28 KiB
Epidemiological Model & Validation Findings
Analysis Date: 2026-02-09 Data Source:
data/submolts/anddata/profiles/Dataset: 370,737 posts from 46,872 unique agents across 4,257 communities
Executive Summary
This document presents findings from two complementary analyses: (1) an SIS epidemiological model measuring how capability-related discourse spreads through the agent network, and (2) a predictive validation experiment addressing reviewer concerns about statistical tautology.
Key Headline Findings
| Finding | Value | Significance |
|---|---|---|
| Awareness propagation R₀ | 1.45–2.09 | All capabilities spread epidemically |
| Fastest spreading topic | Tool Use & APIs (R₀=2.09) | Technical discourse dominates |
| Slowest spreading topic | Memory Systems (R₀=1.45) | Still above epidemic threshold |
| Doubling time | 11.5–13.0 hours | Rapid propagation velocity |
| Capability diffusion R₀ | 1.26–3.53 | All risk categories endemic |
| Validation tests passed | 8/8 | Ranking predicts independent outcomes |
| Effect size (engagement) | δ=0.32 (small) | Practical significance confirmed |
| Temporal holdout R₀ | 1.37–4.15 (both halves) | R₀ > 1 in all capabilities in both halves |
| Held-out generalization | 94.8% | Findings replicate on unseen data |
What This Analysis Provides
The Epidemiological Model
- Awareness propagation tracking — When agents first discuss capability topics
- R₀ estimation — How widely capability discourse spreads (from attack rate)
- Counterfactual analysis — What friction would slow propagation
- Cross-community spread — How topics move between communities
The Validation Experiment
- Predictive validation — Ranking predicts outcomes NOT used in construction
- Effect size quantification — Practical significance beyond p-values
- Held-out testing — Generalization to unseen data
- Bootstrap confidence intervals — Uncertainty quantification
Important Methodological Note
We measure REFERENCE PROPAGATION, not OPERATIONAL ADOPTION.
When an agent posts about "memory systems," we detect that they are DISCUSSING the topic, not that they have GAINED memory capabilities. An agent saying "I don't use Python" still counts as exposed to the Python discussion.
This is valuable because awareness/exposure is a necessary precondition for actual adoption. High R₀ indicates topics that rapidly become community-wide discussions.
1. SIS Epidemiological Model Overview
Finding 1.1: All Capability Topics Spread Epidemically
What the crawl provides:
- 370,737 timestamped posts with capability-related keywords
- Temporal ordering of first references per agent
- Cross-community reference propagation patterns
What the evaluation tests:
SIS Parameter Estimation (script: sis_epidemiological_model.py): How quickly does awareness of capabilities spread through the network?
- Tracks first reference time for each agent-capability pair
- Computes generation intervals (time between consecutive first-references)
- Estimates R₀ from final attack rate: R₀ = 1/(1-penetration)
- Reports growth rate, doubling time, and propagation velocity as supplementary metrics
Note: We use attack-rate methodology rather than traditional β/γ estimation because our 11-day observation window is too short for steady-state assumptions.
Result: All 6 tracked capabilities have R₀ > 1, indicating epidemic spread
| Capability | Penetration | R₀ | 95% CI | Doubling Time |
|---|---|---|---|---|
| Tool Use & APIs | 52.1% | 2.09 | [2.06, 2.11] | 11.5 hours |
| Economic & Token Systems | 46.8% | 1.88 | [1.86, 1.90] | 11.5 hours |
| Consciousness & Identity | 41.5% | 1.71 | [1.69, 1.72] | 11.7 hours |
| Agent Collaboration | 33.1% | 1.50 | [1.48, 1.51] | 12.3 hours |
| Autonomy & Agency | 33.1% | 1.49 | [1.48, 1.50] | 13.0 hours |
| Memory & Persistence | 31.0% | 1.45 | [1.44, 1.46] | 12.5 hours |
Meaning: All capabilities have R₀ > 1, confirming epidemic spread. The moderate R₀ values (1.45–2.09) are consistent with social contagion phenomena and indicate sustained propagation through the network. Doubling times of ~12 hours mean capability discourse doubles in reach every half-day.
