DD-014
Conditional Attention and Contrarian Policies
Exact conditional-policy planner theorem, independently reproduced equilibrium census, and a preserved larger-class counterexample.
Completed finite studyChecked independently
The question
Does the value of deliberate non-use survive when agents can condition actions jointly on their private clue and the shared clue?
What we found
- DD-C-0066: In the complete deterministic agreement-respecting label-equivariant disagreement class for three labels, if p>=1/3 then replacing any third-option contrarian with a private-dominant role weakly improves discovery, strictly when p>1/3; consequently conditional policies do not improve the DD-012 unconditional planner optimum, and its zero-or-one public-dominant role classification survives.
- DD-C-0067: In DD-014's 75-cell rational census, the 775 anonymous conditional-policy profiles contain 115 weak and 53 strict equilibria; 23 cells have a positive gap between planner discovery and the best weak-equilibrium discovery, 49 cells have only planner-optimal weak equilibria, and none has a positive conditional planner gain over the unconditional class.
- DD-C-0068: In each of DD-014's four two-agent, two-label raw-policy audit cells, all sixteen deterministic policy tables permit discovery one through complementary constant policies, strictly above the best embedded unconditional private/public profile by between 1/16 and 1/9; hence the restricted three-label conditional result cannot be promoted to an unrestricted deterministic-policy claim.
What this result covers
Checked independently. The formal evidence wording and registry state remain available below.
Technical details
- Study phase
complete-bounded-study- Registry status
complete-bounded-study- Evidence status
one-verified-theorem-one-independently-reproduced-census-and-one-independently-reproduced-negative-result
Reproducible evidence
20260721T222047Z_DD-014_f5f099a8_ea0276dd16
Files and data
What comes next
preserve the run and scope boundary; complete PR merge, CI, Pages, and live-route validation before the focused paper