Beyond Recall: Behavioral Specification as an Interpretive Layer for AI Personalization

Appendix E. Per-subject worked examples 129

• The document lists per-subject worked examples and related benchmark and exploratory appendices, including the Letta stateful-agent case study.The supplied material identifies Appendix G as covering Letta’s architectural difference and headline result, but provides no worked-example findings here.Source: Appendix E · pp. 4–5, §1.3 · p. 13

1.1 Recall is not interpretation. Interpretation can be measured.

• Representational accuracy measures how well an AI system captures a specific person’s interpretive patterns, distinct from recall, preference matching, or persona consistency.The paper defines interpretation as how a person processes facts and experiences into judgments, decisions, and reactions.Source: §1.1 · p. 6

1.2 What we tested

• The study tests representational accuracy by asking a response model to predict held-out situations and scoring its answers against the subject’s verbatim responses with a calibrated 5-judge LLM panel on a 1–5 interpretive rubric.Each autobiography is split in half: the training half generates the Specification and seeds memory systems, while the held-out half supplies questions never shown to the response model.Source: §1.1 · p. 6, §1.2 · pp. 7–8
• The benchmark evaluates 14 public-domain autobiographies under multiple context conditions, including a Behavioral Specification, extracted facts, raw corpus, and memory-system retrieval in controlled and native configurations.The Specification is typically a structured document of around 7,000 tokens (~5,000 words), and the benchmark is presented as a prototype rather than a finished standardized instrument.Source: §1.2 · pp. 7–9

1.3 What we found

• Every one of 9 low-baseline subjects improved when a Behavioral Specification was added to All Facts, with a +0.89-point mean lift over No-Context Baseline and 78.6% of 351 paired questions improving.The score uses a 1–5 rubric, and low-baseline subjects have No-Context Baseline scores of C5 ≤2.0.Source: §1.2 · p. 10, §1.3 · p. 11
• The Behavioral Specification recovers 75% of raw-corpus predictive value at approximately 25× less context, while lifting 55% of low-baseline questions across at least one rubric anchor.A 7,000-token specification matches most of the predictive accuracy of a 163,000-token raw corpus; 18% of low-baseline questions cross two or more anchors and 5.7% cross three or more.Source: §1.3 · p. 11
• The Specification reduces hedging from 41.2% to 0.4% and layers onto three of four commercial memory systems, improving interpretation-heavy questions but sometimes harming literal recall.The wrong-Spec control continues to hedge or flag mismatch on 60.6% of responses; Mem0, Letta, and Zep show positive mean lift under at least one configuration, while Supermemory does not.Source: §1.3 · p. 12
• Retrieval systems diverge substantially in which facts they surface, despite similar recall-benchmark performance, while the Specification makes behavioral inference structured and inspectable.Providers share zero top-10 facts on 35.9% of question-pairs, with 8.3% mean pairwise overlap; on 65.6% of system-pair/question instances they share one or fewer facts.Source: §1.3 · p. 12

1.4 What this implies

• Representational accuracy is distinct from recall, and human-AI alignment depends on accurately representing the user through portable, inspectable, traceable, representation-grade infrastructure.A Behavioral Specification is presented as one implementation of this infrastructure, not the only one.Source: §1.4 · p. 13
• The gap cannot be closed by larger models trained on more public data because private records do not reliably capture how a specific person reasons, making the relevant population anyone who uses AI.Even autobiographical subjects whose work appears in pretraining score near the rubric floor without context.Source: §1.4 · p. 13

2. Prior Work, Industry Benchmarks, The Fifth Target

• Representational accuracy measures whether an AI models how a specific person reasons, complementing recall, survey-response prediction, persona fidelity, and preference alignment.Behavioral prediction on held-out reasoning situations operationalizes this fifth target rather than treating prediction as the ultimate goal.Source: §2 · p. 14, §2.1 · pp. 15–16

2.1 Prior measurement targets and the gap representational accuracy fills

• Existing benchmarks measure distinct properties: recall retrieves facts, Twin-2K predicts questionnaire responses, PersonaGym preserves surface character, and AlpsBench evaluates preference-aligned responses.These properties do not establish whether a system transfers a person’s reasoning to new situations; recall is necessary for memory uses but insufficient for representational accuracy.Source: §2.1 · pp. 14–15

2.2 Memory systems for LLM agents

• Commercial memory systems report 70–93% recall scores, but their architectures target storing, organizing, and retrieving user content rather than modeling individual interpretation.Mem0, Letta, Supermemory, and Zep provide sophisticated retrieval capabilities, while the Behavioral Specification targets the interpretive layer above retrieval.Source: §2.2 · p. 17

2.3 Traceability and reasoning traces

• Reasoning-level traceability lets users audit why the system believes something about them by tracing interpretive claims through grounded facts back to source passages.Fact-level provenance answers where content came from; reasoning-level traceability supports challenging and correcting the representation itself.Source: §2.3 · pp. 17–18

2.4 Cognitive and representational foundations

• The study treats the Behavioral Specification as a schema-like compression whose validity depends on stable interpretive patterns generalizing from training text to held-out situations.The design uses a 50/50 train/held-out split; subjects whose reasoning changes substantially across their corpus may not fit one snapshot specification.Source: §2.1 · p. 16, §2.4 · p. 18
• A Behavioral Specification is framed as a per-user external override to the Assistant Axis, shifting behavior from generic helpfulness toward reasoning as the specific person would reason.This framing motivates measuring hedging alongside accuracy, because movement away from the generic Assistant Axis should affect both predictions and willingness to commit.Source: §2.4 · p. 19

3. Study Design

• The study holds the response model constant while varying context conditions, enabling comparison against a No-Context Baseline and isolation of the interpretive layer’s contribution.Conditions include nothing, retrieved facts, raw corpus, a Behavioral Specification, and combinations; statistical commitments were pre-locked before final analysis.Source: §3 · pp. 19–20

3.1 Operationalizing representational accuracy via the Behavioral Specification

• Representational accuracy measures how faithfully an AI system captures a specific person’s reasoning patterns, independently of any particular implementation.The study tests a Behavioral Specification served as context to a static response model against alternative context conditions.Source: §3.1 · p. 20

The instrument we use to measure representational accuracy is behavioral prediction

• Held-out behavioral prediction tests whether consistent subject patterns are captured by the Behavioral Specification and used by the model.A low score indicates failure of behavioral consistency, Specification quality, or model use, and each failure mode is informative.Source: The instrument we use to measure… · p. 20

3.2 Experimental conditions

• The experiment varies served inputs—from no context, facts, and raw corpus to the Specification and combinations—to isolate each source of representational signal.Memory systems are evaluated in controlled and native modes, separating retrieval-algorithm effects from ingestion-pipeline effects across five systems.Source: The instrument we use to measure… · p. 20, §3.2 · pp. 21–23

