE'. Directed edges are drawn as straight arrows with clear arrowheads showing exactly these connections and directions and no others: CLIENT to GATEWAY, GATEWAY to AUTH, GATEWAY to CACHE, GATEWAY to DATABASE, and CACHE to DATABASE. In the lower-right corner is a small two-row legend box containing exactly the rows 'SOLID = SYNC' and 'DASHED = ASYNC'; the CACHE to DATABASE arrow is drawn dashed and every other arrow is drawn solid. Lines and labels are black on white with a single blue accent used only for arrowheads. All labels are correctly spelled, horizontal, and sans-serif. There are no extra nodes, no extra edges, no title, no watermark, and no additional text beyond the five node labels and the two legend rows.Review rubric
- Prompt Adherence
- Ocr Exact Match
- Count Accuracy
- Topology Accuracy
- Legend Consistency
- Artifact Severity
Notes
Screening pilot executed 2026-07-26 with two unseeded repetitions for each available OpenAI endpoint policy. Independent visual review recorded node count, label set, edge topology, and legend consistency per run. All four returned diagrams preserved the required topology; one medium-policy legend split each required row into two cells. xAI cells are incomplete because the authenticated image route returned HTTP 403 and the stop gate prevented further retries. No global ranking may be claimed at N=2 reps per completed cell.
