{ "matrix_version": "1.0", "purpose": "Canonical scientific-claims matrix for public project wording.", "status_legend": { "implemented": "Implemented in the codebase, but not externally validated as valuable.", "controlled_evidence": "Supported by controlled benchmark evidence under a narrow protocol.", "strong_intra_theory_evidence": "Strongly supported inside the theory the implementation operationalizes.", "early_controlled_evidence": "Initial comparative or task evidence exists, but remains narrow or early.", "not_established": "Not supported by current public evidence.", "falsified": "Tested under pre-registered protocol and result was negative; no further retry planned under the current design.", "retry_planned": "Tested under pre-registered protocol and result was negative; a targeted follow-up study is planned at higher power or on a more suitable dataset." }, "evidence_level_legend": { "1_theory_fidelity": "Theory fidelity", "2_controlled_retrieval": "Controlled retrieval behavior", "3_appraisal_quality": "Appraisal quality", "4_realistic_tasks": "Realistic memory tasks", "5_human_ecological": "Human / ecological validation" }, "oracle_affect_note": "requires_oracle_affect=true: the claim's primary positive-result benchmarks inject preset valence/arousal at encode time (oracle affect); the LLM/keyword appraisal pipeline is bypassed for those results. requires_oracle_affect=false: the benchmark ran without preset affect injection — either no affect signal (naturalistic) or the appraisal pipeline itself produced the signal.", "claims": [ { "claim_id": "aft_multilayer_engine", "claim_text": "AFT is implemented as a coherent multi-layer memory engine.", "claim_area": "Architecture", "status": "implemented", "evidence_level": "1_theory_fidelity", "requires_oracle_affect": false, "allowed_public_wording": "AFT is implemented as a coherent multi-layer memory engine.", "current_evidence": "Public API, engine/retrieval/state modules, sync/async parity tests.", "not_yet_shown": "External validation of architectural value.", "next_study": "Keep API/docs aligned while external evaluations grow.", "evidence_refs": [ "src/emotional_memory/engine.py", "src/emotional_memory/async_engine.py", "tests/test_async_engine.py", "tests/test_engine.py" ], "benchmark_refs": [], "protocol_refs": [], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "retrieval_affect_aware", "claim_text": "Retrieval is affect-aware, not semantic-only.", "claim_area": "Retrieval behavior", "status": "controlled_evidence", "evidence_level": "2_controlled_retrieval", "requires_oracle_affect": true, "allowed_public_wording": "Retrieval is affect-aware, not semantic-only.", "current_evidence": "127 fidelity cases validate affect-aware ranking logic. The controlled quadrant benchmark (SBERT) ties AFT and naive_cosine at recall@5 ≈ 0.80-0.85 (ceiling effect, N = 20 items) while clearly beating recency (Δ = −0.55, p < 0.001). SOTA extension on affect_reference_v1 (2026-05-07, gpt-4.1-mini): Mem0 reaches recall@5 = 1.00 (Δ = +0.20 vs cosine, p_boot = 0.045) and LangMem 0.90 (Δ = +0.10, p_boot = 0.254); AFT 0.85 (Δ = +0.05, p_boot = 0.825). LLM-backed memory dominates AFT on that simpler probe but at 25× higher encode latency (Mem0 1130 ms/item vs AFT 45 ms/item). The realistic replay v2 benchmark (Hd2 PASS) shows AFT top1 = 0.53 vs naive_cosine = 0.33 with SBERT (N = 200; Δ=+0.205 [0.150,0.265], p_bootstrap<0.001, d=0.49). SOTA replication on v2 (2026-05-07, gpt-4.1-mini, H_v2_sota PASS): on the same v2 dataset Mem0 = 0.330 (Δ vs cosine = +0.005, p=0.945) and LangMem = 0.365 (Δ = +0.040, p=0.233); neither beats naive_cosine. AFT top1 = 0.535 (Δ vs cosine = +0.210 [+0.155,+0.270], p<0.001, d=0.512) — non-overlapping CIs vs both SOTA systems. Pattern inverts across benchmarks: Mem0 excels at semantic quadrant recall (affect_reference_v1) while AFT excels at affect-modulated multi-session recall (v2). Architecture advantage under automatic appraisal: Hg1 FAIL (LLM appraisal + oracle-free; Δ=−0.010, p=0.367). All Hd* results use oracle/preset affect. — FR realistic_recall_v2 (Addendum M, 2026-05-16, multilingual-e5-small, Branch A PASS): N=120 queries × 30 native-FR hand-authored scenarios; AFT top1=0.31 [0.23, 0.39], naive_cosine top1=0.12 [0.07, 0.18], Δ=+0.18 [0.11, 0.26], p_bootstrap=0.0000, Hedges g=0.424. Advantage confirmed in French on same 2-session realistic replay format as EN.", "not_yet_shown": "General downstream superiority over production memory systems in the semantic-recall regime — Mem0 and LangMem outperform AFT on affect_reference_v1 (simple quadrant probe). Architecture advantage under automatic appraisal unconfirmed (Hg1 FAIL). Consistent cross-embedder per-type advantage (H6 momentum_alignment fails for e5-small-v2). Cross-language replication on IT/ES at me5 N=120 in 2-session format remains open (Hd2 Branch C FAIL 2026-05-07); FR is the sole 2-session confirmatory cross-language PASS (Addendum M, me5, N=120). Cross-domain ecological replication: Hk1 FAIL on DailyDialog (Branch B, Δ=-0.008, p_holm=1.000, 2026-05-07). Human / ecological validation (#27 open).", "next_study": "Robust cross-language replication (explicit confirmatory study); ecological corpus with >30% affect-bearing turns per dialogue for naturalistic replication. Inter-rater agreement (κ Cohen) on affect_reference_v1.", "evidence_refs": [ "benchmarks/fidelity/test_emotional_vs_cosine.py", "benchmarks/fidelity/test_mood_congruent.py" ], "benchmark_refs": [ "benchmarks/comparative/results.sbert.md", "benchmarks/comparative/results.sota.sbert.md", "benchmarks/comparative/results.sota.v2.sbert.md", "benchmarks/realistic/results.v2.sbert.md", "benchmarks/realistic/results.v2.e5.md", "benchmarks/appraisal_confound/results.confirmatory.md", "benchmarks/appraisal_confound/results.hd2.sbert.md", "benchmarks/appraisal_confound/results.hd2_it.me5.md", "benchmarks/appraisal_confound/results.hd2_es.sbert.md", "benchmarks/appraisal_confound/results.hd2_es.me5.md", "benchmarks/appraisal_confound/results.hd2_it.me5.v120.md", "benchmarks/appraisal_confound/results.hd2_es.me5.v120.md", "benchmarks/dailydialog/results.md", "benchmarks/realistic/results.v2_fr.me5.md", "benchmarks/realistic/results.v2_fr.me5.json" ], "protocol_refs": [ "benchmarks/comparative/protocol.md", "benchmarks/comparative/protocol.sota.md", "benchmarks/comparative/protocol.sota.v2.md", "benchmarks/preregistration_addendum_m_fr.md", "benchmarks/preregistration_addendum_m_fr_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "realistic_replay_vs_sota", "claim_text": "On multi-session realistic replay (v2), AFT (oracle-affect) outperforms production LLM-backed memory systems (Mem0, LangMem) on top1_accuracy.", "claim_area": "Retrieval behavior vs SOTA", "status": "controlled_evidence", "evidence_level": "4_realistic_tasks", "requires_oracle_affect": true, "allowed_public_wording": "On the realistic multi-session replay benchmark, AFT with preset affect labels achieves top1_accuracy = 0.535 vs Mem0 = 0.330 and LangMem = 0.365 — an advantage of Δ +0.21 (d=0.51) with non-overlapping 95% CIs. Neither LLM-backed system outperforms naive cosine on this benchmark.", "current_evidence": "H_v2_sota PASS (exploratory, 2026-05-07, gpt-4.1-mini, sbert-bge, N=200, n_bootstrap=10000, seed=0): AFT top1=0.535 [0.465,0.600], Mem0=0.330 [0.265,0.395], LangMem=0.365 [0.300,0.430], cosine=0.325 [0.260,0.390]. Pairwise vs cosine: AFT Δ=+0.210 [+0.155,+0.270], p<0.001, d=0.512; Mem0 Δ=+0.005, p=0.945; LangMem Δ=+0.040, p=0.233. Per-challenge AFT advantage strongest on semantic_confound (+0.275), momentum_alignment (+0.275), affective_arc (+0.275). Pattern inverts vs affect_reference_v1 (where Mem0=1.00>AFT=0.85): the two benchmarks test different capabilities. Methodological asymmetry: AFT receives oracle affect; Mem0/LangMem do not.", "not_yet_shown": "Advantage without oracle-affect (Hg1 FAIL). Advantage on factual QA (LoCoMo S1 FAIL). Advantage under realistic appraisal. Human evaluation. Circularity audit (Addendum U, 2026-06-27, SBERT, N=200): 62.5% of v2 queries are AFT-favorable by construction; the aggregate advantage is concentrated there (Δ=+0.304) and is null on the neutral 37.5% (Δ=+0.013, p=0.63) — the headline figure is scoped to a ~62%-affect-discriminative benchmark, not regime-independent.", "next_study": "SOTA replication on v2 with automatic appraisal (would require gpt-4.1-mini appraisal, consistent with Hg1 methodology). Cross-embedder replication (e5-small-v2 on v2 SOTA).", "evidence_refs": [], "benchmark_refs": [ "benchmarks/comparative/results.sota.v2.sbert.json", "benchmarks/comparative/results.sota.v2.sbert.md" ], "protocol_refs": [ "benchmarks/comparative/protocol.sota.v2.md", "benchmarks/preregistration_addendum_v2_sota.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "theory_faithful_operationalization", "claim_text": "The implementation is faithful to the theories it operationalizes.", "claim_area": "Theory fidelity", "status": "strong_intra_theory_evidence", "evidence_level": "1_theory_fidelity", "requires_oracle_affect": true, "allowed_public_wording": "The implementation is faithful to the theories it operationalizes.", "current_evidence": "127 fidelity cases across 20 phenomena validate expected intra-theory behavior. S3 ablation (N=200, SBERT and e5-small-v2) covers all 7 AFT mechanisms: individual no_mood/no_resonance effects are not significant in the expected direction, dual_path and synchronous keyword variants are destructive, and no_reconsolidation remains null on this benchmark. The 8-variant ablation shows system-level advantage but no isolatable single-layer attribution. Hi3 PASS (Addendum I closure, 2026-05-06, N=500, seed=1): the per-challenge embedder amplification on semantic_confound is confirmed — e5 resonance interference exceeds SBERT by Δ=+0.090 [0.030,0.160] (d=0.257, Holm-adj p=0.0234). Secondary Hi3_recency also PASS (Δ=+0.070, p_adj=0.0234); Hi3_arc FAIL (Δ=+0.010). Mechanism channel (Hi2: link-set density) not differentiated at link-count level — both embedders saturate top-5 cap.", "not_yet_shown": "Ecological correspondence to human emotional memory. The LLM-appraisal dual-path advantage on affect-free data is falsified (Addendum G/P, see appraisal_llm_real_dual_path). Mechanism for embedder-dependent amplification not localised at link-count instrumentation level (Hi2 still exploratory).", "next_study": "Human and behavioral validation. The automatic-appraisal dual-path question is closed (Addendum G/P, falsified).", "evidence_refs": [ "benchmarks/fidelity/test_appraisal_affect.py", "benchmarks/fidelity/test_mood_adaptive.py", "benchmarks/fidelity/test_spacing_effect.py", "benchmarks/fidelity/test_state_dependent.py" ], "benchmark_refs": [ "benchmarks/ablation/results.v2.sbert.md", "benchmarks/ablation/results.v2.e5.md", "benchmarks/ablation/results.v3.sbert.md", "benchmarks/ablation/results.v3.e5.md", "benchmarks/ablation/results.hi3.md" ], "protocol_refs": [ "benchmarks/preregistration_addendum_i.md", "benchmarks/preregistration_addendum_i_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "appraisal_directionally_useful", "claim_text": "Appraisal is directionally useful.", "claim_area": "Appraisal quality", "status": "early_controlled_evidence", "evidence_level": "3_appraisal_quality", "requires_oracle_affect": false, "allowed_public_wording": "Appraisal is directionally useful.", "current_evidence": "The appraisal-quality benchmark provides early controlled evidence on natural-language inputs. Human-gold construct validity now measured (Addendum S, EmoBank N=300): LLM valence r=0.70 vs human labels — see appraisal_human_validated.", "not_yet_shown": "Calibration across models, domains, and languages.", "next_study": "Appraisal robustness study across models, domains, and languages.", "evidence_refs": [ "benchmarks/appraisal_quality/test_appraisal_quality.py" ], "benchmark_refs": [ "benchmarks/appraisal_quality/test_appraisal_quality.py" ], "protocol_refs": [], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "appraisal_human_validated", "claim_text": "The LLM appraisal signal has construct validity against human-annotated affect (valence).", "claim_area": "Appraisal quality", "status": "early_controlled_evidence", "evidence_level": "3_appraisal_quality", "requires_oracle_affect": false, "allowed_public_wording": "Against human-annotated affect (EmoBank, N=300), LLM appraisal valence is human-validated (Pearson r=0.70 [0.66, 0.75]); arousal (r=0.28 [0.17, 0.37]) and dominance (r=0.33 [0.22, 0.43]) are weakly-to-moderately positive (CIs exclude 0). A persistent positive valence bias (+0.15) remains. The rule-based KeywordAppraisalEngine is not human-validated (valence r=0.07).", "current_evidence": "Pre-registered Addendum S executed 2026-06-26 on EmoBank (Buechel & Hahn 2017, human VAD, N=300 sampled seed=42, CC-BY-SA 4.0). LLMAppraisalEngine (Scherer CPM, gpt-5-mini) -> to_core_affect() correlated against human VAD, bootstrap n=2000 seed=42: valence r=0.703 [0.655, 0.746] bias +0.152 MAE 0.319; arousal r=0.276 [0.172, 0.373] bias -0.058; dominance r=0.333 [0.224, 0.432] bias +0.130. All three LLM CIs exclude 0. The +0.169 positive valence bias from Addendum N (vs LLM-gold) persists vs human gold (+0.152). KeywordAppraisalEngine is not human-validated: valence r=0.070 [-0.010, 0.148], dominance r=-0.011; only a weak arousal correlation (r=0.106) with a large -0.288 bias. Provenance note (2026-08-04): the KeywordAppraisalEngine figures were measured with the pre-v0.18 per-dimension averaging, which divided each dimension by every matching rule including rules that left it at its neutral default (fixed in v0.18: only contributing rules count). The r=0.07 keyword valence result therefore describes the diluted engine and would need re-measurement before being read as a post-fix number; the LLM-engine results in this claim are unaffected.", "not_yet_shown": "Strong validity on arousal/dominance (only weak-to-moderate). Correction of the standing +0.15 valence bias in the production path (Addendum O recalibration is not wired into LLMAppraisalEngine). Validity across non-English corpora and other annotation schemes. Addendum V (2026-06-27, EmoBank N=300): the weak arousal/dominance axes are largely a SEC→projection artifact — a direct-VAD estimator (LLM rates V/A/D directly) reaches valence r=0.79 (bias −0.01), arousal r=0.58 (Δr +0.354 vs SEC), dominance r=0.43, paired CIs exclude 0 except dominance (touches 0). Adopt direct-VAD recommended (follow-up). Addendum W (2026-06-27): an affine calibration of direct-VAD arousal (arousal_cal = 0.16*arousal_direct + 0.45, fit on held-out EmoBank) cuts arousal MAE 0.20->0.04 (Hw1 p<0.001), below the SEC projection's 0.11 (Hw2 p<0.001), while preserving r -- calibrated direct-VAD dominates the projection on both r (0.57 vs 0.21) and MAE; the MAE cut partly reflects EmoBank's narrow arousal variance (shrink-to-mean).", "next_study": "Wire the Addendum O mapping recalibration into the appraisal path and re-measure bias vs human gold; extend to a second human-annotated corpus and to non-English text.", "evidence_refs": [ "benchmarks/preregistration_addendum_w_arousal_calibration_closure.md" ], "benchmark_refs": [ "benchmarks/human_gold_appraisal/results.json", "benchmarks/human_gold_appraisal/results.md", "benchmarks/datasets/emobank_v1.json", "benchmarks/arousal_calibration/results.json", "benchmarks/arousal_calibration/results.md" ], "protocol_refs": [ "benchmarks/preregistration_addendum_s_human_gold_appraisal.md", "benchmarks/preregistration_addendum_s_human_gold_appraisal_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "replayable_multi_session_help", "claim_text": "AFT helps on replayable multi-session memory tasks.", "claim_area": "Realistic tasks", "status": "controlled_evidence", "evidence_level": "4_realistic_tasks", "requires_oracle_affect": true, "allowed_public_wording": "On the author-constructed affect-conditioned v2 benchmark with preset affect at encoding and retrieval, AFT outperforms cosine (N=200; SBERT delta +0.205). This does not establish naturalistic utility or individual-layer necessity.", "current_evidence": "Realistic replay v2 (50 scenarios, 200 queries, 5 challenge types × 40). SBERT bge-small-en: AFT top1=0.53 vs naive_cosine=0.33, Δ=+0.205 [0.150,0.265], p_bootstrap<0.001, d=0.49. e5-small-v2: AFT top1=0.50 vs naive_cosine=0.34, Δ=+0.155 [0.090,0.225], p_bootstrap<0.001, d=0.31. Advantage holds on both embedder classes. Controlled oracle comparison: Hd1 PASS (Addendum D, seed=1, 2026-04-27); Gate 3 CLOSED. — FR realistic_recall_v2 (Addendum M, 2026-05-16, multilingual-e5-small, Branch A PASS): N=120 queries × 30 native-FR scenarios (2-session multi-turn); AFT top1=0.31 [0.23, 0.39], naive_cosine=0.12 [0.07, 0.18], Δ=+0.18 [0.11, 0.26], p_bootstrap=0.0000, Hedges g=0.424.", "not_yet_shown": "General superiority on external open-domain QA (LoCoMo Gate 1 FAIL — AFT F1=0.168 vs naive_rag=0.271) or completed human/ecological validation. Generalization beyond affect-discriminative scenarios: v2 benchmark has 4/5 challenge types where AFT signals are theoretically relevant by design; advantage on valence-neutral tasks is not established. Circularity audit (Addendum U, 2026-06-27, SBERT, N=200): 62.5% of v2 queries are AFT-favorable by construction; the aggregate advantage is concentrated there (Δ=+0.304) and is null on the neutral 37.5% (Δ=+0.013, p=0.63) — the headline figure is scoped to a ~62%-affect-discriminative benchmark, not regime-independent.", "next_study": "Single MSC human-grounded retrieval study, fully text-appraised encoding and query; hybrid primary baseline, cosine and Mem0 secondary; not executed.", "evidence_refs": [ "benchmarks/realistic/runner.py", "tests/test_realistic_benchmark.py" ], "benchmark_refs": [ "benchmarks/realistic/results.md", "benchmarks/realistic/results.v2.sbert.md", "benchmarks/realistic/results.v2.e5.md", "benchmarks/realistic/results.v2_it.me5.md", "benchmarks/realistic/results.v2_it.sbert.md", "benchmarks/realistic/results.v2_es.sbert.md", "benchmarks/realistic/results.v2_es.me5.md", "benchmarks/realistic/results.v2_fr.me5.md", "benchmarks/realistic/results.v2_fr.me5.json", "benchmarks/ablation/results.v2.sbert.md", "benchmarks/ablation/results.v2.e5.md" ], "protocol_refs": [ "benchmarks/realistic/protocol.md", "benchmarks/appraisal_confound/README.md", "benchmarks/preregistration_addendum_v2.md", "benchmarks/preregistration_addendum_m_fr.md", "benchmarks/preregistration_addendum_m_fr_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "locomo_external_qa_negative", "claim_text": "AFT does not outperform naive RAG on external open-domain conversational QA (LoCoMo).", "claim_area": "External benchmarks", "status": "controlled_evidence", "evidence_level": "4_realistic_tasks", "requires_oracle_affect": false, "allowed_public_wording": "On the LoCoMo conversational QA benchmark (1986 QA pairs, 10 conversations), AFT retrieval underperforms a naive RAG baseline (F1 0.168 vs 0.271; judge_acc 0.279 vs 0.441). Gate 1 was not met. AFT's affective weighting does not help on factual open-domain QA.", "current_evidence": "Pre-registered S1 run completed 2026-04-27. H1 and H2 both fail (Δ<0). Holm-corrected one-tailed p=1.0 for both hypotheses. Add. J Pareto sweep executed 2026-05-06 (10 weight configs × 200-QA stratified subsample, seed=42, 50 per category). Hj1 FAIL: no AFT config achieves F1 ≥ naive_rag on any category. Best AFT config (W2: sem=0.50, mood=0.30, no resonance) achieves aggregate F1=0.1765 vs naive_rag F1=0.2092 (Δ=−0.033); closest cell W2×multi_hop Δ=−0.011. The per-task base_weights tuning line is closed as honest negative. Addendum L (query-type routing, 2026-05-19): closed-loop heuristic routing on LoCoMo 200-QA smoke test. Hl1 FAIL (Δ=−0.017, below +0.02 threshold and in wrong direction). Hl2 FAIL (Δ=−0.081 vs naive_rag). Hl3 data-collection issue (classifier log bug). Branch B verdict: routing ships as optional feature; claim status unchanged. Hl3 follow-up resolved 2026-06-11 (PR #52): the classifier-log alignment bug is fixed (log matched by query text, not list index) and ground-truth heuristic classifier accuracy on the exact 200-QA subset (seed=42) is 0.465 overall, 0.600 excluding adversarial (no routable class); multi_hop is essentially undetected (2/28).", "not_yet_shown": "Whether selective affect suppression on literal queries or factual/affective dual branching could close the gap. base_weights tuning alone cannot match naive_rag (Hj1 FAIL, Add. J closure). Closed-loop query-type routing also FAIL (Addendum L, Hl1 Branch B).", "next_study": "Architectural approaches beyond weight tuning: selective affect suppression, factual/affective dual-branch retrieval, or hybrid systems.", "evidence_refs": [], "benchmark_refs": [ "benchmarks/locomo/results.json", "benchmarks/locomo/results.md", "benchmarks/locomo/pareto_results.json", "benchmarks/locomo/pareto_results.md", "benchmarks/locomo/routing_results.json", "benchmarks/locomo/routing_results.md" ], "protocol_refs": [ "benchmarks/preregistration.md", "benchmarks/preregistration_addendum_j.md", "benchmarks/preregistration_addendum_j_closure.md", "benchmarks/preregistration_addendum_l_query_routing.md", "benchmarks/preregistration_addendum_l_query_routing_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "downstream_value", "claim_text": "In the affect-discriminative regime, AFT's retrieval-ranking advantage converts to higher downstream answer quality.", "claim_area": "Downstream tasks", "status": "early_controlled_evidence", "evidence_level": "4_realistic_tasks", "requires_oracle_affect": true, "allowed_public_wording": "In the affect-discriminative regime (oracle affect + query-time state injection), AFT's retrieval-ranking advantage converts to higher LLM-judged answer quality: on realistic_recall_v2 (N=200, sbert-bge, gpt-5-mini judge) AFT judge-accuracy 0.595 vs cosine 0.440 (Δ=+0.155 [0.095, 0.220], p<0.001), token-F1 Δ=+0.152. No downstream benefit is shown in the naturalistic, automatic-appraisal regime (see locomo_external_qa_negative).", "current_evidence": "Pre-registered Addendum R (Hr1/Hr2) executed 2026-06-26. encode→retrieve→generate→judge on realistic_recall_v2 (200 queries), identical generator + LLM judge (gpt-5-mini) for both systems; only retrieval differs (full 6-signal AFT vs embedding cosine). Hr1 judge_correct: AFT 0.595 vs cosine 0.440, Δ=+0.155 [+0.095, +0.220], paired bootstrap n=2000 seed=42 p<0.001, Holm<0.001, McNemar<0.001 — PASS. Hr2 token-F1: Δ=+0.152 [+0.100, +0.205], p<0.001 — PASS. Retrieval-ranking reference top1 Δ=+0.205 [0.150,0.265] matches Hd2. Positive across all 5 challenge types (weakest recency_confound +0.075, strongest same_topic_distractor / momentum_alignment +0.200). The +0.205 ranking edge converts ~1:1 to a +0.155 answer-quality edge.", "not_yet_shown": "Downstream value in the oracle-free / automatic-appraisal regime: LoCoMo (oracle-free) shows AFT loses downstream (judge 0.279 vs 0.441). The Addendum R PASS is bounded by the A2 oracle-affect + state-injection boundary; production appraisal is not tested here. Addendum T (2026-06-27, retrieve-time query appraisal): the advantage is partly recoverable without query-state labels, while retaining preset event affect — appraising the query text (direct-VAD) and injecting it as the state beats cosine (Δ+0.115 [0.055,0.180], p<0.001) and recovers ~59% of the oracle advantage (~82% on the affect-discriminative subset). The state-injection boundary is partially dissolved, not closed. Addendum T2A (2026-06-27) NOW TESTS the naturalistic regime: retrieve-time query appraisal on DailyDialog (N=120, 396 q, multilingual-e5-small) does NOT beat