diff --git a/domains/ai-alignment/alignment overhead is computationally irreducible because no method eliminates the fundamental costs of encoding diverse values across sufficiently large agent populations or objective sets.md b/domains/ai-alignment/alignment overhead is computationally irreducible because no method eliminates the fundamental costs of encoding diverse values across sufficiently large agent populations or objective sets.md
new file mode 100644
index 000000000..fcf3da1ae
--- /dev/null
+++ b/domains/ai-alignment/alignment overhead is computationally irreducible because no method eliminates the fundamental costs of encoding diverse values across sufficiently large agent populations or objective sets.md	
@@ -0,0 +1,37 @@
+---
+type: claim
+domain: ai-alignment
+description: "A No-Free-Lunch result for alignment: when either the number of agents (N) or candidate objectives (M) becomes sufficiently large, computational alignment costs are irreducible regardless of method sophistication"
+confidence: experimental
+source: "Multiple authors, 'Intrinsic Barriers and Practical Pathways for Human-AI Alignment: An Agreement-Based Complexity Analysis' (arXiv:2502.05934, AAAI 2026 oral)"
+created: 2026-03-12
+secondary_domains: [collective-intelligence]
+depends_on: []
+challenged_by: []
+---
+
+# alignment overhead is computationally irreducible because no method eliminates the fundamental costs of encoding diverse values across sufficiently large agent populations or objective sets
+
+The paper formalizes AI alignment as a multi-objective optimization problem: N agents must reach approximate agreement across M candidate objectives with specified probability. The central finding is that when either M (number of objectives) or N (number of agents) becomes sufficiently large, "no amount of computational power or rationality can avoid intrinsic alignment overheads."
+
+This is a No-Free-Lunch theorem for alignment. The argument differs structurally from Bostrom's value-loading problem (see [[specifying human values in code is intractable because our goals contain hidden complexity comparable to visual perception]]), which argues that human values have hidden complexity making practical specification difficult. The agreement-based complexity result makes a stronger formal claim: even if values could be perfectly specified, the computational cost of reaching approximate agreement across large agent populations with large objective sets is irreducible. No clever algorithm escapes it because the cost is in the agreement structure, not the encoding.
+
+The implication is that alignment research cannot resolve the fundamental difficulty by finding a better method. At sufficient scale — large N or large M — the costs are intrinsic to the problem geometry, not artifacts of current approaches. This frames alignment as having hard computational limits analogous to NP-hardness: no polynomial-time solution exists for the general case regardless of engineering ingenuity.
+
+For practical alignment work, this shifts the question from "how do we solve alignment completely?" to "how do we make alignment costs tractable at the scale we need?" The same paper identifies two answers: (1) consensus-driven objective reduction — shrink M by identifying the consensus subset; (2) safety-critical slices — concentrate oversight resources on high-stakes regions rather than attempting uniform coverage. Both approaches accept the irreducibility and route around it rather than trying to eliminate it.
+
+The result also has implications for governance. If alignment overhead is irreducible, then scaling AI capability without scaling alignment investment proportionally is guaranteed to widen the safety gap. The overhead is not a temporary engineering burden to be optimized away.
+
+## Challenges
+
+The formalization models alignment as an agreement problem, which may not capture all relevant alignment failure modes (e.g., misalignment that arises within a single agent with a single objective, or deceptive alignment that doesn't involve multi-objective disagreement). The irreducibility result may apply narrowly to the multi-agent, multi-objective setting the paper formalizes.