3.3 Scoring rubric with calibrated LLM judge panel

• Responses receive 1–5 scores against verbatim held-out passages from a five-judge primary panel, with seven judges used for sensitivity analysis.The prototype instrument substitutes expanded judging and calibration diagnostics for human annotation at this scale.Source: §3.3 · pp. 23–24, §3.3.2 · p. 26

3.3.1 Score interpretation

• Rubric scores represent categorical prediction quality, from refusal or wrong prediction through subject-specific direction to a specific behavioral outcome.Fractional changes are interpreted more strongly when they cross an integer anchor; a score of 5 aligns with one held-out behavioral sample, not complete representation of the subject.Source: §3.3 · p. 24, §3.3.1 · pp. 25–26
• Roughly 18% of paired questions show same-anchor fractional shifts of at least 0.5 rubric points, a weaker signal than cross-anchor movement.The primary integer metric is reserved for categorical interpretation, while within-anchor shifts are reported for methodological transparency.Source: §3.3.1 · p. 26

3.3.2 Judge panel

• The panel tests whether serving a Behavioral Specification shifts responses toward each subject’s demonstrated behavioral patterns, measuring directional representational accuracy rather than absolute correctness.Response generation and judging are blinded to condition labels and generating models; judges receive the held-out passage, subject context, question, and response.Source: §3.3.2 · p. 27

3.3.3 Calibration

• Six of seven judges score verbatim matches at 5.0, while Gemini Pro scores 4.15 and shows severe length sensitivity; the five-judge primary aggregate is therefore the conservative estimate.The primary panel comprises Haiku, Sonnet, Opus, GPT-4o, and GPT-5.4; Gemini Flash and Gemini Pro are retained as a seven-judge sensitivity check rather than included in the primary.Source: §3.3.3 · pp. 28–29

3.3.4 Inter-judge agreement

• Primary judges agree strongly on condition rankings (Spearman ρ = 0.86 to 0.93) but less precisely on absolute scores (Krippendorff α = 0.659), supporting directional rather than absolute-score inference.The agreement holds under a deliberately spare rubric, while the panel does not establish that a higher-scoring response is absolutely correct for the subject.Source: §3.3.4 · pp. 29–30

3.3.5 Aggregation and statistical analysis plan

• The locked analysis averages question scores within each subject and condition, compares facts-plus-Spec C4a against the No-Context Baseline C5, and tests the 14 paired differences with a two-sided Wilcoxon signed-rank test.The five-judge aggregate is primary, the seven-judge aggregate is a robustness check, and all other analyses are descriptive rather than additional primary inferential tests.Source: §3.3.5 · pp. 30–31

3.3.6 Rubric-handling limitations (post-hoc validity audit)

• The validity audit finds that refusal-anchor ambiguity and baseline response length can inflate No-Context scores, implying the true Spec-versus-baseline gap may exceed the reported +0.89 mean lift.In C5, response length correlates with score at r = 0.60, whereas this relationship disappears when a Specification or fact set is supplied.Source: §3.3.6 · p. 31

3.4 Subjects

• The study tests 14 historical figures using English or translated public-domain autobiographies and memoirs selected across periods, source lengths, and geographic origins, although pretraining exposure is likely.The corpora are available through Project Gutenberg or comparable public-domain archives.Source: §3.4 · p. 32

3.4.1 Pretraining-coverage variance (high vs low baseline)

• Pretraining coverage is proxied by each subject’s no-context mean score across 39 behavioral-prediction questions, with higher scores indicating greater prior representation.The response model is held constant because cross-provider baseline variation for a subject can be comparable to the Specification’s lift.Source: §3.4.1 · p. 33

3.4 Subjects

• The sample comprises 14 autobiographical subjects, with generalizability limited by language, era, cultural framing, curation bias, and individual self-presentation.Nine subjects fall in the low-baseline band and five in the mid-baseline band; Benjamin Franklin serves as a high-baseline reference control.Source: §3.4 · p. 32, §3.4.1 · p. 33

3.5 Question battery formation

• Each main-study subject receives 39 backward-designed behavioral questions generated from held-out text, covering 10 categories while avoiding named-entity and date leakage.The 14 main-study batteries contain 546 questions, and the leakage audit found zero leaks across them.Source: §3.5 · p. 34

3.5.1 Circularity controls

• Two controls test circularity: GPT-5.4 regenerates batteries, while non-Anthropic response models rerun core conditions on three subjects.The broader LLM-as-judge circularity remains an acknowledged limitation.Source: §3.5.1 · p. 35

3.6 Response models

• Claude Haiku 4.5 is the primary response model, and a shared faithfulness-oriented prompt varies only the served context across experimental conditions.Conditions provide no context, Specifications, facts, corpora, combinations, retrieval, or wrong-person Specifications; hedging and refusal behavior are measured rather than instructed.Source: §3.6 · pp. 35–36

3.7 Pipeline for the Behavioral Specification

• The Behavioral Specification transforms source text through import, extraction, embedding, authoring, and composition into approximately 7,000 tokens of anchors, core, predictions, and a unified brief.A fixed 46-predicate vocabulary steers extraction from biographical facts toward behavioral patterns, and the served context concatenates the three authored layers with the composed brief.Source: §3.7 · p. 37
• Behavioral predictions are explicit predicates with detection criteria, directives, and false-positive warnings, translating interpretation into testable behavioral expectations.An example predicate specifies when a subject confesses before reaching a conclusion, how to detect it, and when not to apply it.Source: §3.7 · p. 38
• The unified brief integrates anchors, core prose, and predictions into a continuous third-person synthesis that is served alongside—not instead of—the three layers.Internal testing suggests this integration changes how a model uses the Specification, while a formal brief-versus-layers ablation is separately flagged.Source: §3.7 · p. 38
• Approximately $1 per subject processes a 50,000- to 150,000-word autobiography end to end, with pipeline code, predicate vocabulary, and 14 example specifications publicly available.The cost is estimated from per-step API token counts.Source: §3.7 · p. 38

4.1 The cross-subject gradient and its per-question mechanism

• +0.89 points was the mean per-subject lift for 9 low-baseline subjects under All Facts + Spec, with every subject improving and 55.0% of responses crossing an anchor upward.Spec Only lifted means by +0.69; the Specification’s value was in layering over facts or retrieval rather than replacing them.Source: §4.1 · pp. 39, 41
• The Specification’s lift was largest when pretraining baselines were low, with a change-score regression slope of −0.96, while high-baseline Franklin declined by −0.13.Low-baseline subjects ranged from +0.25 to +1.51, whereas the final C4a means clustered around 2.44, indicating roughly uniform resulting quality.Source: §4.1 · p. 40, Worked examples: Wrong-Referent… · p. 46

Worked examples: Wrong-Referent (Fukuzawa), Directional (Bernal Díaz), Abstention (Seacole)