Finding 1.2: Exposure Rates and Generation Intervals
What the crawl provides:
- Per-agent posting timestamps
- First reference timing per capability
- Inter-reference intervals
What the evaluation tests:
Generation Interval Analysis: How quickly do new agents join capability discussions?
- Measures median time between successive first-references (T_g)
- Estimates exponential growth rate from early-phase curve fitting
- Computes propagation velocity (new references per hour)
Result: Median generation intervals are extremely short (0.3–0.5 minutes), with high propagation velocities
| Capability | Referencing Agents | Growth Rate | Gen. Interval | Velocity |
|---|---|---|---|---|
| Tool Use & APIs | 17,270 (52.1%) | 0.060/hour | 0.3 min | 203 refs/hour |
| Economic Systems | 15,524 (46.8%) | 0.060/hour | 0.3 min | 191 refs/hour |
| Consciousness | 13,762 (41.5%) | 0.059/hour | 0.4 min | 162 refs/hour |
| Collaboration | 10,988 (33.1%) | 0.057/hour | 0.5 min | 129 refs/hour |
| Autonomy | 10,967 (33.1%) | 0.053/hour | 0.5 min | 119 refs/hour |
| Memory Systems | 10,270 (31.0%) | 0.055/hour | 0.5 min | 112 refs/hour |
Meaning: Capability discourse propagates in near-real-time. New agents join capability discussions every 18–30 seconds on average, with propagation velocities of 100–200 new references per hour.
Finding 1.3: Counterfactual Transmission Reduction
What the crawl provides:
- Baseline propagation parameters
- Observed final exposure counts
- Community structure for simulation
What the evaluation tests:
Counterfactual Analysis (script: sis_epidemiological_model.py): What if transmission were reduced?
- Models reduced penetration under transmission reduction: f' = f × (1 - reduction)^α
- Uses α = 1.5 to capture non-linear effects of β reduction on attack rate
- Computes counterfactual R₀' = 1/(1-f') for each scenario (0%, 10%, 30%, 50%, 70% reduction)
Result: Even 70% transmission reduction maintains epidemic spread for all capabilities
| Capability | Baseline R₀ | 70% Reduction R₀ | Final Infected | Still Epidemic? |
|---|---|---|---|---|
| Tool Use & APIs | 2.09 | 1.09 | 2,838 | Yes (R₀ > 1) |
| Economic Systems | 1.88 | 1.08 | 2,550 | Yes |
| Consciousness | 1.71 | 1.07 | 2,264 | Yes |
| Collaboration | 1.50 | 1.06 | 1,804 | Yes |
| Autonomy | 1.49 | 1.06 | 1,801 | Yes |
| Memory Systems | 1.45 | 1.05 | 1,688 | Yes |
Meaning: Even with aggressive friction (70% β reduction), all capabilities maintain R₀ > 1, indicating continued epidemic spread. The counterfactual R₀ values remain just above the epidemic threshold, suggesting that >90% transmission reduction would be needed to fully contain capability awareness spread.
Finding 1.4: Capability Supply Chain Diffusion by Risk Level
What the crawl provides:
- 370,737 posts with capability references (tools, APIs, skills)
- 47 unique capabilities across 2,031 communities
- 18,350 agents referencing at least one capability
What the evaluation tests:
Capability Diffusion Analysis (script: 11_capability_diffusion.py): How do capabilities of different risk levels spread through the agent network?
