Known Limitations of Affective Field Theory¶

This section documents the acknowledged limits of the current implementation. Transparency about limits is integral to any claim of scientific rigor.

Interpretation correction (2026-10-04): Addendum T's encoding path still uses dataset valence/arousal; only the query oracle is removed. Its result remains on the curated corpus. Addendum U's 62.5% counts queries in a partition using affect separation and cosine errors, not the causal proportion of explained benefit. Neither non-significance nor appraisal correlation establishes equivalence or a causal explanation of retrieval failures. See the current evidence page for the planned external human-grounded study; it has not been executed.

1. Limits of the Affective Model¶

1.1 Affect dimensionality¶

Since v0.8.2, CoreAffect is a 3D PAD (Pleasure-Arousal-Dominance) value object (Mehrabian & Russell 1974), implemented as a continuous coordinate (valence, arousal, dominance). The dominance dimension was promoted from a MoodField heuristic to a first-class field in commit 8b9ddbe. Remaining limitations:

Contested circumplex structure: Russell (2003) himself revised the theory toward "core affect" as a contextual construction. The current model uses the classic 1980 form + Mehrabian dominance extension, not the constructionist revisions.
Limited emotional granularity: the Plutchik mapping to 8 primary emotions is a discretization of the continuous space. Complex emotions (nostalgia, Schadenfreude, awe) have no direct representation.
Dominance estimation: KeywordAppraisalEngine infers dominance from coping_potential vocabulary; LLMAppraisalEngine delegates to the Scherer CPM prompt. Cross-cultural validity of dominance estimates is not evaluated. Update (Addendum V, 2026-06-27): against human gold (EmoBank, N=300) the production SEC→linear-projection is the bottleneck on the non-valence axes (arousal r=0.23, dominance r=0.31). Asking the LLM for V/A/D directly lifts arousal to r=0.58 (Δr +0.354), dominance to r=0.43, and valence to r=0.79 with near-zero bias — adopting a direct-VAD appraisal schema is recommended as a follow-up. See benchmarks/preregistration_addendum_v_direct_vad_closure.md. Update (Addendum W, 2026-06-27): the one residual direct-VAD caveat — arousal had the better correlation but a worse MAE than the projection (0.19 vs 0.11) — is resolved by an affine calibration (arousal_cal = 0.16·arousal_direct + 0.45, fit vs EmoBank): held-out arousal MAE drops to 0.04, so calibrated direct-VAD now dominates the production projection on both r (0.57 vs 0.21) and MAE (0.04 vs 0.11) (Hw1/Hw2 PASS, Gw OK). Caveat: the large MAE cut partly reflects EmoBank's narrow arousal variance (shrink-to-mean); the durable claim is dominance-on-both-axes via r-preservation, not the absolute 0.04. See benchmarks/preregistration_addendum_w_arousal_calibration_closure.md.

1.2 Language dependence¶

Validated robustly in English only: the headline realistic-recall advantage (SBERT d=0.49, e5-small-v2 d=0.31, N=200) is robust and multi-embedder. Italian (Hd2_IT, N=80) shows a significant hit@k advantage but top1 non-significant; a pre-registered power top-up to N=120 on multilingual-e5-small returns Δ=+0.058 (p=0.276) — FAIL at declared power. Spanish (Hd2_ES, SBERT, N=80): Δ=+0.138, p=0.045 — a single directional positive, not power-replicated; me5 FAIL at both N=80 (Δ=+0.113, p=0.110) and N=120 (Δ=0.000, p=1.00). See power top-up closure benchmarks/preregistration_addendum_hd2_powertopup_closure.md.
KeywordAppraisalEngine is English-only: the LLMAppraisalEngine handles multilingual input, but appraisal quality varies by language.
Cross-lingual validation: EN/IT/ES/FR confirmed, broader scope open: English (robust), French (Hm1 PASS, me5 N=120), and Spanish-SBERT exploratory (N=80) are validated; Italian and Spanish me5 FAIL at N=120 power. Generalisation to other languages is not established. Larger multilingual embedders (e.g., BGE-M3) are a natural next step.

