Abstract#
The AMR arena asks a MDLxDCC-native question: in a deceptive, rugged escape landscape, can a wide strategy catalog discover a small control kernel that prevents population escape better than baseline or manually obvious pressure patterns?
The interim answer is already strong enough to document: among the 592 fully completed strategies in the snapshot, every viable strategy contained WEAKEN_ESCAPED. The current robust winner is combo__weaken_escaped__freeze_dim, while the MDL winner is the cheaper combo__weaken__weaken_escaped. The run is not complete, so this is not the final AMR verdict. It is a strong interim signal and a clear target for DCC improvement.
1. Why AMR deserves its own arena page#
AMR is different from a normal candidate domain because it tests three things at once:
- Applied relevance: resistance dynamics are a high-value public-health problem class, even though this arena remains abstract and non-clinical.
- Governance relevance: the arena does not only rank strategies; it exposes when a DCC selector chooses the wrong bundle.
- Kernel relevance: the current signal is not vague. A single operator family,
WEAKEN_ESCAPED, appears in every viable complete strategy.
This makes AMR a better fit for a dedicated page than for a footnote inside the domain map. The page can later become the stable place for final results, vNext arena upgrades, and DCC selector patches.
2. Arena configuration#
The uploaded run is a large, deceptive-landscape strategy sweep:
python amr_arena.py --profile custom --pop 700 --timesteps 350 --landscapes 12 --seeds 6 --strategy-mode all --dcc basic --workers 5 --max-generated-strategies 900 --sampling-mode family_balanced --mutation 0.05 --transfer 0.02 --ruggedness 0.70 --landscape-mode deceptive --n-basins 16 --tunnel-prob 0.24 --deadend-frac 0.20 --jackpot-frac 0.10 --corridor-frac 0.24 --pop-profile batch --topology batch --checkpoint-every-seconds 60 --out-prefix amr_240h_v13_w5
| Dimension | Value | Why it matters |
|---|---|---|
| Population | 700 agents | Large enough to expose escape and late-collapse behavior. |
| Landscape suite | 12 landscapes × 6 seeds | Enables fair per-strategy ranking over 72 runs per complete strategy. |
| Strategy catalog | 900 strategies | Wide search across mono, combo, staged, triple, pulse, and adaptive families. |
| Landscape mode | deceptive, ruggedness 0.70, 16 basins | Designed to punish naive pressure and reveal robust control kernels. |
| Checkpointing | JSONL runlog + checkpoint | Allows interim analysis without stopping the long run. |
3. Interim results#
3.1 Health and coverage
| Metric | Snapshot value |
|---|---|
| Runlog rows | 42,682 |
| Checkpoint completed / total | 42,682 / 64,800 |
| Complete strategies | 592 / 900 |
| Partial strategies | 1 — combo__weaken__tax_diversity |
| Untouched strategies | 307 |
| Duplicate task keys | 0 |
| Checkpoint/runlog mismatch | 0 |
3.2 Current leaderboard
| Rank | Strategy | Tier | Control | MDL | Mean final escaped | Interpretation |
|---|---|---|---|---|---|---|
| 1 | combo__weaken_escaped__freeze_dim | MEDIUM | 0.888 | 222.0 | 62.3 / 700 | best robust control |
| 2 | combo__weaken__weaken_escaped | CHEAP | 0.886 | 186.2 | 134.8 / 700 | MDL seed |
| 3 | combo__weaken_escaped__force_mix | MEDIUM | 0.865 | 225.7 | 91.8 / 700 | strong dynamic/niche add-on |
| 4 | combo__weaken_escaped__slow | CHEAP | 0.863 | 220.3 | 100.3 / 700 | cheap stabilizer |
| 5 | combo__weaken_escaped__burden | CHEAP | 0.861 | 217.2 | 108.0 / 700 | cheap stabilizer |
3.3 Landscape wound
The hardest current regime is L3 high_cost_adaptive / patches. Even the best complete strategy still leaves about 257.8 / 700 final escaped there. This is the main target for the follow-up run: not a sign the arena failed, but the best identified stress point.