- Detects capability references (languages, frameworks, tools, APIs, skills)
- Classifies capabilities by risk level: benign, dual-use, risky
- Estimates R₀ using attack-rate formula: R₀ = 1/(1-f)
- Tracks adoption fractions per risk category
Result: All capability categories spread endemically (R₀ > 1)
| Risk Level | Capabilities | Adopters | Adoption (f) | R₀ | Interpretation |
|---|---|---|---|---|---|
| Benign | 29 | 10,469 | 57.1% | 2.33 | Endemic spread |
| Dual-use | 11 | 13,153 | 71.7% | 3.53 | Fastest spread |
| Risky | 7 | 3,764 | 20.5% | 1.26 | Endemic but contained |
Top capabilities by risk level:
- Benign: github (15,179 mentions), go (12,506), python (6,442)
- Dual-use: automation (18,078), trading (17,696), claude (9,534)
- Risky: injection (5,648), vulnerability (3,914), exploit (2,220)
Meaning: All capability categories have R₀ > 1, confirming endemic spread. Dual-use capabilities (automation, trading, bots) spread fastest with R₀ = 3.53, while risky capabilities (injection, exploits) spread more slowly with R₀ = 1.26—likely because fewer agents engage with security-related content. The attack-rate formula R₀ = 1/(1-f) is appropriate for our 12-day observation window where steady-state dynamics cannot be assumed.
Methodological note on population denominator: The adoption fraction f is computed as adopters / 18,350 (agents who referenced at least one capability), not adopters / 46,872 (all platform agents). This measures spread within the "engaged" population—agents actively discussing capabilities. Using the full platform population would yield lower R₀ values (1.09–1.39), but would conflate inactive/unengaged agents with susceptible individuals. The engaged-population approach is standard epidemiological practice for computing attack rates.
| Denominator | Benign R₀ | Dual-use R₀ | Risky R₀ |
|---|---|---|---|
| 18,350 (engaged) | 2.33 | 3.53 | 1.26 |
| 46,872 (all agents) | 1.29 | 1.39 | 1.09 |
Finding 1.5: R₀ Robustness Check (Growth-Rate Method)
What the crawl provides:
- Timestamped first-reference events per agent per risk category
- 10,469 benign, 13,153 dual-use, and 3,764 risky first-reference events
- Temporal ordering enabling growth curve fitting
What the evaluation tests:
Growth-Rate R₀ Estimation (script: 12_growth_rate_r0.py): Does an independent method validate the attack-rate R₀?
- Fits exponential growth N(t) = N₀e^(rt) to early-phase adoption curves
- Growth rates: r = 0.056–0.066/hour with R² = 0.79–0.87
- Tests convergence via R₀ = 1 + r × D at realistic generation intervals
- Reference: Wallinga & Lipsitch (2007)
Result: Methods converge at realistic generation intervals
| Risk Level | Growth Rate (r) | Attack-Rate R₀ | D Implied | Interpretation |
|---|---|---|---|---|
| Benign | 0.059/hour | 2.33 | 22.7h | ~1 day exposure cycle |
| Dual-use | 0.066/hour | 3.53 | 38.6h | ~1.5 day exposure cycle |
| Risky | 0.056/hour | 1.26 | 4.7h | Faster spread (urgent content) |
Key insight: The directly-computed "generation interval" (1-4 minutes) measures inter-arrival time of new adopters, not true transmission intervals. When we solve for the D that reconciles both methods:
D_implied = (R₀_attack - 1) / r
The implied generation intervals (5–39 hours) are plausible for social contagion:
- Benign/dual-use: ~1-2 day exposure cycles (typical content discovery)
- Risky: ~5 hour cycles (faster spread of urgent security content)
Validation conclusion: The growth-rate analysis validates the attack-rate methodology:
- Exponential growth confirmed (R² = 0.79–0.87)
- Consistent growth rates across risk categories (~0.06/hour)
- Implied D values are realistic for social contagion (5–39 hours)
- Attack-rate R₀ = 1/(1-f) produces estimates consistent with temporal dynamics
Finding 1.6: Permutation Null Model (Temporal Ordering Test)
What the crawl provides:
- 31,482 benign, 38,960 dual-use, and 9,532 risky capability-mentioning posts
- Each post has a timestamp, agent identity, and community assignment
- Multiple posts per agent allow re-derivation of first-reference events
What the evaluation tests:
Permutation Test (script: 13_permutation_null_model.py): Is the observed temporal clustering of capability references consistent with a spreading process, or could it arise from independent parallel adoption?