2. Limits of Empirical Validation¶

2.1 No validation with human users¶

The 127 psychological fidelity tests validate that the system behaves coherently with the theories it implements (for example, that retrieval is mood-congruent and that decay follows a power law). They do not validate that the system's behavior corresponds to how real human beings form and retrieve emotional memories.

This is the critical distinction between intra-theoretical validation (tested) and ecological validation (not tested). A system that faithfully implements a wrong theory passes all fidelity tests.

2.2 Synthetic comparative benchmarks¶

The repository includes a controlled comparative benchmark today (benchmarks/comparative/) on affect_reference_v1, a synthetic 258-example affect-labeled dataset with 4 mood-congruent retrieval queries. This benchmark is useful for measuring whether AFT changes ranking in the theoretically expected direction, but it has important limits:

Small and synthetic setup: it does not represent multi-session conversations or realistic agentic tasks.
Protocol oriented to affect-aware retrieval: this comparative benchmark measures recall@k per affective quadrant, not downstream answer quality. (Downstream answer quality has since been measured separately, in the affect-discriminative oracle-affect regime — Addendum R, §"Downstream value": AFT LLM-judged answer accuracy 0.595 vs cosine 0.440, Δ=+0.155, p<0.001.)
Explicit affect for affect-aware adapters: AFT receives the query's affective context via valence/arousal; generalist baselines can ignore it entirely.
Non-uniform latencies: the numbers mix local stores, optional dependencies, and systems designed for different purposes.

Accordingly, current comparative results should be read as early controlled evidence, not as proof of general superiority over production-grade memory systems.

2.3 Public datasets still small¶

The repository now ships two public dataset families:

affect_reference_v1.jsonl: synthetic benchmark targeted at mood-congruent retrieval
realistic_recall_v1.json: multi-session replay benchmark, still small and scripted

This improves reproducibility relative to earlier versions. The realistic benchmark now rejects trivial queries (candidate_count <= top_k) and uses top1_accuracy as the primary metric rather than relying on hit@k alone. Even so, it is not yet a standardized, large, genuinely multi-turn dataset capable of supporting strong between-system comparisons on emotional memory.

Comparisons with Mem0 and LangMem are no longer hypothetical: the repository ships adapters and reproducible results on the controlled benchmark. These comparisons, however, remain limited by the synthetic protocol, by the different functional surfaces of the systems, and by the absence of human eval or realistic downstream tasks. Letta remains unevaluated without access to the cloud service / API key.

The multi-session replay benchmark (v1.4) has been expanded to 50 scenarios / 100 queries with a rebalanced challenge-type distribution. AFT separates clearly from a pure recency baseline and leads naive_cosine on aggregate top1 (0.70 vs 0.50, N = 100, sbert-bge). On the semantic_confound subset (N = 30), AFT top1 = 0.73 vs naive 0.47, Δ = +0.27 [0.10, 0.43], p_adj = 0.006 — the first per-challenge result to survive Holm correction. Other per-subset results remain below the significance threshold. The earlier hash embedder regression (AFT 0.12 vs naive 0.25) is confirmed as an embedder artefact in benchmarks/realistic/challenge_subset_pairwise.json. The benchmark still does not support strong claims of general superiority over semantic-only retrieval in fully naturalistic multi-turn scenarios.

2.4 Oracle-affect circularity in Hd* studies¶

The primary architecture comparison studies (Hd1, Hd2, Hd2_IT, Hd2_ES) compare aft_noAppraisal — AFT using preset valence/arousal values from the benchmark dataset — against naive_cosine which has no access to any affective signal.

This means the Hd* numbers do not test whether AFT with automatic appraisal beats naive cosine. They test: "given a perfect oracle of affect, does the AFT retrieval architecture (mood field, momentum, resonance, decay) utilize that oracle better than a pure cosine baseline?"