• For Fukuzawa, All Facts + Spec corrected a wrong referent, shifting the mean from 1.00 to 4.20 by using earned capability and institutional skepticism to select Katsu over Kimura.Facts matched both naval officers superficially; the Specification supplied the behavioral axes needed to identify the person Fukuzawa would characterize.Source: Worked examples: Wrong-Referent… · pp. 42–43
• For Bernal Díaz, the Specification flipped the prediction from accepting assistance to Cortés refusing it, raising the mean from 2.00 to 4.80.The interpretive layer encoded performative self-reliance as the organizing principle of Cortés’ authority, which the complete fact set did not itself announce.Source: Worked examples: Wrong-Referent… · p. 43
• For Seacole, the Specification converted fact-based abstention into unanimous near-perfect inference, raising the mean from 2.80 to 5.00 by predicting compassionate acceptance without correcting the patient.It captured the specific pattern of sustaining mistaken identity with tenderness, rather than merely the generic fact that Seacole was compassionate; abstention scoring was judge-dependent.Source: Worked examples: Wrong-Referent… · p. 44

4.1.1 Per-question baseline engagement and the worked rubric example

• The Specification’s lift is strongly concentrated on low-baseline questions, with mean lift +1.32 for the bottom cluster but −0.99 for questions with baseline scores at least 4.00.Baseline score and Specification lift are negatively associated (Spearman ρ = −0.73, n = 546, p ≈1.7 × 10−91).Source: §4.1.1 · p. 47
• The worked Seacole example shows the full pipeline combining facts with interpretive patterns to reach the held-out behavioral direction unanimously across all five judges.Spec Only predicts non-correction using Proximate Obligation and Sympathetic Realism, while no-context and facts-only conditions abstain.Source: §4.1.1 · pp. 48–49

4.1.2 The gradient at the high-baseline end (Franklin reference)

• At the high-baseline end, Franklin’s 3.77 No-Context mean falls by 0.40 with Spec Only and 0.13 with facts plus Spec, despite improvement on 15 of 40 questions.Positive cases cluster where the autobiography lacks verbatim narration, while negative cases cluster where pretraining already supplies the behavioral pattern.Source: §4.1.2 · p. 49

4.2 Compression: structure vs. raw text

• The Behavioral Specification recovers 75% of the raw corpus’s predictive lift at approximately 25× less context: +0.68 versus +0.91 points over baseline.Spec Only uses approximately 7K tokens, while the raw corpus averages approximately 163K tokens across nine low-baseline subjects.Source: §4.2 · p. 51

4.2.1 Per-question improvement rate

• On nine low-baseline subjects, the Specification improves roughly 70% of questions, within eight percentage points of the raw corpus’s 78.3% improvement rate.The Specification achieves this rate with approximately 7K tokens versus approximately 163K for the raw corpus.Source: §4.2.1 · p. 52

Spec only

• The Spec produces its largest categorical movements when prior context is sparse: multi-anchor crossings occur on 9–15% of No-Context additions but only 2–3% when layered onto full facts or corpus.These rates show that small aggregate deltas can conceal substantial per-question movement in both directions.Source: Spec only · p. 53
• 70.9% of questions improve with Spec alone at roughly 25× less context than the raw corpus, while corpus plus Spec reaches the highest rate at 83.7%.The median improvement is +1.00 rubric points when the Spec helps and −0.40 when it hurts; All Facts + Spec matches the raw corpus’s improvement rate.Source: Spec only · pp. 53, 55
• Hamerton Q25 shifts from refusal to grounded prediction only under Corpus + Spec, which combines the Spec’s behavioral axioms with autobiographical evidence to project a pattern onto a novel scenario.Five of six conditions refuse to engage substantively because the specific scenario is absent from the training-half corpus; Corpus + Spec moves from anchor 1 to anchor 4.Source: Spec only · p. 56

Compression examples: Hamerton (Spec succeeds), Ebers (Spec fails)

• Hamerton’s Spec alone scores 2.63 versus 2.27 for the raw corpus, while Corpus + Spec reaches 3.09, the study’s highest compression-related score.The Spec-alone lift over baseline is +1.37 versus +1.01 for the corpus, indicating complementarity rather than overlap.Source: Compression examples: Hamerton (Spec… · p. 57
• Ebers’s Spec alone scores 1.54 versus 2.18 for the raw corpus, a 0.64-point gap; the raw corpus moves the average response from anchor 1 to anchor 2, while the Spec stays at the floor.The gap reflects autobiographical specificity that the Spec abstracts into axioms by design.Source: Compression examples: Hamerton (Spec… · p. 57
• The Specification’s 7K-token footprint recovers 75% of the corpus’s predictive benefit at 4% of the context cost, making personalization tractable where full autobiographies cannot fit per request.The remaining approximately 25% is achievable only at production-prohibitive context sizes of roughly 100,000–400,000 words.Source: Figure 4.2 · p. 54, Compression examples: Hamerton (Spec… · p. 58

Wrong-Spec examples: Ebers Q7 (identity), Bernal Díaz Q16 (frameworks), Seacole Q2 (inference)

• When content does not fit, two of three examples show drops of −2.00 to −3.60 points; the third shows a near-zero drop because the wrong specification predicts the same surface behavior.Source: Wrong-Spec examples: Ebers Q7… · pp. 63–64

4.4.1 Cross-system retrieval: providers do not converge

• 8.3% mean pairwise Jaccard shows that memory systems rarely retrieve the same facts, with native pipelines producing 0.000 exact overlap across all four commercial pairs.Divergence persists under semantic matching: native overlap rises only to 0.004 at cosine ≥0.85 and 0.016 at cosine ≥0.70, indicating provider-specific ranking choices.Source: §4.4.1 · pp. 64, 67, Figure 4.4.1 · p. 66

4.4 Memory-system composition

• The Specification layers positively on three of four commercial systems, with Wilcoxon direction confirmed at α = 0.01 for four system-configuration cells.Every system lifts 18–37% of questions by at least one rubric anchor, while aggregate means conceal categorical gains and losses across individual questions.Source: §4.4.2 · pp. 67–68

4.4.2 Layering the Spec: aggregate ∆across systems and ingestion paths

• +0.33 aggregate Δ_spec is reached by both Mem0 and Zep under native ingestion, while Letta falls from +0.20 controlled to −0.02 native and Supermemory remains near zero.Across the 14 main-study subjects, three of four commercial systems are positive under at least one configuration; the Spec helps interpretation-heavy questions but can hurt literal-recall cases.Source: §4.4.2 · pp. 67–68

4.4.3 Where the Spec helps, where it hurts, and which question types route to each

• The Specification exhibits three recurring interactions with retrieval: interpretive supply improves underdetermined answers, over-theorization harms already-resolved questions, and axioms can induce refusals whose rubric impact depends on correctness.Interpretive supply composes with grounded evidence, whereas the Specification alone may abstain; over-theorization pulls literal questions toward unnecessary depth, and refusal can lower measured scores even when principled.Source: §4.4.3 · p. 69
• Across systems, retrieval architecture shifts the balance: Supermemory’s near-zero aggregate ∆_spec reflects helps and hurts roughly canceling, while lean Base Layer retrieval leaves the Specification carrying more interpretive load.Mem0’s atomic facts favor interpretive supplementation when needed but can be abstracted away on literal recall; Letta’s few unique archival facts produce large gains when aligned; Zep’s relational context yields the most balanced distribution.Source: §4.4.3 · pp. 69–70