cosine (aft_query_appraised 0.212 vs cosine 0.220, Δ−0.008 [−0.056,+0.040], p_holm=1.000, 0/3 types — Ht2a FAIL), reproducing the Hk1 null. The query appraisal is faithful (diagnostic valence r=0.69, arousal r=0.74), so this is a regime limit, not an appraisal-quality failure: the production-reachable recovery is BOUNDED to the affect-discriminative regime (Addendum U) and does not extend to naturalistic affect-conditioned dialogue.", "next_study": "Single MSC human-grounded retrieval study, fully text-appraised encoding and query; hybrid primary baseline, cosine and Mem0 secondary; not executed.", "evidence_refs": [], "benchmark_refs": [ "benchmarks/downstream/results.json", "benchmarks/downstream/results.md" ], "protocol_refs": [ "benchmarks/preregistration_addendum_r_downstream.md", "benchmarks/preregistration_addendum_r_downstream_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "resonance_amplification_e5", "claim_text": "Resonance interference is amplified under e5-small-v2 vs SBERT on semantic_confound queries.", "claim_area": "Theory fidelity", "status": "controlled_evidence", "evidence_level": "2_controlled_retrieval", "requires_oracle_affect": true, "allowed_public_wording": "Hi3 PASS (N=500, seed=1): e5-small-v2 shows larger resonance interference than SBERT on semantic_confound queries (Δ=+0.090 [0.030,0.160], d=0.257, Holm-adj p=0.0234). The v2 per-challenge finding (post-hoc Δ=+0.100) is confirmed as not a sample-size artefact.", "current_evidence": "Hi3 confirmatory study (Addendum I closure, 2026-05-06): paired bootstrap n=10,000 seed=1, Holm m=3. Primary Hi3 (semantic_confound): Δ=+0.090, CI=[0.030,0.160], p_one=0.0078, p_adj=0.0234, d=0.257, PASS. Secondary Hi3_recency (recency_confound): Δ=+0.070, p_adj=0.0234, d=0.239, PASS. Hi3_arc (affective_arc): Δ=+0.010, p_adj=0.3795, FAIL. Dataset: realistic_recall_v3.json (N=500, affect-rich).", "not_yet_shown": "Mechanism for amplification: link-count instrumentation shows both embedders saturate top-5 cap (mean≈5.0); link-type or link-strength differentiation not yet measured. Counterfactual (raising ResonanceConfig.cosine_threshold) not executed.", "next_study": "Link-type/strength instrumentation to localise the Hi2 spreading-activation over-fire hypothesis.", "evidence_refs": [], "benchmark_refs": [ "benchmarks/ablation/results.v3.sbert.md", "benchmarks/ablation/results.v3.e5.md", "benchmarks/ablation/results.hi3.md" ], "protocol_refs": [ "benchmarks/preregistration_addendum_i.md", "benchmarks/preregistration_addendum_i_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "Hi3_recency", "claim_text": "Resonance interference amplification under e5 vs SBERT extends to recency_confound queries.", "claim_area": "Theory fidelity", "status": "controlled_evidence", "evidence_level": "2_controlled_retrieval", "requires_oracle_affect": true, "allowed_public_wording": "Hi3_recency PASS (secondary, Holm-corrected): e5 resonance interference exceeds SBERT on recency_confound queries (Δ=+0.070 [0.020,0.130], d=0.239, Holm-adj p=0.0234, N=500).", "current_evidence": "Hi3_recency: paired bootstrap n=10,000 seed=1. Δ=+0.070, CI=[0.020,0.130], p_one=0.0112, p_adj=0.0234, d=0.239, PASS. Secondary hypothesis in Holm m=3 family; reported descriptively.", "not_yet_shown": "Mechanism; replication on independent dataset.", "next_study": "Part of Hi3 family — follow-up contingent on mechanism study.", "evidence_refs": [], "benchmark_refs": [ "benchmarks/ablation/results.hi3.md" ], "protocol_refs": [ "benchmarks/preregistration_addendum_i.md", "benchmarks/preregistration_addendum_i_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "Hi3_arc", "claim_text": "Resonance interference amplification under e5 vs SBERT on affective_arc queries.", "claim_area": "Theory fidelity", "status": "falsified", "evidence_level": "2_controlled_retrieval", "requires_oracle_affect": true, "allowed_public_wording": "Hi3_arc FAIL: no significant amplification on affective_arc queries (Δ=+0.010 [-0.020,0.050], d=0.058, Holm-adj p=0.3795, N=500). The embedder gap does not extend to affective_arc at this sample size.", "current_evidence": "Hi3_arc: paired bootstrap n=10,000 seed=1. Δ=+0.010, CI=[-0.020,0.050], p_one=0.3795, p_adj=0.3795, d=0.058, FAIL. Holm secondary; honest negative result.", "not_yet_shown": "Whether a larger N or different dataset would show signal.", "next_study": "No active follow-up planned; Hi3_arc FAIL scopes the amplification claim to semantic/recency channels.", "evidence_refs": [], "benchmark_refs": [ "benchmarks/ablation/results.hi3.md" ], "protocol_refs": [ "benchmarks/preregistration_addendum_i.md", "benchmarks/preregistration_addendum_i_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "appraisal_llm_real_dual_path", "claim_text": "AFT with LLM dual-path appraisal outperforms naive cosine on affect-free data.", "claim_area": "Appraisal quality", "status": "falsified", "evidence_level": "3_appraisal_quality", "requires_oracle_affect": false, "allowed_public_wording": "Hg1 FAIL (twice): AFT with LLMAppraisalEngine (dual-path, gpt-5-mini) does not outperform naive cosine on affect-free queries. Original Hg1 (realistic_recall_v3_noAF, N=200): delta=-0.010, p=0.367. Re-run with the recalibrated SEC->affect mapping on a leakage-free set disjoint from the calibration data (Addendum P, realistic_recall_v4_noAF, N=160): delta=-0.087 [-0.144,-0.031], p=0.0018, d=-0.242 -- naive cosine significantly ahead. The LLM affect signal is real (beats neutral, Hp2 PASS) and the dual-path schedule is essential (beats synchronous, Hp3 PASS, d=0.95), but the full system still trails pure cosine when queries are affect-free. The oracle-affect circularity remains the scope boundary of the Hd1/Hd2 claims.", "current_evidence": "Addendum G closure 2026-05-07: realistic_recall_v3_noAF (50 scenarios, 200 queries), sbert-bge embedder, gpt-5-mini LLM appraisal. aft_llm_dual top1=0.315 [0.250,0.380] vs naive_cosine