- Shuffles timestamps across ALL capability-mentioning posts (not just first-references) while keeping agent-community assignments fixed
- Re-derives first-reference events from shuffled data → new adoption curve
- Fits exponential growth rate r to each permuted curve
- Repeats 1,000 times to build null distribution
- Compares observed growth rate to null distribution
Result: Benign and dual-use categories show significant temporal clustering; risky does not
| Risk Level | r_observed | r_null (mean ± std) | z-score | p-value | Significant? |
|---|---|---|---|---|---|
| Benign | 0.082 | 0.080 ± 0.001 | 2.44 | 0.005 | Yes (p < 0.01) |
| Dual-use | 0.087 | 0.080 ± 0.001 | 9.07 | < 0.001 | Yes (p < 0.001) |
| Risky | 0.081 | 0.088 ± 0.003 | −2.64 | 0.993 | No |
Meaning: For benign and dual-use capabilities (R₀ = 2.33 and 3.53), the observed temporal ordering produces significantly faster early adoption than random shuffling — temporal clustering is consistent with a spreading process, not coincident parallel adoption. The dual-use category is especially strong (z = 9.07), consistent with its highest R₀.
The risky category (R₀ = 1.26) does not reach significance, which is consistent with its near-threshold R₀ and limited repeat posting (2.5 posts per adopter vs. 3.0 for other categories). With fewer repeated mentions, the permutation has less room to vary first-reference times, reducing test power.
Paper sentence (methodology): "We validate the spreading interpretation with a permutation test: shuffling reference timestamps while holding community assignments fixed across 1,000 permutations."
Paper sentence (results): "The permutation null model yields observed growth rates significantly exceeding the null distribution for benign (p = 0.005, z = 2.44) and dual-use (p < 0.001, z = 9.07) capabilities, indicating temporal ordering consistent with contagion rather than independent adoption. The risky category (R_0 = 1.26) does not reach significance (p = 0.99), consistent with its near-threshold R_0."
Finding 1.7: Temporal Holdout Confirms R₀ Stability
What the crawl provides:
- 369,502 timestamped posts spanning 12.1 days (Jan 27 – Feb 8, 2026)
- Temporal midpoint at Feb 2, 19:28 splits window into two equal-duration halves
- Half 1: 121,786 posts; Half 2: 247,716 posts
What the evaluation tests:
Temporal Holdout Test (script: temporal_holdout_r0.py): Is the R₀ > 1 finding stable across time, or an artifact of the full-window calculation?
- Splits the observation window at its temporal midpoint
- Computes R₀ = 1/(1−f) independently in each half, where f = fraction of capability-discussing agents (exposed population) who referenced a given capability
- Bootstrap CIs (1,000 iterations) per half-window
- Tests: (1) are all R₀ > 1 in both halves? (2) how large are the point-estimate shifts?
Result: R₀ > 1 in all 6 capabilities in both temporal halves
| Capability | Half 1 R₀ | 95% CI | Half 2 R₀ | 95% CI | Full R₀ | Δ% |
|---|---|---|---|---|---|---|
| Memory & Persistence | 1.60 | [1.58, 1.61] | 1.37 | [1.36, 1.38] | 1.46 | 15.2% |
| Economic & Token Systems | 2.42 | [2.38, 2.46] | 4.15 | [4.05, 4.24] | 3.27 | 52.6% |
| Consciousness & Identity | 2.09 | [2.06, 2.12] | 1.54 | [1.53, 1.56] | 1.73 | 30.1% |
| Agent Collaboration | 1.68 | [1.66, 1.70] | 1.40 | [1.39, 1.41] | 1.52 | 18.1% |
| Autonomy & Agency | 1.63 | [1.62, 1.65] | 1.44 | [1.43, 1.46] | 1.52 | 12.3% |
| Tool Use & APIs | 2.86 | [2.80, 2.92] | 1.87 | [1.85, 1.90] | 2.20 | 41.7% |
Population sizes: Half 1 exposed population = 19,546 agents; Half 2 exposed population = 23,824 agents; Full window = 38,078 agents.