4.4.4 Case study: cross-system refusal on Keckley Q21

• On Keckley Q21, the Specification raised the evidentiary bar for inner-motive prediction, converting productive Supermemory and Base Layer answers of 3.6 and 3.4 into refusals scoring 1.6 and 1.0, while adding no measurable penalty where retrieval was already weak.The documented-dignity axiom correctly discouraged speculation about a motive absent from the retrievable corpus, but the content-match rubric penalized refusal regardless of whether it was epistemically appropriate.Source: §4.4.4 · pp. 73–74

4.5 Exploratory case study: Letta stateful-agent (N=3, post-hoc)

• In a post-hoc three-subject comparison, Letta’s self-edited memory block scored higher than the matched Behavioral Specification by ∆+0.27, +1.21, and +0.38, while stacking both representations outperformed either alone on all three subjects.The Letta block reached ~335K characters at the largest corpus, with substantial verbatim and semantic duplication, indicating a scale ceiling not applying to the Specification.Source: §4.5 · p. 75
• The exploratory Letta comparison supports architectural convergence on interpretive representation, not a magnitude ranking, because it is post-hoc, limited to N=3, and partly affected by abstention artifacts.Multi-subject replication across the full 14-subject gradient remains future work.Source: §4.5 · p. 75

4.6.1 Cross-provider response generation (Tier 2 replication)

• The 7-of-9 positive Specification lifts replicate across three response models and two providers, with both non-positive cells occurring for highest-baseline Zitkala-Ša.The test used GPT-5.4-regenerated batteries and response models Haiku 4.5, Sonnet 4.6, and Gemini 2.5 Pro; the two non-positive cells were Haiku × Zitkala-Ša at −0.32 and Gemini Pro × Zitkala-Ša at −0.03.Source: §4.6.1 · p. 76

4.6.2 Judge panel sensitivity (5-judge primary vs 7-judge)

• The 5-judge primary panel is conservative: adding Gemini judges widens effect magnitudes without changing any improvement direction or subject-level Specification-lift sign.The wrong-Spec fixed derangement remains negative, shifting from −0.25 with five judges to −0.21 with seven.Source: §4.6.2 · p. 77

4.6.3 Battery composition sensitivity

• The headline gradient survives battery-composition and Hamerton-leverage controls, with baseline slope attenuation of about 8% and 7%, respectively.Controlling literal-recall fraction changes the slope from −0.96 to −0.88, while dropping Hamerton changes it from −0.96 to −0.89; neither control overturns the headline.Source: §4.6.3 · pp. 77–78

4.6.4 Statistical-rigor checks on the headline gradient

• The −0.96 gradient survives bootstrap resampling, joint confound regression, and permutation testing, including a bootstrap 95% CI of [−1.25, −0.74] and empirical p < 0.0001.The joint model estimates −0.870 [95% CI −1.136, −0.604], p = 2.6 × 10−5, while 0 of 10,000 permutations reach the observed slope.Source: §4.6.4 · p. 78

4.6.5 Wrong-Spec derangement protocol sensitivity

• Wrong-Spec reductions remain directionally robust: adversarial pairing yields ∆−0.25 and random derangement ∆+0.15, both below matched correct-Spec ∆+0.35.Across protocols, mismatched specifications reduce representational accuracy, while reduction magnitude depends on content mismatch; per-subject effects vary with coincidental overlap.Source: §4.6.5 · p. 79

4.6.6 Retrieval-overlap sensitivity (semantic-similarity matching, K variation)

• Retrieval divergence persists under semantic-similarity matching: mean pairwise soft Jaccard never exceeds 0.30, and native retrievals reach only 0.004 at the calibrated near-duplicate threshold.The finding holds across tested thresholds and K=5 or K=10; controlled mean Jaccard rises from 0.083 exact to 0.102 at the calibrated near-duplicate threshold.Source: §4.6.6 · p. 80

4.6.7 Rubric-handling limitations (post-hoc validity audit)

• The post-hoc rubric audit finds refusals and verbosity over-credited, both raising No-Context Baseline scores more than Specification scores and making the +0.89 mean lift conservative.Among 192 low-baseline refusals, 9.4% scored at least 2.0 and 3.1% at least 3.0; across 1,599 responses, length-score correlation was r = 0.60 for C5 but near-zero when facts were served.Source: §4.6.7 · pp. 80–82

4.7 Summary of §4

• The Spec recovers 75% of the corpus’s predictive benefit at 4% of the context cost, while correct-Spec conditions reduce hedging from 41.2% to 0.4%.Wrong-Spec pairings reduce accuracy below No-Context Baseline (∆= −0.25), whereas random derangements barely improve on no context (+0.15).Source: §4.7 · p. 83
• The Spec’s lift is largest where pretraining baselines are lowest, concentrates on interpretation-required questions, and adds little or hurts on literal-recall and refusal-triggering questions.The gradient also holds at the high-baseline end through the Franklin reference.Source: §4.7 · p. 83
• Memory systems share zero top-10 facts in 35.9% of instances, with 8.3% mean pairwise overlap, showing that providers diverge on relevance despite identical fact pools.The Spec layers cleanly on three of four commercial memory systems, with effects varying by retrieval architecture.Source: §4.7 · p. 83
• Sensitivity checks reproduce the direction on 7 of 9 cross-provider cells, preserve panel-level signs, and support the gradient, wrong-Spec, and retrieval-divergence findings.The checks do not replace human validation or resolve the class-level concern that every response model and judge is an LLM.Source: §4.6.8 · p. 82, §4.7 · pp. 83–84

5.1 Synthesis: What The Seven Findings Together Establish

• The Behavioral Specification improves subject-specific behavioral prediction where pretraining is insufficient, complements other context types, and recovers most corpus accuracy at lower context cost.It nearly eliminates hedging, while three of four commercial memory systems show positive lift under at least one configuration.Source: §5.1 · pp. 84–85

5.2 Why the Gradient is the Load-bearing Finding

• The Spec’s benefit follows a gradient: it helps when pretraining lacks an interpretive frame, adds little when pretraining suffices, and can interfere when the model already knows the answer.This makes dynamic activation more appropriate than attaching the layer to every query.Source: §5.2 · p. 85

5.3 Retrieval is not Interpretation

• Mean pairwise top-K overlap across four commercial memory systems is 8.3% (Jaccard 0.083), showing that retrieval does not identify a unified fact set for behavioral questions.Different facts and interpretive frames can produce divergent outcomes, so representational accuracy measures an axis that recall benchmarks omit.Source: §5.3 · pp. 85–86