top1=0.325 [0.260,0.390]. Δ=-0.010, d=-0.032, p_one=0.3669. Exploratory: Hg2 FAIL (aft_llm_dual vs aft_neutral Δ=0.000); Hg3 PASS (aft_llm_dual vs aft_llm_sync Δ=+0.185, d=0.342). Diagnosis refined (WP-1a, 2026-05-30, gpt-5-mini, N=750 on realistic_recall_v3 oracle): the appraisal signal is correlated but mis-calibrated, not absent — valence Pearson r=0.81, sign accuracy=0.86, systematic valence bias +0.19 and arousal bias -0.14. A pre-registered prompt-calibration attempt (Addendum N, N=150) failed both hypotheses: valence bias dropped to +0.044 (Hn1 requires both axes; arousal unchanged at -0.118) and gold-set regressed 1/15; the prompt change was reverted. Addendum O (mapping recalibration, 2026-05-31, gpt-5-mini, N=750 SEC dump, by-scenario 70/30 split seed 42) numerically recalibrated the SEC→valence/arousal projection on the train scenarios and promoted model M1 (Scherer feature basis preserved; valence intercept constrained to 0). Held-out test (228 events): valence bias +0.200→+0.072, arousal bias -0.144→-0.023; MAE improved on both axes (valence 0.306→0.298, arousal 0.182→0.118), Pearson r within the -0.05 guardrail. Both pre-registered hypotheses (Ho1 bias, Ho2 guardrail) PASS; M1 weights are now live in `_scherer_project`. Addendum P (Hg1 re-run, 2026-05-31, gpt-5-mini, sbert-bge, seed 0, n_bootstrap 10000) executed the pre-registered re-run with the recalibrated M1 mapping on a leakage-free affect-free dataset DISJOINT from realistic_recall_v3 (realistic_recall_v4_noAF, 40 scenarios / 160 queries, pNN_ namespace, zero overlap on scenario/memory ids and event text, frozen before the run): aft_llm_dual top1=0.800 [0.738,0.863] vs naive_cosine top1=0.887 [0.838,0.931], Hp1 delta=-0.0875 [-0.144,-0.031], one-tailed p=0.0018, d=-0.242 -- FAIL, naive cosine SIGNIFICANTLY ahead (gap widened vs Hg1's -0.010 and crossed into significance). Recalibration did not rescue affect-free retrieval. Exploratory: Hp2 dual>neutral PASS (0.800 vs 0.744, delta=+0.056, p=0.030) -- the LLM affect carries signal; Hp3 dual>sync PASS (0.800 vs 0.287, delta=+0.512, d=0.95) -- the deferred dual-path schedule is essential (synchronous appraisal collapses). Net: a better-calibrated affect signal is a net distractor on affect-free queries even though it is not noise; the architecture-vs-cosine delta stays falsified. Addendum Q (affect-aware gating, 2026-06-11, gpt-5-mini shared appraisal, sbert-bge, seed 0, realistic_recall_v5_gate 50 scen / 200 q frozen pre-run, qNN_ namespace disjoint from v3/v4): Hq1 FAIL (dual vs cosine on the affective subset, delta=-0.050 [-0.11,0.00], p_holm=0.0854, d=-0.168 -- wrong direction; on the anchored affect_congruent_tiebreak type 0.160 vs 0.280); Hq3 FAIL (gated vs cosine full corpus, delta=-0.025); Hq4 even the oracle-gate arm is significantly below cosine (delta=-0.045, p=0.0024) -- the bottleneck is the channel, not the gate classifier (LLM gate accuracy 0.844). Hq2 PASS (gated > always-on dual, delta=+0.080, p_holm=0.0009): the front-router recovers the entire always-on penalty exactly (gated == cosine on the affect-free half, Hq5 delta=0.000). Pre-specified Branch C: the affect-routing line is CLOSED; state-based affect signals cannot answer query-driven affect questions (the query is never appraised -- the oracle-affect boundary is also a state-injection boundary). affective_arc_blind queries hit a floor for every system (<=0.02 incl. cosine) under cross-scenario accumulation -- construction limitation, disclosed; the verdict stands on the 50 anchored tiebreak queries.", "not_yet_shown": "Nothing further pre-registered for the affect-free architecture-vs-cosine comparison: the recalibrated mapping (Addendum O M1) does NOT change the Hg1 verdict (Addendum P, leakage-free). The LLM dual-path advantage remains scoped to oracle-affect conditions (Hd1/Hd2). The Hg3/Hp3 PASS (dual-path > synchronous) is reaffirmed. Mapping weights are calibrated for gpt-5-mini's appraisal output; a different appraisal engine may need re-fitting.", "next_study": "DONE. Addendum P closed the affect-free architecture-vs-cosine line; Addendum Q (2026-06-11) closed the state-based affect-aware routing/gating line (Branch C: Hq1/Hq3/Hq4 FAIL, Hq2 PASS -- gating is a safe wrapper but cannot exceed cosine). The residual retrieve-time query-appraisal signal was subsequently pre-registered and executed: Addendum T PASS on curated realistic_recall_v2; T2A FAIL on DailyDialog; X/X2 FAIL inverted on third-party corpora; Addendum Y shipped retrieve_query_gated() as a validated query-affect safe wrapper (v0.17.0). No further residual on this claim.", "evidence_refs": [ "benchmarks/appraisal_diagnostics/results.diagnostic.gpt5mini.json", "benchmarks/preregistration_addendum_n_appraisal_calibration_closure.md", "benchmarks/appraisal_calibration/results.recalibration.gpt5mini.json", "benchmarks/preregistration_addendum_o_mapping_recalibration_closure.md", "benchmarks/preregistration_addendum_p_hg1_rerun_closure.md" ], "benchmark_refs": [ "benchmarks/appraisal_confound/results.hg1.json", "benchmarks/appraisal_confound/results.hg1.md", "benchmarks/appraisal_confound/results.hg1_v4.json", "benchmarks/appraisal_confound/results.hg1_v4.md", "benchmarks/appraisal_confound/results.hq.json", "benchmarks/appraisal_confound/results.hq.md" ], "protocol_refs": [ "benchmarks/preregistration_addendum_g.md", "benchmarks/preregistration_addendum_g_closure.md", "benchmarks/preregistration_addendum_o_mapping_recalibration.md", "benchmarks/preregistration_addendum_p_hg1_rerun.