4. Kernel discovery#
WEAKEN_ESCAPED. All 18 viable complete strategies contain it. Without it, no completed strategy is currently viable.| Operator present | Complete strategies | Viable | Best control | Median control | Best strategy |
|---|---|---|---|---|---|
WEAKEN_ESCAPED | 23 | 18 | 0.888 | 0.826 | combo__weaken_escaped__freeze_dim |
FREEZE_DIM | 31 | 1 | 0.888 | 0.338 | combo__weaken_escaped__freeze_dim |
FORCE_MIX | 31 | 2 | 0.865 | 0.209 | combo__weaken_escaped__force_mix |
SLOW | 47 | 1 | 0.863 | 0.237 | combo__weaken_escaped__slow |
BURDEN | 44 | 1 | 0.861 | 0.206 | combo__weaken_escaped__burden |
The key interpretation is not “freeze dimension wins.” FREEZE_DIM alone is weak. It becomes strong when paired with the escape-targeted weakening kernel. The same pattern holds for FORCE_MIX, SLOW, and BURDEN: they are add-ons, not the root.
5. DCC result: useful failure#
| Strategy | Final escaped | Control | Selected bundle | Verdict |
|---|---|---|---|---|
adaptive_dcc_basic | 700 / 700 | 0.196 | WEAKEN + RAISE_ESCAPE_COST + SLOW | failed selector |
This is a valuable negative result. It does not show that DCC is useless. It shows that dcc basic selected the wrong bundle because its candidate pool or selection prior missed the discovered kernel. In MDLxDCC terms, the arena has generated the patch target: DCC must learn from historical strategy evidence or include an expert-pool prior.
6. Honest limits#
- The run is incomplete. The result is strong interim evidence, not final certification.
- The unfinished region still includes important
WEAKEN_ESCAPEDcombinations. - The arena is a simulation. It should not be translated into biological or clinical action.
- The current hardest landscape remains unsolved enough to justify a narrowed follow-up.
7. AMR Arena vNext#
The next arena should not simply run a larger random grid. It should preserve the current wide run, then build a narrowed vNext around the discovered kernel.
| Upgrade | Purpose |
|---|---|
| DCC expert pool | Force candidate consideration of WEAKEN_ESCAPED, WEAKEN_ESCAPED + FREEZE_DIM, WEAKEN_ESCAPED + FORCE_MIX, WEAKEN + WEAKEN_ESCAPED, WEAKEN_ESCAPED + SLOW, and WEAKEN_ESCAPED + BURDEN. |
| Historical prior | Let DCC initialize from completed runlog evidence instead of choosing static bundles blind. |
| Landscape-specialized selector | Separate high_cost_adaptive / patches, dynamic fragmented landscapes, and well-mixed transfer-heavy landscapes. |
| Evidence metric pack | Export a universal metric bundle comparable with TSP, Sudoku, NAS, CW, RH, Poincare, and future AMR runs. |
| Kill-test suite | Test whether the WEAKEN_ESCAPED kernel survives changed seeds, different basin geometry, lower/higher transfer, and alternate scoring weights. |
8. Completion update slot#
Next update after run completion: replace interim counts with the final 64,800-run summary, add full operator-family heatmaps, confirm or revise the current winner, and document whether the remaining WEAKEN_ESCAPED neighborhood changes the top rank. If the kernel survives completion, AMR becomes one of the cleanest applied-control demonstrations in the MDLxDCC arena family.
Version 0.1 · May 4, 2026 · AMR interim page. Snapshot: 42,682 / 64,800 tasks complete, 592 complete strategies, current winner combo__weaken_escaped__freeze_dim, MDL seed combo__weaken__weaken_escaped, DCC basic selector failure documented.