The dataset's affect values were hand-crafted by the author with AFT theory in mind (valence/arousal chosen to discriminate the target memory). This introduces two distinct concerns:

Operationalization gap: real-world AFT deployments use LLMAppraisalEngine or KeywordAppraisalEngine, not preset oracle values. The Ha2/Hb2 results show that even keyword appraisal collapses performance (Δ = −0.39). The architecture advantage under automatic appraisal is now established as negative — see the Addendum G / P results below.
Dataset circularity: affect labels were designed to favor affective discrimination. Scenarios where valence is not discriminative would reduce the oracle advantage. This proportion is now audited (Addendum U, 2026-06-27, SBERT, N=200): 62.5% of queries are AFT-favorable by construction (the gold is not cosine-top-1 and is the affect-closest candidate to the query state). The entire aggregate advantage lives there (Δ=+0.304 [+0.224, +0.384] in that cell) and is null on the neutral 37.5% (Δ=+0.013 [0.000, +0.040], p=0.63). The advantage is real where affect discriminates but is confined to that regime; the +0.205/+0.18 headline figures are scoped to a benchmark that is ~62% affect-discriminative by construction, not a regime-independent effect. See benchmarks/preregistration_addendum_u_circularity_audit_closure.md.

Addendum G (executed, FAIL). The architecture comparison was run with LLMAppraisalEngine on an affect-free dataset (realistic_recall_v3_noAF, 50 scenarios, 200 queries, sbert-bge, gpt-5-mini): dual-path AFT did not beat naive cosine (top1 0.315 vs 0.325, Δ=−0.010, p_one=0.367).

Addendum O + P (recalibration, then re-run — still FAIL). WP-1a diagnosed the appraisal signal as mis-calibrated, not absent (valence Pearson r=0.81). Addendum O numerically recalibrated the Scherer SEC→valence/arousal projection (model M1, valence bias +0.200→+0.072 on held-out scenarios — a calibration PASS). Addendum P then re-ran Hg1 with M1 on a leakage-free affect-free dataset disjoint from the calibration data (realistic_recall_v4_noAF, 40 scenarios, 160 queries, frozen before the run): naive cosine was significantly ahead — top1 0.800 vs 0.887 (Δ=−0.0875 [−0.144, −0.031], pone=0.0018, d=−0.242). Exploratory contrasts show the signal is real but not enough: LLM affect beats fixed-neutral (Hp2, +0.056, p=0.030) and the deferred dual-path schedule is essential (Hp3, +0.512, d=0.95; synchronous appraisal collapses to 0.287). A better-calibrated affect signal is a _net distractor on affect-free queries where semantics alone is already highly discriminative.

Conclusion. Hd* results should be read as "architecture potential under oracle affect", not "architecture advantage in production". The oracle-affect scope is a hard boundary; calibration quality does not, by itself, convert into an affect-free retrieval gain. See benchmarks/preregistration_addendum_p_hg1_rerun_closure.md.

State-injection boundary (Addendum Q, 2026-06-11). The oracle-affect boundary is sharper than calibration: in the Hd-family protocols each query carries a state field injected into the engine before retrieval, i.e. the benchmark performs the query↔state alignment. Addendum Q tested whether the session trajectory supplies that alignment for free on affect-discriminative queries (LLM-inferred affect, gated front-router, realistic_recall_v5_gate): it does not — the affect channel loses to cosine even there (tiebreak 0.160 vs 0.280, Hq1 FAIL), and even a perfect gate stays below cosine (Hq4, Δ=−0.045, p=0.0024). Gating only neutralizes the always-on penalty (Hq2 PASS, +0.080; gated == cosine on affect-free queries). Because retrieval signals are state-based and the query is never appraised, query-driven affect questions are architecturally out of reach for the current signal set; the affect-routing line is closed. See benchmarks/preregistration_addendum_q_affect_gating_closure.md.