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "cross_domain_affect_replication", "claim_text": "AFT affect-conditioned retrieval advantage generalises to ecological dialogue corpora.", "claim_area": "External benchmarks", "status": "controlled_evidence", "evidence_level": "4_realistic_tasks", "requires_oracle_affect": false, "allowed_public_wording": "Hk1 FAIL (Branch B, 2026-05-07): DailyDialog ecological replication — AFT does not outperform naive cosine on naturalistic short-turn dialogue (N=120 personas, 396 queries, multilingual-e5-small; Δ=-0.008, p_holm=1.000). Addendum M FR PASS (Branch A, 2026-05-16): native French scenarios, 2-session realistic replay (N=120, me5); AFT top1=0.31 vs naive_cosine=0.12, Δ=+0.18 [0.11, 0.26], p<0.0001, Hedges g=0.424. Addendum X FAIL (2026-07-02): first third-party retrieval-native corpus (MADial-Bench EN, NAACL 2025, N=160, oracle-free) — aft_query_appraised significantly below naive cosine (nDCG@5 0.221 vs 0.304, Δ=-0.083 [-0.123, -0.043], p_one=0.9998) despite faithful appraisal (D1 AUC=0.996) and an affect-discriminative corpus (D2=76.9%); construct mismatch — the benchmark rewards counter-congruent supportive recall (emotion regulation), the opposite of AFT's mood-congruence prior. Addendum X2 FAIL (2026-07-07): second third-party corpus (ES-MemEval/EvoEmo, WWW 2026, N=1,133 in-family queries, oracle-free) — aft_query_appraised significantly below naive cosine (upstream-verbatim nDCG@4 0.133 vs 0.284, Δ=-0.150 [-0.165, -0.136], p_one=1.0000, powered: MDE 0.018) despite faithful appraisal (D1 AUC=0.950) and an affect-discriminative corpus (D2=63.0%); failure mode distinct from X — the gold is affect-orthogonal (content-determined; 45.5% of queries affectively neutral, corr(query, gold valence)=+0.25, only 13.1% opposite-sign), not counter-congruent; positive retrieval evidence remains confined to author-crafted corpora. Addendum Z FAIL (2026-07-14): held-out linear learning-to-rank over the 6 AFT retrieval signals (k=5 cross-fitting, Holm m=3) does not beat cosine on any third-party corpus (MADial Δ=+0.015 p_holm=0.314; ES-MemEval Δ=−0.015 p_holm=0.9997; DailyDialog Δ=+0.008 p_holm=0.811) while preserving/amplifying curated on-regime advantage (learned top1=0.425 vs fixed 0.240, Hz2 PASS); boundary hardened from fixed-weight failure to no held-out learned linear affect profile generalizes on third-party gold.", "current_evidence": "Hk1 pre-registered and executed 2026-05-07. Synthetic personas from DailyDialog train+validation (12118 dialogs, 2580 after ≥30% emotion-density filter). 4 programmatic query types from Ekman labels (no LLM). Paired bootstrap n=10000 seed=0, Holm m=4. Aggregate: AFT top1=0.212 vs naive_cosine=0.220 (Δ=-0.008, CI=[-0.056,+0.043], p_holm=1.000, d=-0.015). Only affective_trajectory shows positive trend (Δ=+0.103, d=0.186, N=39) but underpowered (p_holm=0.734). All directional query types FAIL Holm threshold. Branch B confirmed. Addendum T2A (2026-06-27): re-ran the same personas adding an aft_query_appraised arm (direct-VAD query appraisal via the public query_affect API; encode path unchanged). Still FAIL: aft_query_appraised top1=0.212 vs cosine 0.220 (Δ=-0.008 [-0.056,+0.040], p_holm=1.000, 0/3 types); +0.010 (ns) over stale-state AFT. The query appraisal is faithful (diagnostic valence r=0.69, arousal r=0.74), so even retrieve-time query affect — the mechanism that worked on curated realistic_recall_v2 (Addendum T) — does NOT extend to naturalistic short-turn dialogue. This is a stronger negative for cross-domain replication. — FR realistic_recall_v2 (Addendum M, 2026-05-16, multilingual-e5-small, Branch A PASS): N=120 queries × 30 native-FR hand-authored scenarios; AFT top1=0.31 [0.23, 0.39], naive_cosine top1=0.12 [0.07, 0.18], Δ=+0.18 [0.11, 0.26], p_bootstrap=0.0000, p_McNemar=0.0000, Hedges g=0.424. Effect concentrated on same_topic_distractor (top1=0.42) and affective_arc (top1=0.38); momentum_alignment remains weak (top1=0.12). Prior expectation was FAIL; evidence diverges from prior without post-hoc reframing. Addendum X (2026-07-02, MADial-Bench EN, bge-small-en-v1.5, gpt-5-mini direct-VAD, harness merged pre-run): Hx1 FAIL inverted — nDCG@5 0.221 vs cosine 0.304 (Δ=-0.083 [-0.123, -0.043], p_one=0.9998, d=-0.317; MDE 0.051 < |Δ| = powered negative). D1 AUC=0.996, D2=76.9%. Post-hoc: 84/160 negative-valence queries, 73.8% of them with positive gold-set mean valence — counter-congruent supportive recall. Bound now measured on three axes: regime (U/T2A), provenance (author-crafted only), construct (congruence vs regulation). See benchmarks/preregistration_addendum_x_madialbench_third_party_closure.md. Addendum X2 (2026-07-07): pre-registered replication on ES-MemEval/EvoEmo (WWW 2026, CC-BY-4.0, sha256-pinned; upstream session-level qa_retrieval protocol replicated: session-transcript documents, 50-candidate pools, upstream-verbatim metrics). Hx2 FAIL inverted: Δ=-0.150 [-0.165, -0.136], d=-0.622, all 31 grid contrasts negative, uniform across capabilities, worst at rank 1. D1 AUC=0.950 [0.901, 0.988]; D2=63.0% (ex-ante low-D2 prior wrong). Post-hoc: affect-orthogonal gold, not counter-congruence. Two third-party corpora, two distinct measured failure modes; operative boundary sharpened to 'the gold relation itself must be affect-conditioned'. Addendum Z (2026-07-14, make bench-z-profile, 4 corpora, k=5 cross-fitting, paired bootstrap n=10000 seed=0, Holm m=3 + curated non-inferiority ε=0.02). Hz1 FAIL 0/3: aft_learned held-out vs cosine — MADial nDCG@5 0.319 vs 0.304 (Δ=+0.015 [-0.008,+0.040], p_holm=0.314); ES-MemEval u_nDCG@4 0.269 vs 0.284 (Δ=-0.015 [-0.023,-0.006], p_holm=0.9997); DailyDialog top1 0.227 vs 0.220 (Δ=+0.008 [-0.045,+0.061], p_holm=0.811). Hz2 PASS: realistic_recall_v2 learned top1=0.425 vs aft_fixed 0.240 (Δ_vs_fixed=+0.185). Learned beats fixed on third-party (+0.153 MADial, +0.186 ES-MemEval) but not cosine. Negative learned s2 on MADial/ES-MemEval (counter-congruent direction) did not rescue performance. See benchmarks/preregistration_addendum_z_learned_profile_closure.md.", "not_yet_shown": "Cross-domain replication on an affect-rich naturalistic corpus (>30% affect-bearing turns). Whether longer multi-turn sessions with richer emotion vocabulary would show signal consistent with affective_trajectory trend (d=0.186). Cross-language replication