Skip to content

Research Foundations

The library's theoretical basis is documented in the following research papers:

  • Philosophical Foundations — From Greek origins to enactivism
  • Neuroscience — Emotional memory from a cognitive neuroscience perspective
  • Computational Models — Survey of computational emotion models
  • State of the Art — LLM memory systems and affective computing
  • Design Principles — Affective Field Theory design rationale
  • Bibliography — Full reference list
  • Related Work — 33-paper survey and comparative table positioning AFT
  • Limitations — What is validated today, and what is not
  • Current Evidence — Study ladder and claim/evidence matrix
  • Scientific Quality Bar — Pre-committed gates before upgrading public claims
  • Dominance Design (G7) — Option A: CoreAffect 3D design note (shipped v0.8.2)
  • Multilingual Follow-up (Addendum H + M) — Cross-embedder robustness: Italian AFT advantage stable across SBERT and me5-small; French Hm1 PASS (Addendum M, 2026-05-16)
  • Addendum L — Query-Type Routing — Closed-loop heuristic routing on LoCoMo: Hl1 Branch B FAIL (smoke test, 2026-05-19)
  • Addendum R — Downstream task (Hr1/Hr2) — encode→retrieve→generate→judge on realistic_recall_v2 (N=200, oracle affect): AFT judge-accuracy 0.595 vs cosine 0.440 (Δ=+0.155, p<0.001) — PASS, bounded to the oracle-affect regime (2026-06-26)
  • Addendum S — Human-gold appraisal — LLM appraisal vs EmoBank human VAD (N=300): valence human-validated r=0.70 [0.66, 0.75]; arousal/dominance weak; +0.15 bias persists; keyword engine not validated (2026-06-26)
  • Addendum U — Circularity audit — realistic_recall_v2 decomposition: ~62.5% of queries are AFT-favorable by construction; the advantage concentrates there and is null on the affect-neutral remainder (2026-06-27)
  • Addendum V — Direct-VAD appraisal — direct-LLM valence/arousal/dominance beats the Scherer SEC→projection on EmoBank human gold (valence r=0.79, arousal r=0.58; Hv1/Hv3 PASS); shipped as opt-in DIRECT_VAD_SCHEMA (2026-06-27)
  • Addendum T — Retrieve-time query appraisal — appraising the query at retrieve-time (no oracle) beats cosine on curated realistic_recall_v2 (Ht1 PASS, Δ=+0.115, ~59% of the oracle advantage recovered); production-reachable, bounded to the affect-discriminative regime (2026-06-27)
  • Addendum T2A — Naturalistic query-appraisal re-test — the same mechanism on naturalistic DailyDialog does not beat cosine (Ht2a FAIL, Δ=−0.008, p_holm=1.000); a faithful diagnostic (valence r=0.69, arousal r=0.74) shows this is a regime limit, not an appraisal-quality failure (2026-06-27)
  • Addendum W — Arousal affine calibration — affine calibration of direct-VAD arousal on held-out EmoBank cuts MAE 0.20→0.04 while preserving r (dominates the SEC projection on both axes); adopted as a measurement/reporting transform only — library integration evaluated and declined (would compress arousal into [0.45, 0.61] and silently break the decay floor and s3) (2026-06-28)
  • Addendum X — Third-party retrieval (MADial-Bench) — first released third-party retrieval-native emotional corpus (NAACL 2025, N=160, oracle-free): Hx1 FAIL, inverted — cosine significantly ahead (nDCG@5 0.304 vs 0.221, Δ=−0.083, powered negative) despite near-perfect appraisal (D1 AUC=0.996) and an affect-discriminative corpus (D2=76.9%). Post-hoc: the benchmark rewards counter-congruent supportive recall (emotion regulation) — a construct boundary on top of the regime (U/T2A) and provenance bounds (2026-07-02)
  • Addendum X2 — Third-party retrieval (ES-MemEval/EvoEmo) — second released third-party corpus (WWW 2026, N=1,133 in-family QA queries, oracle-free): Hx2 FAIL, inverted — cosine significantly ahead (upstream-verbatim nDCG@4 0.284 vs 0.133, Δ=−0.150, powered: MDE 0.018) despite faithful appraisal (D1 AUC=0.950) and an affect-discriminative corpus (D2=63.0%). Post-hoc: failure mode distinct from X — affect-orthogonal content-determined gold (45.5% neutral queries, corr(query, gold)=+0.25), not counter-congruence; the provenance bound hardens — the gold relation itself must be affect-conditioned (2026-07-07)
  • Addendum Y — Query-affect-conditioned gate — Branch A PASS: gate the affect channel on the appraised query — route neutral queries (|valence|<0.2) to pure cosine, else to the affect-conditioned arm. Hg1 (recover, ES-MemEval) gated>aft +0.076 and Hg2 (preserve, curated) gated>cosine +0.095, both Holm m=2. The gate recovers exactly the neutral-query component of the off-regime penalty (~50% on ES-MemEval, 0% on MADial where no query is neutral) and preserves the on-regime gain — a validated safe wrapper, not a fix for the X/X2 gold-relation boundary. Shipped as retrieve_query_gated() in v0.17.0 (2026-07-07)
  • Addendum Z — Held-out learned retrieval profile — Branch B: Hz1 FAIL (0/3 break corpora; no held-out learned linear profile beats cosine), Hz2 PASS (curated learned top1=0.425 vs fixed 0.240). Cross-fitted pairwise LTR over the 6 retrieval signals; negative learned s2 on MADial/ES-MemEval did not rescue third-party performance. Hardens the boundary from fixed-weight failure to no held-out learned linear profile generalizes on third-party gold (2026-07-14)
  • Addendum AB — Semantic shared-history exploratory pilot — OPEN unexecuted protocol (Proposed ADR 0002 / issue #193): freeze a CPU BAAI/bge-small-en-v1.5 rerun of the existing 36-query / 12-scenario development set before any v2 sizing. Not confirmation; no scores in this document (2026-11-12)
  • Semantic pilot before shared-history v2 (ADR 0002) — Proposed; do not size v2 from hash-only discordance; runner PR #186 is merged; exploratory-pilot protocol drafted in Addendum AB (2026-11-12); acceptance is a separate owner decision
  • Audit (2026-04) — Critical self-review of strengths, gaps, and priority order
  • claim_validation_matrix.json — Canonical machine-readable source for public scientific claims