Update (Addendum T, 2026-06-27) — the boundary is partially dissolved. Replacing the oracle query.state with the query's affect appraised at retrieve-time (direct-VAD on the query text, then injected as the state) makes AFT beat cosine without query-state labels, while retaining preset event affect: Δ=+0.115 [0.055, 0.180], p<0.001 on realistic_recall_v2 (N=200, SBERT), recovering ~59% of the oracle advantage overall and ~82% on the affect-discriminative subset. The appraised query affect tracks the oracle state (valence r=0.80, arousal r=0.56). So the headline query-state contribution is partly recoverable from text — query-state labels are not required for the affect-discriminative regime. Caveat: bounded to that regime (Addendum U). See benchmarks/preregistration_addendum_t_query_appraisal_closure.md.

Update (Addendum T2A, 2026-06-27) — the recovery is regime-bound. The natural follow-up tested the same retrieve-time query-appraisal mechanism on naturalistic affect-conditioned dialogue (DailyDialog, N=120 personas / 396 queries, multilingual-e5-small) via the public query_affect API. It does not beat cosine: aft_query_appraised top1=0.212 vs cosine 0.220 (Δ=−0.008 [−0.056, +0.040], p_holm=1.000, 0/3 directional types — Ht2a FAIL), reproducing the Hk1 null and recovering only +0.010 (ns) over the stale-state AFT arm. Crucially the query appraisal is faithful here too (diagnostic valence r=0.69, arousal r=0.74), so the failure is a regime limit, not an appraisal-quality failure: where lexical/semantic overlap already carries the signal (naturalistic, affect-sparse QA), affect does not discriminate. The production-reachability established in Addendum T is therefore confined to the affect-discriminative regime and does not generalize to naturalistic dialogue. See benchmarks/preregistration_addendum_t2a_naturalistic_query_appraisal_closure.md.

Update (Addendum X, 2026-07-02) — the mechanism fails on third-party gold sets, for a new reason. The first test on a released third-party retrieval-native corpus (MADial-Bench EN, NAACL 2025: 160 emotion-triggered recall queries with gold memory sets, fully oracle-free, harness merged before the scored run) is a decisive FAIL, inverted: cosine significantly beats aftquery_appraised on nDCG@5 (0.304 vs 0.221, Δ=−0.083 [−0.123, −0.043], p_one=0.9998, d=−0.317; MDE 0.051 < |Δ| → powered negative). The diagnostics exonerate the usual suspects: the appraisal is near-perfect against the third-party emotion labels (D1 AUC=0.996) and the corpus is _more affect-discriminative than the curated v2 (D2 76.9% vs 62.5%). Post-hoc (exploratory): 84/160 queries carry negative appraised valence, and for 73.8% of those the gold-set mean valence is positive — the benchmark operationalizes emotion-triggered recall as interpersonal emotion regulation (recall a positive memory to support a distressed user), i.e. counter-congruent recall, while AFT's affect channel implements mood-congruent retrieval (Bower 1981). The affect signal did what the theory says and pushed the gold memories down. The query-appraisal advantage is therefore bounded on three measured axes: regime (U/T2A), provenance (positive evidence only on author-crafted corpora), and construct (congruence vs supportive counter-congruence). A support-mode retrieval profile is the theory-level follow-up; it must be designed from the emotion-regulation literature and validated on held-out data, not tuned on this corpus. See benchmarks/preregistration_addendum_x_madialbench_third_party_closure.md.