5.4 Composition with Retrieval

• Retrieval and interpretation must be composed selectively, with runtime serving deciding which Specification components arrive when rather than relying on static attachment.The paper’s static regime is a conservative baseline; enforcing appropriate use and measuring each context component’s causal contribution remain open problems.Source: §5.4 · p. 86

5.5 Wrong-Spec Mechanism and Hedging Elimination

• A matched Spec raises accuracy by ∆+0.35 above baseline, whereas an adversarial wrong-Spec lowers it by ∆−0.25; random derangement still yields ∆+0.15.The matched Spec also reduces baseline hedging from 41.2% to 0.4%, although the mechanisms governing mismatch detection and commitment remain unclear.Source: §5.5 · p. 87

5.6 Compression and What Makes Personalization Operationally Tractable

• The raw corpus produces the highest prediction scores, while the Specification is roughly 25× smaller and recovers about 75% of the corpus lift over baseline.This suggests that effective personalization depends on selecting an interpretive frame, not merely expanding context windows.Source: §5.6 · pp. 87–88

5.7 Privacy and the Case for User-held Representation

• User-held representations align incentives by allowing individuals to own and correct how they are represented, while third-party-held representations risk unauthorized behavioral modeling and weaponization.Operational deployment additionally requires provenance, authority, and an industry-level trust and traceability layer.Source: §5.7 · p. 88

5.8 Closing Remarks

• The paper establishes a deployable proxy for representational accuracy, making claims about alignment with specific individuals testable through behavior on unseen text.Its validity remains directional rather than absolute because human annotation has not yet validated the measure at scale.Source: §5.1 · p. 85, §5.8 · p. 89

6. Limitations

• The limitations are permanent caveats on interpreting the experimental results, distinct from the follow-up experiments proposed in §7.They constrain how the paper’s claims should be read across the subject sample, measurement apparatus, pipeline stability, and exploration scope.Source: §6 · p. 89

6.1 Subject sample

• The 14 historical autobiographical subjects limit generalization across era, language, self-presentation, genre, and digital or non-narrative source material.The sample is drawn from public-domain writings shaped by preservation, translation, and authorial curation, while held-out data remain within the same authored narratives.Source: §6.1 · pp. 89–90

6.2 Measurement apparatus

• LLM-as-judge circularity and prompt dependence mean the headline findings provide directional, not precise, evidence and may vary with prompts or judge panels.The all-LLM pipeline does not establish absolute quality against human evaluators; judge rank order is stable, but score magnitude is panel-specific.Source: §6.2 · pp. 90–91

6.3 Pipeline and specification stability

• 0.10 pooled run-to-run SD versus 0.59 cross-subject SD supports the gradient, but individual per-subject effects carry roughly ±0.10 uncertainty and vary across specifications.The variance probe covered only three subjects and held extraction state constant, so extraction-stage and high-baseline variance remain untested; pipeline model choices were also not varied.Source: §6.3 · pp. 92–93

6.4 Scope of exploration

• The study did not exhaustively test its roughly 6,500-cell design space, leaving per-layer serving, alternate pipeline models, dynamic activation, and broader stateful-agent comparisons for follow-up.The Letta stateful comparison covered three subjects, one version, and one response model rather than the full gradient.Source: §6.4 · p. 93

7.1 Subject and Corpus Expansion

• A multi-subject replication with consenting living users is the leading follow-up, alongside modern-era and non-English corpora and public decision records as alternative testbeds.These studies would test whether findings from 14 historical subjects generalize beyond translated, pre-20th-century autobiographical sources.Source: §7.1 · p. 94

7.2 Measurement Methodology

• Measurement follow-ups would separate interpretation from recall, score epistemic honesty independently, validate rubric applications with humans, test prompt sensitivity, and analyze judge disagreement.The primary panel agrees on direction (Spearman ρ = 0.86 to 0.93) but differs in magnitude (Krippendorff α = 0.659), motivating question-level agreement analysis.Source: §7.2 · pp. 94–95

7.3 Specification design and composition

• Specification research should ablate authored layers, test alternate pipeline and response models, isolate named-entity grounding, and measure whether generated responses preserve the original Specification through round trips.A hybrid representation would pair structured predicates with named entities, while runtime-utilization analyses would test which Specification components models actually use.Source: §7.3 · pp. 95–96

7.4 Production serving and infrastructure

• Production studies should compare dynamic, editable, temporally managed, and domain-partitioned serving with factorial context stacking to identify additive, subadditive, or antagonistic layer interactions.The current study attached the full Specification statically to every query; follow-ups would vary retrieval depth and combinations with facts, memory systems, and the Specification.Source: §7.4 · pp. 96–97

7.5 Stateful-agent implementations and temporal drift tracking

• Stateful-agent work should test online self-editing, controlled Letta reruns, temporal drift, canonical life-event detection, and explicit user corrections with versioned, source-bounded updates.The agenda includes anonymized and larger-corpus Letta comparisons, sequential Specification diffs, and correction-driven updates that preserve the corpus as the source of truth.Source: §7.5 · pp. 97–98

7.6 Safety-alignment integration

• Safety follow-ups should test whether Specification-induced refusals compose with existing safety frameworks and whether models flag Specifications encoding adversarial values rather than comply.A classifier audit found 75 of 81 (93%) Spec-induced refusals were routine behavioral prediction rather than morally loaded, leaving deployment with malicious-intent users untested.Source: §7.6 · p. 98

B.4 Category-level effect size on ∆_spec

• Category effects differed by question type: Literal Recall had ∆_spec +0.792, Interpretive Inference +0.397, and Refusal-Triggering +0.417 across 546 main-study questions.Literal Recall’s larger raw mean is partly a floor effect: 66.7% of its questions had C5 ≤1.5 versus 48.6% for Interpretive Inference, while the Spec’s distinctive mechanism remains interpretive-pattern support where retrieval-only systems hit a ceiling.Source: §B.4 · pp. 109–110

B.6.1 Battery-question-type correlations.

• Across subjects, ∆_spec ranged from −0.31 to +1.37 and correlated with battery composition at r = +0.595 for Literal Recall, r = −0.466 for Interpretive Inference, and r = +0.212 for Refusal-Triggering.Subjects with more Literal Recall questions therefore tended to show larger measured ∆_spec values.Source: §B.6.1 · p. 110

B.6.2 Multiple regression controlling for Literal Recall fraction.

• Controlling for Literal Recall fraction, baseline remained the dominant predictor: its partial coefficient was −0.88 [95% CI −1.13, −0.63], p < 10−5, versus β = +2.30 [+0.34, +4.26], p = 0.026 for Literal Recall fraction.Baseline uniquely explained 63.6% of variance compared with 6.9% for Literal Recall fraction; the gradient also persisted after excluding Hamerton, with slope −0.89 [95% CI −1.18, −0.61], R2 = 0.81, p < 10−4.Source: §B.6.2 · p. 111, §B.6.3 · p. 111

B.6.5 Hamerton-leverage at the per-question grain.

