Appendix F: Prompt Lineage
The input of the model has three layers, which are kept apart throughout this thesis. The system prompt is the strategy text; it is the only layer that the prompt optimiser changes and the only one called a prompt version. The static floor map, generated by the engine, is appended once to the system prompt at the start of an episode. The per-turn observation, also generated by the engine, is the user message of each call (Section 5.2). The second and third layers are identical for every prompt version; a change to them is a change of the observation, not of the prompt.
Table F.1 lists the system prompts of the conditions reported in Chapter 6 with their role; the earlier hand-written prompt of Table 6.5, row 1, is stored in the run logs of that row. Prompts with a listing number are reproduced in Appendix A. The other evolved prompts are stored in the repository and in the result files of Appendix D.
Table F.1: System prompts. Characters as sent to the model.
| Prompt | Characters | Role | Used in |
|---|---|---|---|
| Single-floor prompt | 2,526 | Combat rules, how to read the observation, five prompt rules | Chapter 6, every single-floor result of the final configuration (Listing A.1) |
| Single-floor prompt at 60 % | 2,526 | The single-floor prompt with its HP limit (rule 3 of that prompt) at 60 % instead of 40 %; no other change | Section 6.1.2, tower settings on single floors |
| Tower prompt | 6,035 | As above, with scouting, shop rules and cross-floor reading | Chapter 6, final tower observation (Listing A.2) |
| Tower prompts of the first and second tower observation | 3,629 and 3,725 | Earlier versions of the tower prompt (Appendix A.2); the second is stored in each run log | Tables 6.9 and 6.13 |
| Rules-only prompt | 826 | Game rules, goal and answer format only | Prompt optimisation study, seed and reference (Listing A.3) |
| Evolved prompt, paired test | 3,152 | Rewrite by the optimiser for DeepSeek-V4-Flash (no thinking), bare observation | Section 6.5.2 (Listing A.4) |
| Other evolved prompts of the first study | 826 to 2,423 | Best prompt of each other run; five runs returned the seed unchanged | Table 6.6 |
| Evolved complex-floor prompt | 3,238 | Best prompt of the run that started from the single-floor prompt on the complex floor (Qwen3.6-35B) | Section 6.5.2 |
| Best prompts of the second study | 3,292 ((1+1) loop), 2,958 (Pareto loop) | Best of eight candidates each, DeepSeek-V4-Flash with thinking, bare observation | Table 6.8 |
| Evolved-route prompt | 2,682 | Best prompt of the (1+1) loop on the complex floor with the validator off; it names no position | Section 6.5.4 |
| Solver-route prompt (with survival rule) | 5,473 | Published solver's route as an allowlist, door order and phases, rule never to take a lethal fight | Compiled-knowledge condition (Listing A.5) |
| Exact-route prompt | 2,101 | The 21 decisions of the exact optimal route with expected HP | Section 6.6 (Listing A.6) |