Update (Addendum X2, 2026-07-07) — the second third-party corpus fails for a second, distinct reason. The replication reserved by the X closure ran the same oracle-free mechanism on ES-MemEval/EvoEmo (WWW 2026, CC-BY-4.0: longitudinal emotional-support QA, N=1,133 in-family queries over 401 session documents, 50-candidate pools replicating the upstream evaluation; harness merged before the scored run). Hx2 FAIL, inverted and powered by an order of magnitude: cosine ahead on upstream-verbatim nDCG@4 (0.284 vs 0.133, Δ=−0.150 [−0.165, −0.136], p_one=1.0000, d=−0.622; MDE 0.018), consistent across all 31 grid contrasts and all capabilities. Diagnostics again exonerate the mechanism's inputs: appraisal faithful (D1 AUC=0.950 [0.901, 0.988]) and — contrary to the pre-registered low-D2 prior — the corpus is affect-discriminative (D2=63.0%). Post-hoc (exploratory): unlike MADial-Bench, the gold is not counter-congruent (corr(query valence, gold valence)=+0.25; only 13.1% opposite-sign vs X's 40%) — it is affect-orthogonal: 45.5% of queries are affectively neutral and the right session is determined by content, so the affect channel buys ranking variance with no gold-directed signal. Two third-party corpora, two distinct failure modes (counter-congruent construct on X; affect-orthogonal gold on X2); what they share is that the gold relation was not authored around mood-congruence. The operative boundary is therefore sharper than the D2 regime criterion: the gold relation itself must be affect-conditioned, which to date has only been observed in author-crafted corpora. See benchmarks/preregistration_addendum_x2_esmemeval_third_party_closure.md.

Update (Addendum Y, 2026-07-07) — a query-affect gate removes the self-inflicted penalty, not the boundary. Gating the affect channel on the appraised query (neutral query -> pure cosine, else the affect-conditioned arm) is a validated safe wrapper (Branch A: Hg1 recover + Hg2 preserve both PASS): it recovers exactly the neutral-query component of the off-regime penalty (~50% on ES-MemEval’s 45.5%-neutral queries; 0% on MADial, where every distressed-user query is affect-carrying) and preserves the full on-regime gain (+0.095 vs cosine on curated). It does not move the X/X2 boundary: affect-carrying queries whose gold is counter-congruent (X) or content-determined (X2) stay un-gateable. Shipped as retrieve_query_gated() in v0.17.0 (EmotionalMemoryConfig.query_affect_gate_tau, default 0.2). See benchmarks/preregistration_addendum_y_query_affect_gate_closure.md.

Update (Addendum Z, 2026-07-14) — learned linear weights do not generalize on third-party gold. The last untested lever — a held-out linear learning-to-rank profile 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; Branch B). The fixed weight vector was mis-specified (learned beats fixed by +0.153/+0.186 on MADial/ES-MemEval), but optimal held-out linear weighting is not sufficient to close the cosine gap where gold is counter-congruent (X) or affect-orthogonal (X2). On curated, learning preserves and amplifies the on-regime advantage (learned top1=0.425 vs fixed 0.240, Hz2 PASS). Negative learned s2 on MADial/ES-MemEval matches the counter-congruent direction but does not translate into a win. The boundary upgrades from "these fixed weights fail" to "no held-out learned linear affect profile generalizes out-of-sample on third-party corpora." See benchmarks/preregistration_addendum_z_learned_profile_closure.md.

2.5 Resonance magnitude amplification on e5-small-v2¶

S3 ablation (no_resonance variant, e5-small-v2) found that removing the ResonanceLink layer improves top1 accuracy: Δ = +0.075 [0.040, 0.115], p_boot=0.0002, p_adj=0.001 (Holm-corrected, n_bootstrap=10000). Per-challenge decomposition (Addendum I, benchmarks/preregistration_addendum_i.md) confirms this is magnitude amplification, not sign reversal: Δ≥0 on all five challenge types, concentrated in semantic_confound (e5 Δ=+0.125 vs SBERT Δ=+0.025).

The SBERT embedder shows Δ = +0.030 (raw p_boot=0.022, family-corrected NS p_holm=0.109) for the same ablation — a non-significant result under Holm correction. The effect is thus embedder-dependent: statistically significant on e5 but not on SBERT.

Hi3 confirmatory closure (2026-05-06, N=500, seed=1, Holm m=3): the cross-embedder amplification on semantic_confound is confirmed at the pre-registered threshold. Primary Hi3 PASS (Δ=+0.090 [0.030, 0.160], d=0.257, Holm-adj p=0.0234). Secondary Hi3_recency PASS (Δ=+0.070, p_adj=0.0234) extends the amplification to recency_confound; Hi3_arc FAIL (Δ=+0.010, p_adj=0.380) scopes it: the embedder gap does not extend to affective_arc queries. The v2 finding is not a sample-size artefact. Closure document: benchmarks/preregistration_addendum_i_closure.md.