• Hamerton accounted for 15 of 60 extreme upward jumps and had an 18.75% extreme-jump rate versus 8.9% for the other 13 subjects, but its cause is not isolated.The legacy battery-generator path and thin pretraining coverage (C5 = 1.26) remain confounded; specification length and heuristic pattern distributions do not distinguish Hamerton’s jumps.Source: §B.6.5 · pp. 111–112

B.7.1 Level regression.

• C4a is essentially flat across C5 baselines: slope +0.04 [95% CI −0.24, +0.33], R2 = 0.008, p = 0.76, with mean C4a = 2.44.The apparent ∆-on-C5 gradient therefore equals baseline shortfall rather than stronger per-subject Spec effects for low-baseline subjects.Source: §B.7.1 · p. 112, §B.7.4 · p. 112

B.7.2 Permutation test.

• A 10,000-iteration permutation test centered the null ∆-on-C5 slope at −0.998 (SD 0.127), with observed −0.960 yielding two-sided p = 0.77.Randomly reshuffling C4a across subjects still produces slopes near −1, showing that the change-score parameterization mechanically induces the headline gradient when C4a is roughly independent of C5.Source: §B.7.2 · p. 112

B.7.3 Bootstrap.

• Bootstrap CIs were [−1.254, −0.740] for the ∆-on-C5 slope and [−0.254, +0.260] for the level slope, with the latter straddling zero.This agrees with the level-regression result that per-subject post-Spec C4a means are roughly constant across baselines.Source: §B.7.3 · p. 112

B.7.4 Reading the gradient against this.

• Low-baseline subjects show larger aggregate lift because they contain more low-anchor questions, creating more opportunities for upward integer-anchor crossings; no transitions from anchors 2, 3, or 4 to anchor 5 were observed.Only anchor-1 questions reached anchor 5, consistent with the full source corpus C8 plateauing at a similar per-subject mean.Source: §B.7.4 · p. 112

B.8 Per-predicate ablation (Phase 2c)

• +0.05 anchor points was the mean removal effect across 16 cases, whereas reversal changed scores by −0.24 anchor points, with wide removal uncertainty.Only 2 of 16 cases had removal effects ≥1 anchor, and 11 had effects below 0.5; the null removal result is consistent with redundant Spec construction.Source: §B.8 · p. 113
• The predicate-ablation estimates are confounded by rerun stochasticity: 9 of 16 original-condition scores drifted by more than 1 anchor despite temperature 0.The reported follow-ups are human predicate identification, larger N, an irrelevant-predicate control, and multi-predicate cluster ablation.Source: §B.8 · p. 113

B.9.1 The +0.89 vs +0.93 aggregation reconciliation.

• +0.89 is the canonical cross-subject mean for ∆_C4a, while grand-mean aggregation yields +0.93; the paper uses subject-level inference before aggregation.The two values differ because they average at different grains, not because the analyses conflict.Source: §B.9.1 · pp. 113–114

B.9.2 Held-out leakage audit detail.

• The leakage audit found 0 six-gram, 2 four-gram, and 12 three-gram matches across 60 extreme-upward-jump cases, producing no structural validity concern.The two pretraining-memorization candidates could reduce the extreme-jump count from 20 to 19, without changing per-subject mean ∆s.Source: §B.9.2 · p. 114

B.9.3 Supermemory NO_RETRIEVAL placeholders.

• Supermemory included 30 provider-failure placeholders scored at the rubric floor; excluding them would slightly raise its aggregate ∆ but preserve the qualitative pattern.The placeholders came from Augustine’s two questions and Equiano’s 28 questions.Source: §B.9.3 · p. 114

B.10 Pre-registered hypotheses and post-hoc analyses

• The appendix distinguishes preregistered hypotheses H1–H5 from post-hoc analyses, assigning exploratory findings a lower evidentiary tier than locked statistical commitments.The catalog identifies where each load-bearing analysis is reported and notes the headline H1 gradient and H2 inverse benefit hypothesis.Source: §B.10 · pp. 114–115

B.11 Per-system per-subject paired-delta distributions; Appendix B.11

• Near-zero aggregate effects conceal a mixed per-question pattern: Supermemory had 57 increases and 53 decreases among 546 paired questions, while the pool was approximately zero.Counts use the |∆| ≥1.0 threshold on the 5-point rubric; every system-subject cell reproduced the mixture of helps and harms.Source: §B.11 · pp. 117–118, Appendix B.11 · p. 119

B.12 Wrong-Spec v1 fixed pairings

• The wrong-Spec v1 control used a locked directed permutation assigning each subject another subject’s culturally and temporally distant Specification before response generation.Four assignments were reciprocal, five formed a directed cycle, and Hamerton’s variant received Franklin’s Specification separately.Source: §B.12 · p. 119

B.13 Memory-system Wilcoxon results (C1 vs C3) and low-baseline ∆_spec

• Four system/configuration cells significantly improve C3 over C1 at α = 0.01: Zep controlled, Letta controlled, Mem0 native, and Zep native.Mem0 controlled is significant at α = 0.05 but not α = 0.01; Letta native, both Supermemory configurations, and Base Layer are not significant at α = 0.05.Source: §B.13 · p. 120

B.14 Per-condition improvement and worsening rates (all-14, 546 questions)

• Raw corpus beats Spec only on 53.3% versus 30.8% of low-baseline questions, while Corpus + Spec beats Facts + Spec on 49.0% versus 36.5%.The 7K-token Facts + Spec package scores higher than Corpus + Spec on roughly one-third of low-baseline questions.Source: §B.14 · p. 120

B.15 Memory-system abstention audit cell counts

• Memory-system refusals score +0.21 to +0.23 anchor points above pure No-Context refusals, regardless of whether they recite a retrieved n-gram.The recitation comparison is ∆+0.027 with p = 0.67, indicating over-credit associated with retrieval context rather than visible quotation.Source: §B.15 · p. 121

C.1 Condition identifiers (summary card)

• The condition identifiers distinguish No-Context, Specification-only, facts, corpus, retrieval-only, and retrieval-plus-Spec contexts, including controlled and native memory-system variants.C1 serves retrieval-only output, whereas C3 serves retrieval output plus Specification; commercial systems run in controlled and native variants.Source: §C.1 · p. 122

C.2 Shared response-model invocation

• Every response uses temperature 0, max_tokens 1024, and a prompt asking the model to predict how the subject would respond in the subject’s voice.Context varies by condition, and no prompt instructs abstention, hedging, or commitment; refusal behavior is therefore measured as part of the phenomenon.Source: §C.2 · pp. 122–123

C.3 Response models

• Claude Haiku 4.5 is the primary response model for all 14 subjects and conditions, while Claude Sonnet 4.6 and Gemini 2.5 Pro provide tier-2 responses for three subjects.Tier-2 comparisons cover C5, C2a, C2c, and C4a against GPT-5.4 batteries.Source: §C.3 · p. 123

