Explore-and-Select Loop — Research Results (2026-07-02)

Source: backlog/260702_explore_select_research.md

Explore-and-Select Loop — Research Results (2026-07-02)

Appendix to 260702_loop_engineering_backlog.md §2 (the plateau finding). Five research

lanes returned; this doc records the proposed rule, the design skeleton, and the source

map. Lanes: ChatGPT 5.5 Pro (session continuation), 3× gpt-5.3-codex-spark literature

sweeps, LazyCodex/OmO-harnessed gpt-5.5 (team-topology question).

1. Proposed rule (ChatGPT Pro) — LOOP-EXPLORE-SELECT-01

Candidate rule text for a future dev-pabcd revision (currently referenced from §11.4 as

a pointer; full adoption pending a normal PABCD pass):

DEFAULT (LOOP-EXPLORE-SELECT-01):
When LOOP-ARCHETYPE-01 classifies a work-phase as Open-ended optimization
and PLATEAU fires, do not run another same-class C→B repair. Start a new
Explore-and-Select PABCD work-phase.

P defines: strategic descriptors, telemetry, candidate budget, archive,
opponent portfolio, evaluator cascade, selection rule, and terminal states.
A audits the loop design and verifier, not only the code plan.
B generates a bounded population of StrategyCard+patch candidates, evaluates
them on paired instances, admits them to a strategic-diversity archive, and
promotes only robust finalists.
C revalidates the selected best-so-far against the fixed verifier, regression
sentinels, timing, and allowed-surface constraints.
D reports DONE only if the objective threshold is met; otherwise report
BUDGET_EXHAUSTED with best-so-far, archive evidence, loss signatures, and the
next exploration direction.

Key design elements (full pseudocode in the ChatGPT session,

https://chatgpt.com/c/6a466d08-a268-83ee-a628-428205a8bd36):

• Unit of evolution = StrategyCard + bounded patch, never "threshold ±5". The

candidate carries archetype, opponent model, strategic hypothesis, expected replay

signature, predicted vs measured behavior descriptors, and a death log entry.

• Two-layer diversity: idea diversity (StrategyCards, EoH-style thoughts) ×

behavior diversity (replay-measured descriptors). A candidate whose measured behavior

contradicts its claimed strategy is re-archived by measurement or killed.

• MAP-Elites-style archive keyed by strategic descriptors (pressure_timing,

economy_posture, map_control, primary_win_plan, opponent_model, map_size_plan) — not

one leaderboard.

• PSRO-lite opponent portfolio: fixed opponents + incumbent + 2-4 frozen elites +

1-3 hand-written exploiters/probes stressing one weakness each.

• Hyperband-style evaluation cascade: stage0 validity → stage1 cheap target slice

(+ rush-regression sentinels) → stage2 archive admission → stage3 full tournament for

top-K only. Novelty is an exploration criterion, never a deployment criterion.

• Plateau = 2+ of: score_plateau, archive_plateau, loss_signature_plateau,

death_class_plateau (native §10), anchor_plateau (native §10), budget_plateau.

• Restart ladder: soft (sample from underfilled cells) → strategic (regenerate from

replay telemetry) → island (reset worst island) → PSRO (change opponent mix / add

exploiter) → evaluator (add telemetry — mandatory first step for NEXT NATION, whose

verifier only says "losing") → PABCD (back to P/Interview).

2. Team-topology verdict (LazyCodex lane, gpt-5.5)

Committee (co-writing, consensus) does NOT attack the plateau — it accelerates

convergence around the incumbent concept; population/league DOES. Division of labor:

committee for review/postmortem/invariants, population for strategy generation, **solo

builder for final integration** (the author's B-phase distrust survives). Diversity must

be mechanical: archetype-pinned prompts with *different objectives per archetype*,

isolated contexts/worktrees, behavior-descriptor novelty gates, deterministic promotion

("do not let agents vote"). Supporting evidence: multi-agent debate does not reliably

beat single agents while model heterogeneity helps (https://arxiv.org/abs/2502.08788);

LLM conformity to majority answers (https://arxiv.org/abs/2410.12428).