Mechanism remains exploratory: link-count instrumentation shows both embedders saturate the top-5 link cap (mean ≈ 5.0), so link density alone does not differentiate the embedders. The Hi2 spreading-activation over-fire hypothesis is not yet localised at link-density granularity; link-type and link-strength distributions remain unmeasured.

Potential explanations:

Geometry incompatibility: e5-small-v2's distance space clusters semantically related items more aggressively than SBERT. Spreading activation over this geometry may amplify noise rather than signal, causing resonance links to hurt rather than help.
Resonance as amplifier: if the spreading-activation BFS traverses links that connect semantically similar but contextually incorrect memories, it boosts distractors in e5's tighter cluster geometry.
Benchmark-specific: realistic_recall_v2 contains challenge types (semantic confound, same-topic distractor) explicitly designed to create confusable neighbors. Resonance may be counterproductive when the primary challenge is distractor rejection rather than recall.

This finding does not refute resonance's theoretical role, but it indicates that the interaction between the spreading-activation mechanism and the embedder's distance geometry is non-trivial and embedder-dependent. It is disclosed in claim_validation_matrix.json under the dedicated claims resonance_amplification_e5 (Hi3 PASS), Hi3_recency (PASS), and Hi3_arc (FAIL).

2.6 Controlled benchmark scope and confirmation bias¶

The realistic_recall_v2 benchmark (50 scenarios, 200 queries, 5 challenge types × 40) was designed to evaluate AFT on scenarios where affective signals are relevant. Four of the five challenge types are by construction pro-AFT:

affective_arc — valence discriminates the emotionally salient target from neutral alternatives
momentum_alignment — emotional momentum is the primary differentiating signal
semantic_confound — surface cosine fails specifically where affect differs
same_topic_distractor — same-topic distractors are partly rejected via affect

Only recency_confound is valence-neutral (the discrimination criterion is temporal, not affective). The aggregate advantage (Δ=+0.205 SBERT, Δ=+0.155 e5) should therefore be read as "AFT advantage when affective context is discriminative", not as general superiority across all retrieval scenarios.

This is complementary to the LoCoMo negative result (Gate 1 FAIL, F1=0.168 vs 0.271): on factual open-domain QA where affect is not discriminative, AFT provides no advantage. The two results are consistent — AFT's benefit is scope-dependent. Scenarios where valence and arousal do not distinguish the target from distractors are under-represented in v2; their effect on aggregate top1 is unknown.

2.7 Human-eval pipeline not yet run with real ratings¶

The repository now ships an executable pipeline in benchmarks/human_eval/ that generates packets, rating templates, and aggregated summaries, with a pilot v1 locked on aft vs naive_cosine across 10 scenarios. This closes the procedural gap but not the empirical one:

No real raters included in the repository: the pilot has not yet been run with human participants.
Empty templates do not count as evidence: the pipeline now rejects ratings.jsonl files left in placeholder state.
No checked-in summary counts as a result: summary.json and summary.md are not part of the evidence surface until the pilot is run with completed ratings.
Inter-rater agreement is wired but not yet measured: Krippendorff's alpha (ordinal) is computed automatically when ratings.jsonl contains ratings from at least 2 raters. The target is alpha ≥ 0.67 with 3+ raters; no real ratings have been collected yet.

2.8 LoCoMo per-task weight tuning closed (Add. J Hj1 FAIL)¶

Following the LoCoMo Gate-1 negative result (G2 FAIL, F1=0.168 vs naive_rag=0.271 on the full benchmark), Addendum J (benchmarks/preregistration_addendum_j.md) pre-registered a per-category Pareto sweep over 10 weight configurations × a 200-QA stratified subsample (seed=42) to test whether per-task base_weights tuning could close the gap. Hj1 FAIL (2026-05-06): no AFT configuration reaches naive_rag parity on any of the four LoCoMo categories (single_hop, multi_hop, temporal, open_domain).