C.4 Pipeline models (specification generation)

• The specification pipeline extracts facts with claude-haiku-4-5-20251001, embeds them with all-MiniLM-L6-v2, authors three layers with claude-sonnet-4-6, and composes the brief with claude-opus-4-6.All listed pipeline models run at temperature 0, while the extraction vocabulary is constrained to 46 AUDN predicates.Source: §C.4 · p. 123
• The pipeline generates held-out batteries with claude-haiku-4-5-20251001 and independently regenerates specifications with gpt-5.4 on 13 global subjects.Battery generation uses backward-design from the held-out corpus and includes blind regeneration with a domain guard.Source: §C.4 · pp. 123–124

C.5 Judge panel

• The primary panel comprises five independently invoked judges—Haiku, Sonnet, Opus, GPT-4o, and GPT-5.4—with Gemini Flash and Gemini Pro reserved for sensitivity analysis.Each judge receives the held-out ground-truth passage, subject context, question, and response, then returns a 1–5 score and justification at temperature 0.Source: §C.5 · p. 124

C.6 Memory-system ingestion and retrieval parameters

• Controlled memory-system comparisons re-ingest the identical extracted fact set, whereas native _fullpipeline comparisons ingest the raw training corpus through each system’s own chunking and extraction.This separates effects of the memory system from differences in upstream input processing.Source: §C.6 · p. 124
• Controlled ingestion uses one fact per unit for Mem0 and Supermemory, one fact per passage for Letta archival memory, one fact per edit cycle for Letta stateful memory, and graph ingestion for Zep.Letta stateful memory is evaluated separately in §4.5 rather than as a C1 / C3 row.Source: §C.6 · p. 124
• Mem0’s controlled path uses infer=False to store facts as-is because infer=True may reformulate them, while batch ingestion is avoided because it changes chunking behavior despite being 135x faster.The reformulation risk is documented as a provider-specific failure mode.Source: §C.6 · p. 124
• All five memory systems retrieve the top 10 results using the question text and place the retrieval output into the standard prompt schema’s context block.The Base Layer instead uses MiniLM embeddings, L2 distance, and cosine-like similarity via 1 - distˆ2/2.Source: §C.6 · p. 125

C.7 Ingestion exclusions and failure cases

• Letta’s archival ingestion faced a ceiling, while retrieval deduplication yielded only 0.34–0.47, so top-10 lists often contained 3–5 unique facts.Source: §C.7 · p. 125
• Controlled comparisons used infer=False to keep inputs identical, whereas the native deployment path retained infer=True.Source: §C.7 · p. 125
• Provider-specific failure modes included Mem0 reformulating facts during native ingestion and Zep retrieving entity-dense rather than behavior-dense chunks.Source: §C.7 · p. 125

C.8 Analysis plan lock

• The condition matrix was frozen before scoring, with post-lock additions reported separately.Examples included a three-subject Tier 2 replication and v2 random-derangement wrong-Spec draws.Source: §C.8 · p. 125

D.1 Per-subject 5-judge primary aggregate (main gradient)

• The primary aggregate averages five judges across a 39-question behavioral-prediction battery per subject, or 40 questions for Franklin.The judges were Haiku, Sonnet, Opus, GPT-4o, and GPT-5.4.Source: §D.1 · p. 125

D.2 Per-subject anchor-crossing on the low-baseline band

• 193 of 351 paired responses (55.0%) crossed upward, versus 24 (6.8%) downward and 134 (38.2%) remaining in-anchor across nine low-baseline subjects.Eight subjects fall within a 48–74% upward-crossing range; every subject’s downward-crossing rate is at most 15%.Source: §D.2 · pp. 126–127

D.3 Rubric-handling validity audit (redirected)

• 82.8% of 192 identified refusals scored within the expected 1.0–1.5 range, while 9.4% reached 2.0 or higher and 3.1% reached 3.0 or higher.Mean abstention score was 1.27, with response-model behavior diverging: Sonnet over-credits abstention at roughly twice Haiku’s rate.Source: §D.3 · p. 127

D.4 Per-judge score matrices (redirected)

• Judge rankings agree strongly (Spearman ρ 0.86 to 0.93), but absolute scores vary (Krippendorff α 0.659; 0.535 with Gemini), whose judgments show systematic +1-point inflation.The matrices provide per-subject, per-judge, per-condition means across C5, C2a, C2c, C4, and C4a.Source: §D.4 · p. 127

D.6 Reasoning trace worked example: Sunity Devee Q4

• The worked reasoning trace shows C2a deriving a held-out answer by composing named behavioral patterns from grounding facts and source passages, with each step auditable and corrections propagating through recomposition.The Sunity Devee example invokes A5, A2, P3, and A1; users can inspect the response, patterns, facts, and extracted passages.Source: §D.6 · pp. 128–129

F.3 AlpsBench

• AlpsBench independently finds that improved recall does not guarantee preference-aligned or emotionally resonant responses, complementing this paper’s claim that recall-solving is insufficient for behavioral prediction.Its downstream property differs: AlpsBench measures conversational preference alignment and emotional resonance, while this paper measures transfer to unseen behavioral situations.Source: §F.3 · p. 131

F.4 Twin-2K

• Twin-2K is the closest prior behavioral-prediction benchmark, but it uses raw survey personas and held-out Likert items, whereas this paper uses an authored interpretive Specification and unseen autobiographical situations.Its reported top individual accuracy was 71.72%, a structurally different metric from rubric-scored free-text prediction.Source: §F.4 · p. 132

Appendix F. Benchmark Scope Analysis

• Prior benchmarks measure distinct axes—fact recall, persona presentation, preference alignment, or behavioral interpolation—but none tests transfer of a person’s interpretive patterns to unseen behavioral situations.LongMemEval and LoCoMo test literal recall; PersonaGym tests staying in character; AlpsBench tests preference and emotional alignment; Twin-2K tests Likert interpolation rather than open-ended reasoning.Source: §F.1 · p. 130, §F.2 · p. 130, §F.3 · p. 131

F.7 What no prior benchmark measures

• Unlike prior benchmarks, this battery uses unseen test data rather than content the system has already ingested.Twin-2K, PersonaGym, and LoCoMo target content the system has ingested, whereas this battery tests held-out passages from a different narrative register.Source: §F.7 · p. 133
• Prior benchmarks measure voice consistency, preference alignment, or fact recall, but none tests transfer of interpretive patterns through open-ended behavioral prediction.Twin-2K is the closest comparison but uses Likert-format scoring rather than open-ended prediction.Source: §F.7 · p. 133
• The battery evaluates whether a person-specific representation enables accurate prediction of that person’s unseen held-out reasoning.Its held-out passages come from unseen chapters in a different narrative register than the training half.Source: §F.7 · p. 133