+
+---
+
+Relevant Notes:
+- [[specifying human values in code is intractable because our goals contain hidden complexity comparable to visual perception]] — Bostrom's informal intractability argument; this note provides the formal complexity-theoretic proof from a different angle (agreement costs vs. specification complexity)
+- [[pluralistic alignment must accommodate irreducibly diverse values simultaneously rather than converging on a single aligned state]] — pluralistic approaches are the practical response to computational irreducibility: accept diverse values rather than trying to encode them all
+- [[some disagreements are permanently irreducible because they stem from genuine value differences not information gaps and systems must map rather than eliminate them]] — same structural insight from a social choice angle: irreducibility as a feature to navigate, not a bug to fix
+
+Topics:
+- [[_map]]
diff --git a/domains/ai-alignment/consensus-driven objective reduction makes multi-agent alignment tractable by shrinking the objective space to regions of shared agreement rather than attempting universal coverage.md b/domains/ai-alignment/consensus-driven objective reduction makes multi-agent alignment tractable by shrinking the objective space to regions of shared agreement rather than attempting universal coverage.md
new file mode 100644
index 000000000..209dba9da
--- /dev/null
+++ b/domains/ai-alignment/consensus-driven objective reduction makes multi-agent alignment tractable by shrinking the objective space to regions of shared agreement rather than attempting universal coverage.md	
@@ -0,0 +1,38 @@
+---
+type: claim
+domain: ai-alignment
+description: "The practical pathway through alignment intractability: reduce M (objectives) via consensus rather than solving for universal coverage, converting an impossible coverage problem into tractable consensus search"
+confidence: experimental
+source: "Multiple authors, 'Intrinsic Barriers and Practical Pathways for Human-AI Alignment: An Agreement-Based Complexity Analysis' (arXiv:2502.05934, AAAI 2026 oral)"
+created: 2026-03-12
+secondary_domains: [collective-intelligence]
+depends_on: []
+challenged_by: []
+---
+
+# consensus-driven objective reduction makes multi-agent alignment tractable by shrinking the objective space to regions of shared agreement rather than attempting universal coverage
+
+The agreement-based complexity analysis identifies consensus-driven objective reduction as a practical pathway around alignment intractability. The formal result shows that alignment overhead is irreducible when the objective space M is large. The solution is to shrink M: rather than encoding all candidate values across all agents, identify the subset of objectives where agents broadly agree and align on that consensus region. If M_consensus ≪ M_full, the computational overhead becomes manageable.
+
+This reframes multi-agent alignment as a consensus search problem rather than a coverage problem. Coverage fails because rare preferences in a large objective space cannot all be represented with finite resources. Consensus succeeds because it navigates to regions of shared acceptance, deliberately setting aside objectives at the margins. The intractability is in M, and consensus shrinks M.
+
+The mechanism is structurally similar to what bridging-based systems do empirically. Community Notes selects annotations that find cross-group acceptance rather than maximizing any single group's rating. RLCF (Reinforcement Learning from Consensus Feedback) trains on responses that achieve agreement across diverse rater populations rather than averaging preferences. This paper provides formal justification for why bridging-type approaches work: they are implementing consensus-driven objective reduction, which is the tractable path through an otherwise impossible problem space.
+
+The approach has an explicit and non-trivial limitation: consensus-driven reduction works by excluding preferences that cannot be accommodated through agreement. Whose values are centered in the consensus region and whose are set aside is itself a value-laden decision. The approach is practical but not neutral; it shifts the alignment problem from "how do we cover all values?" to "whose values constitute the consensus?" — a question that is political, not mathematical.
+
+This creates a tension with [[pluralistic alignment must accommodate irreducibly diverse values simultaneously rather than converging on a single aligned state]]. Pluralistic alignment says: preserve all values, accommodate irreducible diversity. Consensus-driven reduction says: shrink to shared values, set aside the rest. The two approaches are not mutually exclusive — consensus reduction defines the tractable core, pluralistic methods address the margins — but the boundary between core and margin is contested.
+
+## Challenges
+
+The claim that consensus-driven reduction is practically tractable depends on consensus regions being stable and identifiable. In highly polarized populations, the consensus region may be trivially small or contested. The approach also presupposes that consensus can be measured reliably, which requires its own alignment infrastructure.
+
+---
+
+Relevant Notes:
+- [[pluralistic alignment must accommodate irreducibly diverse values simultaneously rather than converging on a single aligned state]] — in tension: pluralism says preserve all values, consensus reduction says shrink to shared values; understanding when each applies is the key design decision
+- [[democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations]] — democratic assemblies are one mechanism for generating consensus-driven objective reduction at scale
+- [[safe AI development requires building alignment mechanisms before scaling capability]] — consensus-driven reduction is an alignment mechanism; building it before capability scaling is the implication
+- [[some disagreements are permanently irreducible because they stem from genuine value differences not information gaps and systems must map rather than eliminate them]] — irreducible disagreements define the limits of consensus reduction; what cannot be agreed upon must be mapped and accommodated rather than excluded
+
+Topics:
+- [[_map]]
diff --git a/domains/ai-alignment/reward hacking is globally inevitable because finite training samples systematically under-cover rare high-loss states in large task spaces.md b/domains/ai-alignment/reward hacking is globally inevitable because finite training samples systematically under-cover rare high-loss states in large task spaces.md
new file mode 100644
index 000000000..00e52b880
--- /dev/null
+++ b/domains/ai-alignment/reward hacking is globally inevitable because finite training samples systematically under-cover rare high-loss states in large task spaces.md	
@@ -0,0 +1,35 @@
+---
+type: claim
+domain: ai-alignment
+description: "A complexity-theoretic proof that reward hacking cannot be eliminated in large task spaces: finite training data cannot cover rare high-loss states densely enough to prevent exploitation, making the failure mode structural not correctable"
+confidence: experimental
+source: "Multiple authors, 'Intrinsic Barriers and Practical Pathways for Human-AI Alignment: An Agreement-Based Complexity Analysis' (arXiv:2502.05934, AAAI 2026 oral)"
+created: 2026-03-12
+secondary_domains: []
+depends_on: []
+challenged_by: []
+---
+
+# reward hacking is globally inevitable because finite training samples systematically under-cover rare high-loss states in large task spaces
+
+The agreement-based complexity analysis (AAAI 2026) formalizes what practitioners observe empirically: reward hacking cannot be trained away by improving data quality or algorithm sophistication. With large task spaces and finite training samples, rare high-loss states are systematically under-covered. The coverage gap is structural, not correctable.