The best configuration (W2: sem=0.50, mood=0.30, no resonance) improves over the AFT default by Δ_F1=+0.044 (0.1765 vs 0.1323) and over W0 by Δ_judge_acc=+0.085, but still trails naive_rag (F1=0.2092) by Δ=−0.033. Aggregating across 10 sweeps and 4 categories produced no Pareto frontier where AFT dominates the baseline, even on categories where affect could plausibly help (temporal arc-like queries).

This forecloses the per-task weight-tuning hypothesis as a path to closing G2. The remaining options are architectural rather than parametric:

Per-category routing: classify queries by type and switch between AFT and naive cosine retrieval based on whether affect is discriminative.
Query-type classifier: detect "factual open-domain" vs "affective arc" queries upstream of the 6-signal scorer and downweight non-discriminative signals dynamically.
Hybrid retrieval: combine naive top-k with AFT re-ranking only when the affect signal exceeds a confidence threshold.

These are larger architectural changes outside v0.10 scope. Full-N (all 1500 LoCoMo queries) replication of W2 was not warranted given the 200-QA result already closes Hj1. See benchmarks/locomo/pareto_results.md and benchmarks/preregistration_addendum_j_closure.md for the full numerical record.

2.9 Cross-seed robustness: characterized (retrieval is near-deterministic)¶

The confirmatory studies pin a single random seed each (ablation seed=0, Hi3 seed=1, the Pareto sweep seed=42), and the confidence intervals reported throughout are bootstrap CIs resampled within a single run, not variance across seeds. To check whether that single-seed convention hides cross-run instability, benchmarks/realistic/multiseed_runner.py (make bench-multiseed) re-runs the realistic replay benchmark across seeds {0, 1, 7, 42, 123}, each in an isolated subprocess invoking the canonical runner, and reports the cross-seed mean/stdev/min/max of top1_accuracy and of the AFT−baseline Δ.

Result (benchmarks/realistic/multiseed_results.md, realistic_recall_v2, hash embedder): retrieval is near-deterministic — not exactly deterministic. The RNG seed itself moves nothing; the genuine residual stochasticity lives in (a) dataset generation seeds (frozen and committed) and (b) bootstrap RNG (already reflected in the reported CIs). But the per-query top-1 outcome is not bit-stable across fresh sweeps: repeating make bench-multiseed six times gave retrieval_deterministic=True in 5/6 runs and False in 1/6 (cross-seed stdev 0.0024), and the absolute aft mean drifted across sweeps (0.120–0.125).

An earlier version of this section asserted "cross-seed stdev and spread are exactly 0.0000 — identical across all five seeds." A fresh validation pass falsified that headline; the corrected scope is near-deterministic, with sub-CI, timing-driven variance on near-ties.

The cause is a timing effect, not RNG. The engine stamps encode/retrieve with real wall-clock time (datetime.now(tz=UTC), see engine.py), and ACT-R decay is a function of now − encoded_at. Because each seed's subprocess is launched at a slightly different instant, a ranking that is already at a numerical tie can tip between seeds even with subprocess isolation — the isolation removes RNG coupling but not wall-clock timing. (Running several benchmarks back-to-back inside a single process amplifies the same effect.) This is correct production behaviour (decay should track real time), not a library defect, and the resulting variance stays within the reported bootstrap CIs — AFT still clears naive_cosine by far more than the spread (Δ ≈ +0.075). A fully time-deterministic benchmark would require threading an injected clock through encode/retrieve; that is deliberately not done here, as it would expand the core API and disturb the 127 fidelity benchmarks for a sub-CI, benchmark-only effect. Finally, the sweep would surface larger variance for a stochastic embedder or a regenerated dataset; extending it to those cases is straightforward future work.