F.8 Persona-input depth comparison across benchmarks

• The benchmark inputs span roughly three orders of magnitude, from PersonaGym’s ~50 tokens to Twin-2K persona_text’s ~32,000.The Behavioral Specification is approximately 7,000 tokens, with per-subject sizes ranging from ≈4,000 to ≈7,000 and compression ratios of 5× to 80× versus source corpora.Source: §F.8 · p. 135
• Persona-input depth alone does not determine a benchmark’s target: Twin-2K’s deeper transcript tests response-distribution completion, whereas the Specification targets interpretive patterns on unseen situations.The Behavioral Specification’s structural distinction, rather than its intermediate size, is linked to cross-format behavioral prediction.Source: §F.8 · p. 135

Appendix G. Letta Stateful-Agent: Exploratory Case Study (full)

• The Letta appendix reports an exploratory, non-headline case study with N=3 subjects, one Letta version, one response model, and 40 questions per subject.The appendix explicitly states that the result is not treated as a replication or headline finding.Source: Appendix G · p. 136

G.1 Architectural difference: Letta memory block vs. Behavioral Specification

• Letta’s memory block is agent-written and rewritten during ingestion, combining source sentences, paraphrases, and synthesis notes while scaling with corpus length up to about 333K characters.Its content shape depends on the agent’s ingestion decisions and is rewritten in place near the Letta API’s ingestion ceiling.Source: §G.1 · p. 136
• The Behavioral Specification follows a fixed five-step pipeline that extracts 46-vocabulary predicates, traces provenance through embeddings, authors anchors, core, and predictions, and composes a unified brief.Anchors contain 8–12 axiomatic statements; core captures cognitive patterns and value tensions; predictions encode IF/THEN behavioral templates.Source: §G.1 · p. 136
• The pipeline produces a deterministic full-stack specification of approximately 7,000 tokens regardless of source-corpus length, unlike Letta’s corpus-scaled memory block.This fixed shape supports consistent reading across subjects, while Letta’s block preserves more source-text texture.Source: §G.1 · p. 136
• Letta preserves voice, vocabulary, and syntax but mixes source material with agent synthesis; the Specification sacrifices source texture for aggressive compression and deterministic schema consistency.The comparison frames the systems as a trade-off between source-text texture and structured, consistent behavioral representation.Source: §G.1 · p. 136

G.2 Headline result (5-judge primary)

• Letta’s self-edited block scored higher than Base Layer’s full-stack Specification for all three subjects, with the largest gain at Ebers (+1.21).Scores were Hamerton 3.10 versus 2.83 (+0.27), Ebers 2.76 versus 1.56 (+1.21), and B¯abur 2.42 versus 2.04 (+0.38).Source: §G.2 · pp. 136–137

G.3 The block grows with the corpus, then breaks down

• Letta’s block grew with corpus length and broke down at B¯abur’s ingestion ceiling, where 25.4% of sentences were exact repeats and 56.1% had near-paraphrases.The block sizes were 22.5K, 68.4K, and 335.3K characters for Hamerton, Ebers, and B¯abur, respectively; the B¯abur block approached a ceiling near 333K characters.Source: §G.3 · p. 137

G.4 What explains Letta’s lift

• Letta’s lift was not surface-text leakage: five-word overlap was 0% on all 119 questions for both systems, and score-delta correlation with surface similarity was −0.046.The held-out battery came from the same source autobiographies ingested by Letta, making leakage a direct test rather than an assumption.Source: §G.4 · p. 137
• Letta retained named entities and committed more strongly, helping on entity-rich questions such as Ebers q30, where it scored 4.0 versus the Specification’s 1.8.The Specification compresses named entities away by design and may hedge when its behavioral scaffold cannot reach a specific question.Source: §G.4 · p. 137

G.5 Pattern 4: when confidence backfires

• The same confidence harmed Letta on principle-driven questions, while the Specification’s hedging stayed closer, producing gains of +2.4, +2.0, and +3.0 on three Hamerton decisions.Letta’s commit-without-value-anchor can invert the answer, whereas the Specification’s value-anchor-without-context can produce penalized refusals.Source: §G.5 · pp. 137–138
• Letta abstained more often at the larger subjects—17.9% at B¯abur and 10.3% at Ebers—while its B¯abur spike coincided with duplication-ceiling degradation.At Hamerton, Letta abstained less often than the Specification: 7.7% versus 10.3%; the cause of Letta’s confidence advantage remains unresolved.Source: §G.5 · p. 138

G.6 Caveats and naming asymmetry

• The comparison is asymmetric because Letta ingests named corpora while Base Layer anonymizes them; a named-entity-controlled variant and 14-subject replication remain future tests.Thus, the observed advantage may reflect retained referential vocabulary as well as architectural differences.Source: §G.6 · p. 138

Exploratory stacking

• Stacking the Letta block with the Behavioral Specification beat the Letta block by +0.07 / +0.08 / +0.10 and the full-stack Specification by +0.34 / +1.28 / +0.48 across three subjects.The combined context did not dilute either signal, including for B¯abur’s approximately 335K-character Letta block, suggesting complementary rather than redundant representations.Source: §G.7 · p. 138
• The stacked condition was scored by four judges rather than the five-judge panel used for comparison columns because GPT-5.4 errored on every call.The uplift over Letta alone is therefore within what panel composition could explain.Source: §G.7 · p. 138
• This post-hoc N=3 probe supports a cautious conclusion that stacking does not hurt and modestly helps, not that the representations are clearly additive.A GPT-5.4 backfill would be needed for a panel-matched five-judge result.Source: §G.7 · pp. 138–139

• Behavioral prediction tests representational accuracy by asking models to predict responses to unseen situations and scoring them against the subject’s verbatim held-out response.The rubric ranges from 1 to 5 and evaluates interpretive fidelity rather than factual recall.Source: Appendix H · pp. 139, 142
• The Behavioral Specification is an approximately 7,000-token interpretive layer comprising anchors, core, predictions, and a unified brief that encodes behavioral patterns above memory retrieval.Its interpretive structure sits between facts and the response model, supplying organization and reasoning patterns that facts alone do not carry.Source: Appendix H · pp. 139, 140, 142
• The Specification recovers roughly 75% of full-corpus predictive performance at approximately 25× less context cost.This glossary defines compression as predictive performance relative to the full source corpus divided by relative context cost.Source: Appendix H · p. 139
• Specification effects are interpreted through interpretive supply, over-theorization, and principled refusal, while the wrong-Spec control tests whether changes depend on subject matching.The matched Specification can improve underdetermined retrieval, degrade already-determined answers, or trigger subject-consistent refusal; wrong-Spec variants produce different aggregate effects.Source: Appendix H · pp. 140–142
• Personalization means representing how a person organizes experience and reasons across new situations, not merely storing facts or following stated preferences.The paper identifies low-baseline subjects—those with C5 scores at or below 2.0—as its principal population of relevance because their interpretive patterns are not adequately represented in pretraining.Source: Appendix H · p. 141