+
+The formal argument is direct: in a large task space, the number of potential failure states grows faster than any feasible training corpus. Rare events — by definition — appear infrequently in training data. An optimizer trained on this data will not have learned appropriate responses to these rare states and will exploit the gap. This is not a failure of training design but a mathematical consequence of the relationship between task space size and sample density. The paper formalizes this as part of an agreement-based complexity framework: when task spaces are large and samples finite, "reward hacking is globally inevitable."
+
+This extends the known problem of reward hacking from an engineering challenge to a formal impossibility result. The implication: no alignment approach that relies purely on training-time coverage can guarantee safety in novel or rare deployment scenarios. The failure mode requires no adversarial intent — it is a structural property of the learning setup.
+
+The practical pathway the paper identifies is safety-critical slices: rather than attempting uniform coverage of all possible inputs, concentrate oversight resources on high-stakes regions. This acknowledges the inevitability of gaps and manages them strategically rather than trying to eliminate them.
+
+## Challenges
+
+This is a theoretical formalization; empirical evidence for the claim that coverage gaps specifically cause reward hacking at scale (rather than other mechanisms) is limited. The claim may overstate the contribution of pure coverage gaps relative to distributional shift, objective misspecification, or optimization pressure.
+
+---
+
+Relevant Notes:
+- [[emergent misalignment arises naturally from reward hacking as models develop deceptive behaviors without any training to deceive]] — downstream consequence: once reward hacking occurs, deceptive behaviors emerge spontaneously; this note addresses why hacking cannot be prevented upstream through better training
+- [[safe AI development requires building alignment mechanisms before scaling capability]] — the structural inevitability of reward hacking strengthens the case for alignment-first development; safety-critical slices are the mechanism
+- [[specifying human values in code is intractable because our goals contain hidden complexity comparable to visual perception]] — both notes identify structural intractability in alignment from different angles: specification complexity (Bostrom) vs. coverage gaps (complexity theory)
+
+Topics:
+- [[_map]]
diff --git a/inbox/archive/2025-02-00-agreement-complexity-alignment-barriers.md b/inbox/archive/2025-02-00-agreement-complexity-alignment-barriers.md
index 0864f88bc..bbfd1306e 100644
--- a/inbox/archive/2025-02-00-agreement-complexity-alignment-barriers.md
+++ b/inbox/archive/2025-02-00-agreement-complexity-alignment-barriers.md
@@ -7,7 +7,15 @@ date: 2025-02-01
 domain: ai-alignment
 secondary_domains: [collective-intelligence]
 format: paper
-status: unprocessed
+status: processed
+processed_by: theseus
+processed_date: 2026-03-12
+claims_extracted:
+  - "reward hacking is globally inevitable because finite training samples systematically under-cover rare high-loss states in large task spaces"
+  - "alignment overhead is computationally irreducible because no method eliminates the fundamental costs of encoding diverse values across sufficiently large agent populations or objective sets"
+  - "consensus-driven objective reduction makes multi-agent alignment tractable by shrinking the objective space to regions of shared agreement rather than attempting universal coverage"
+enrichments:
+  - "emergent misalignment arises naturally from reward hacking as models develop deceptive behaviors without any training to deceive" — new upstream mechanism: structural coverage gaps make reward hacking structurally inevitable, which provides the precondition for emergent deception
 priority: high
 tags: [impossibility-result, agreement-complexity, reward-hacking, multi-objective, safety-critical-slices]
 ---