3. Literature map (spark lanes, condensed)

LLM-guided evolutionary/program search

• FunSearch — program evolution + external evaluator, island diversity:

https://www.nature.com/articles/s41586-023-06924-6

• AlphaEvolve — evolvable code blocks, elite archive:

https://arxiv.org/abs/2506.13131

• EoH — evolve NL "thoughts" + code together: https://arxiv.org/abs/2401.02051
• EoH-S — complementary heuristic *set* (portfolio vs one generic policy):

https://arxiv.org/abs/2508.03082

• MAP-Elites — quality-diversity archive: https://arxiv.org/abs/1504.04909
• Novelty search (abandoning objectives): https://doi.org/10.1162/EVCO_a_00025

Self-play / population / opponent modeling

• PSRO original (unified game-theoretic MARL): https://arxiv.org/abs/1711.00832
• PSRO survey (strategy exploration as core difficulty):

https://arxiv.org/abs/2403.02227

• AlphaStar league + exploiters: https://www.nature.com/articles/s41586-019-1724-z
• TStarBot-X — league training at low compute: https://arxiv.org/abs/2011.13729
• PBT (population-based training): https://arxiv.org/abs/1711.09846
• Meta-game evaluation (seed/map cross-play + bootstrap — directly addresses the

5-seed × 3-map selection bias): https://arxiv.org/abs/2405.00243

Plateau detection / restart & exploration scheduling

• Hyperband (successive-halving budgets): https://arxiv.org/abs/1603.06560
• Stagnation detection in EAs (raise step size on non-improvement):

https://arxiv.org/abs/2004.03266

• Bandit multi-start control for local search: https://arxiv.org/abs/1401.3894
• GP-Hedge portfolio allocation: https://arxiv.org/abs/1009.5419
• MAB restart/reset in CDCL SAT: https://arxiv.org/abs/2404.03753 (+ cold restarts:

https://arxiv.org/abs/2404.16387)

• ME-MAP-Elites (bandit over emitters — widen vs deepen decision):

https://arxiv.org/abs/2007.05352

• Dominated Novelty Search: https://arxiv.org/abs/2502.00593
• Reflexion (+ known limits): https://arxiv.org/abs/2303.11366 · Self-Refine:

https://arxiv.org/abs/2303.17651 · Agent-R (loop-break recovery):

https://arxiv.org/abs/2501.11425

Verification caveats

• Two spark-sourced future-dated arXiv IDs are UNVERIFIED — spot-check before citing:

https://arxiv.org/abs/2605.28273 (Global PSRO / population exploitability),

https://arxiv.org/abs/2602.16928 (LLM-discovered MARL algorithms).

• Spark lane 1 occasionally framed the case as the game of Go rather than a Go-language

bot; method mappings remain valid, wording does not.

4. Operational counters (lane-3 synthesis)

Keep two counters per candidate lineage: no_progress_at_gate and

same_reason_repeats. Reward depth-first continuation while improving; enforce

controlled diversification when failure signatures converge. This is the quantitative

backing for §10 LOOP-PHASE-DEATH-01's N=3 starting value.

5. Single-model divergence smoke test (2026-07-02, gpt-5.5 ×17 parallel)

Question: with ONE model (no model-heterogeneity axis), which divergence schema

actually prevents mode collapse? Task: propose a strategy card (4 behavior axes:

FIRST_ATTACK, ECONOMY, DENIAL_TARGET, WIN_PLAN) for the NEXT NATION anti-economy

plateau. 5 conditions, same model (gpt-5.5, effort medium), no web search.

Condition	Calls	Unique 4-axis tuples	Notes
A. Baseline (identical prompt)	4	2/4	Total collapse: all early/balanced/eco-denial "contain" family; names Contain/Clamp ×4; one mechanism family
B. Cell assignment (Zwicky/MAP)	4	4/4 (forced)	Guaranteed coverage by construction; in-cell quality held even for the odd cell (turtle+hq-snipe → "Tripwire Turtle Snipe")
C. Objective splitting	4	4/4	Spontaneous rush/greedy/mid/hq-snipe coverage; 1 of 4 still fell into the baseline attractor family
D. Verbalized Sampling (1 call, 4 cards + probabilities)	1	4/4	Widest spontaneous coverage — only condition to produce FIRST_ATTACK=none, ECONOMY=turtle, WIN_PLAN=attrition unforced; self-estimated P for the typical answer (0.27-0.42) matched the observed baseline collapse
E. Evidence slicing	4	3/4	ECONOMY collapsed to balanced ×4, but highest groundedness — each card directly answers its evidence; the diagnosis lane, not the breadth lane

Attractor check: the baseline mode ("early balanced contain/eco-denial") reappeared

in C (1/4) and D (P=.27 card) — VS's verbalized probabilities effectively *named* the

mode the baseline collapsed into.

Adopted 1-model recipe (DIVERGE-SPEC draft):

1. VS first — one call, k cards with probabilities, demand low-P tails (cheapest breadth scan).
2. Cell-assign the real candidate batch from undercovered axis cells (guaranteed coverage).
3. Split objectives per cell (optimization pressure per niche; also generates exploiter probes).
4. Evidence slices once real telemetry exists (grounded refinement, not breadth).
5. Post-hoc behavior-distance novelty gate (self-reported tags are gameable — final gate must use measured descriptors).

Caveats: n=4/condition, single run, self-reported axis tags, diversity ≠ quality

(the tournament verifier still owns selection). Smoke-level evidence only.