3. Architectural Limits¶

3.1 Appraisal depends on an external LLM¶

LLMAppraisalEngine requires a call to an external LLM (OpenAI-compatible) to produce an AppraisalVector. This introduces:

Latency: 200–2000 ms per encoding in slow-path mode. Batch throughput is mitigated by encode_batch, which appraises items concurrently (bounded by appraisal_max_concurrency, default 8) rather than serially; per-item latency is unchanged. dual_path_encoding defers appraisal off the encode hot path entirely.
Cost: dependent on provider and model.
Non-determinism: the same text can produce different appraisals across successive calls.
Network dependency: encoding does not work offline without the KeywordAppraisalEngine fallback.

The KeywordAppraisalEngine is a rule-based fallback with limited coverage (~50 keyword classes). It does not generalize to open domains.

Note (v0.18): its per-dimension averaging divided each dimension by every matching rule, including rules that left that dimension at its neutral default, so multi-rule inputs came out systematically attenuated ("I succeeded at the project" → goal_relevance 0.35 instead of 0.70). Only contributing rules are counted now. Every published keyword-engine measurement — the Addendum S human-gold figure (valence r=0.07) and the aft_keyword_synchronous ablation arm (Hf1) — was produced with the diluted engine and would need re-running before being read as a post-fix result. No committed benchmark artefact was regenerated for this fix; the LLM-appraisal results are untouched.

For an engineering cost budget (embedder vs appraisal vs 6-signal scoring), store choice, and measured retrieve breakdowns, see Performance & Scaling.

AffectiveState is no longer only in-process. The repository now supports dedicated AffectiveStateStore implementations, with local backends such as InMemoryAffectiveStateStore and SQLiteAffectiveStateStore, plus an optional RedisAffectiveStateStore backend for shared state.

Important architectural limits remain:

No production-grade validation of the shared backend: Redis is available as an optional backend, but comprehensive validation on real multi-worker deployments is still missing.
No joint memory+state transaction: MemoryStore and AffectiveStateStore remain separated by design; there is no atomic cross-store semantics yet.
Limited distributed synchronization: the Redis backend covers affective-state persistence, not distributed vector memory.

3.3 Thread-safety limited to a single connection¶

SQLiteStore uses a single sqlite3.Connection object with threading.RLock. This serializes accesses but does not allow reader parallelism. For high throughput on a SQLite store, a connection-pool architecture would be more scalable.

3.4 No "enterprise" vector store¶

The available adapters are InMemoryStore (RAM, non-persistent; full-scan cosine with a cached embedding matrix), SQLiteStore (local file, scalable to ~10^6 memories via sqlite-vec ANN), QdrantStore (v0.9), and ChromaStore (v0.9). There are no adapters for Weaviate or Pinecone. The MemoryStore Protocol is duck-typed and contributions are welcome. Decision tree: Performance & Scaling.

4. Theoretical Limits¶

4.1 Partial operationalization of Heidegger¶

MoodField operationalizes Stimmung (background emotional tonality) as an EMA over valence–arousal with decay parameters. This is a necessary computational simplification of a phenomenological concept that in Heidegger is pre-cognitive and structurally tied to being-in-the-world. The mapping is inspired, not faithful.

4.2 AFT novelty credentials are not peer-reviewed¶

As of the v0.9 release, Affective Field Theory has not been formally published in a peer-reviewed venue. Originality claims (in particular the unified integration of Russell + Scherer + Heidegger + Hebb + Collins & Loftus in a single computational model for LLM memory) are plausible but have not yet been validated by the scientific community.

5. Future Work¶

Recently shipped (v0.9)¶

Feature	Status
Qdrant vector-database adapter (`[qdrant]` extra)	✅ shipped v0.9
ChromaDB adapter (`[chroma]` extra)	✅ shipped v0.9
OpenTelemetry spans (`[otel]` extra)	✅ shipped v0.9

Open limits (not tied to a specific release)¶

Limit	Indicative horizon
Broader, comparative affect-aware realistic benchmark	ongoing research
Execute the human-eval pilot with real ratings	research track
BYO appraisal schema (OCC, GRID, custom taxonomies)	v0.10

Document added in v0.5.1. Last updated: v0.17.x (2026-07-14, Addendum Z closure).