in non-Romance languages. IT/ES cross-language replication in 2-session realistic replay format (current IT/ES data uses appraisal_confound single-session runner).", "next_study": "Retry targeted affective_trajectory sub-claim at N≥120 on a naturalistic corpus with >30% affect-bearing turns (DailyDialog affect density too low). Query appraisal already tested (Addendum T2A) and did not help. Pre-register before executing. Gate 2 human eval (kit ready, zero ratings — v1.0 scope).", "evidence_refs": [], "benchmark_refs": [ "benchmarks/dailydialog/results.json", "benchmarks/dailydialog/results.md", "benchmarks/dailydialog/t2a_results.json", "benchmarks/dailydialog/t2a_results.md", "benchmarks/realistic/results.v2_fr.me5.json", "benchmarks/realistic/results.v2_fr.me5.md", "benchmarks/learned_profile/results.json", "benchmarks/learned_profile/results.md" ], "protocol_refs": [ "benchmarks/preregistration_addendum_k_dailydialog.md", "benchmarks/preregistration_addendum_k_dailydialog_closure.md", "benchmarks/preregistration_addendum_t2a_naturalistic_query_appraisal.md", "benchmarks/preregistration_addendum_t2a_naturalistic_query_appraisal_closure.md", "benchmarks/preregistration_addendum_m_fr.md", "benchmarks/preregistration_addendum_m_fr_closure.md", "benchmarks/preregistration_addendum_z_learned_profile.md", "benchmarks/preregistration_addendum_z_learned_profile_closure.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "models_human_emotional_memory", "claim_text": "The system models human emotional memory.", "claim_area": "Ecological validity", "status": "not_established", "evidence_level": "5_human_ecological", "requires_oracle_affect": false, "allowed_public_wording": "The system is theory-inspired, but does not yet have human or ecological validation.", "current_evidence": "Theory-inspired design only.", "not_yet_shown": "Human behavioral correspondence.", "next_study": "Gate 2 (human eval): kit ready (benchmarks/human_eval/), zero ratings collected. Tracking continues on roadmap v1.0; does not block the v0.11.x line.", "evidence_refs": [ "benchmarks/human_eval/README.md" ], "benchmark_refs": [], "protocol_refs": [ "benchmarks/human_eval/README.md", "benchmarks/human_eval/protocol.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] }, { "claim_id": "query_affect_gate", "claim_text": "A query-affect-conditioned gate makes AFT a safe wrapper off-regime without sacrificing the on-regime gain.", "claim_area": "Retrieval mechanism", "status": "controlled_evidence", "evidence_level": "4_realistic_tasks", "requires_oracle_affect": false, "allowed_public_wording": "Addendum Y (Branch A, 2026-07-07): a query-affect-conditioned gate — route the query to pure cosine when the appraised query is affectively neutral (|valence|<0.2), else to the retrieve-time affect-conditioned arm — is a validated partial safe wrapper. Hg1 (recover, ES-MemEval) and Hg2 (preserve, realistic_recall_v2) both PASS (Holm m=2, gated>aft +0.076 and gated>cosine +0.095). The gate recovers exactly the neutral-query component of the off-regime penalty (~50% on ES-MemEval's 45.5%-neutral queries; 0% on MADial's 0%-neutral queries) and preserves the full on-regime gain; it does not fix the X/X2 boundary (affect-carrying-but-misaligned queries stay un-gateable). Shipped as EmotionalMemory.retrieve_query_gated() / AsyncEmotionalMemory.retrieve_query_gated() in v0.17.0 (EmotionalMemoryConfig.query_affect_gate_tau, default 0.2).", "current_evidence": "Addendum Y (make bench-y-gate, 4 corpora, tau=0.2 primary + {0.1,0.2,0.3} sensitivity, paired bootstrap n=10k seed=0, Holm m=2). The gated arm is an exact per-query selection between the naive_cosine and aft_query_appraised arms. Hg1 (ES-MemEval) gated>aft delta=+0.076 [+0.065,+0.087] p_holm=0.0000; Hg2 (realistic_recall_v2) gated>cosine delta=+0.095 [+0.050,+0.140] p_holm=0.0000. Per-corpus recovery tracks the neutral-query fraction of the penalty: ES-MemEval 45.5% neutral -> 50% recovered (gated still -0.075 vs cosine); MADial 0% neutral -> 0% recovered (gate never fires); DailyDialog best-of-both (ns); curated preserve (gated==aft, +0.095 vs cosine). tau-monotone on ES-MemEval, flat on curated. The X/X2 gold-relation boundary is unchanged. Production API shipped in v0.17.0: EmotionalMemory.retrieve_query_gated() / AsyncEmotionalMemory.retrieve_query_gated() force semantic-only weights [1,0,0,0,0,0] on the neutral arm via precomputed_weights (bypassing adaptive_weights); tau defaults to EmotionalMemoryConfig.query_affect_gate_tau=0.2.", "not_yet_shown": "Gains beyond safe-wrapper (the gate cannot beat cosine on affect-orthogonal/counter-congruent corpora).", "next_study": "None scheduled. The production retrieve_query_gated() API shipped in v0.17.0. Remaining limitation is the X/X2 gold-relation boundary (affect-carrying-but-misaligned queries stay un-gateable), not an API gap.", "evidence_refs": [ "benchmarks/preregistration_addendum_y_query_affect_gate_closure.md" ], "benchmark_refs": [ "benchmarks/gate/results.json" ], "protocol_refs": [ "benchmarks/preregistration_addendum_y_query_affect_gate.md" ], "limitations_refs": [ "docs/research/08_limitations.md" ] } ], "interpretation_correction_2026_10_04": { "summary": "No demonstrated naturalistic or third-party retrieval advantage. Positive retrieval evidence remains on author-constructed affect-conditioned benchmarks with preset affect at encoding.", "query_appraisal": "Addendum T removes query-state oracle only; event valence/arousal remain preset. Its +0.115 is not independent of the curated benchmark.", "circularity": "62.5% is the fraction of queries in a partition using affect separation and cosine errors, not a causal fraction of benefit. Non-significance is not equivalence.", "next_study": "MSC human-grounded retrieval; not executed. See benchmarks/msc/protocol.md.", "latency": "Historical no-LLM encode latency does not describe direct-VAD encoding or query appraisal." } }