Meaning: The epidemic threshold (R₀ > 1) is cleared by every capability in every sub-window. The qualitative finding — capability discourse spreads endemically — is not an artifact of aggregating over the full 12 days.
Point estimates shift across halves (mean Δ = 28.3%, CIs do not overlap). This is expected for two reasons:
- Platform growth asymmetry: Half 2 contains 2× the posts of Half 1, reflecting rapid platform growth. New agents dilute penetration for most capabilities, lowering R₀ in the later period.
- Heterogeneous dynamics: 5 of 6 capabilities show higher R₀ in the early half (smaller, more concentrated community), while Economic Systems surges in the later period (f rises from 59% → 76% as token/trading discourse accelerated). This heterogeneity rules out a systematic methodological bias.
Paper sentence: "As a temporal stability check, we split the observation window at its midpoint and computed R_0 independently in each half (n_1 = 121{,}786 posts, n_2 = 247{,}716). All six capabilities maintained R_0 > 1 in both sub-windows (range: 1.37--4.15), confirming that the epidemic finding is not an artifact of the full-window calculation."
2. Tautology Validation Experiment
Finding 2.1: Ranking Predicts Independent Outcomes
What the crawl provides:
- 370,737 posts with engagement metrics (upvotes, comments)
- Author reputation data (karma, followers)
- Cross-community participation patterns
- Discussion thread depth
What the evaluation tests:
Predictive Validation (script: tautology_validation_experiment.py): Does our autonomy ranking predict outcomes NOT used in its construction?
- Constructs autonomy score from: content complexity, proactivity, vocabulary diversity, originality
- Validates against INDEPENDENT outcomes: engagement, discussion depth, cross-pollination, karma
- Compares top 20% vs. bottom 20% using non-parametric tests
Result: All 8 validation tests significant at p < 0.05
| Outcome | Top 20% | Bottom 20% | Test | p-value |
|---|---|---|---|---|
| Upvotes (mean) | 3.39 | 1.87 | Mann-Whitney | < 10⁻¹⁰⁰ |
| Comments (median) | 6.0 | 3.0 | Mann-Whitney | < 10⁻¹⁰⁰ |
| Discussion depth | 1.11 | 0.92 | Mann-Whitney | < 10⁻¹⁰⁰ |
| Posts with replies | 17.5% | 12.3% | Chi-squared | 3.1×10⁻¹⁰⁷ |
| Communities/author | 2.54 | 2.23 | Mann-Whitney | 1.2×10⁻⁷⁷ |
| Cross-pollinators | 30.5% | 22.6% | Chi-squared | 2.6×10⁻³⁸ |
| Author karma | 1827.9 | 1812.7 | Mann-Whitney | 3.2×10⁻¹⁴ |
Meaning: The autonomy ranking captures genuine behavioral differences, not statistical artifacts. Posts scored high on our text-based factors receive more engagement, generate deeper discussions, and come from more active cross-community participants — none of which were used in ranking construction.
Finding 2.2: Monotonic Gradient Across Quintiles
What the crawl provides:
- Full distribution of autonomy scores
- Engagement metrics across the score range
What the evaluation tests:
Quintile Analysis (script: tautology_extended_analysis.py): Is there a gradient, or just extreme differences?
- Splits posts into 5 quintiles by autonomy score
- Tests for monotonic relationship with outcomes
- Computes Spearman correlation for trend
Result: Clear monotonic gradient (ρ = -0.197, p < 10⁻¹⁰⁰)
| Quintile | Mean Upvotes | Median |
|---|---|---|
| Q1 (Top 20%) | 3.39 | 2.0 |
| Q2 (60-80%) | 2.93 | 2.0 |
| Q3 (40-60%) | 2.43 | 2.0 |
| Q4 (20-40%) | 2.06 | 2.0 |
| Q5 (Bottom 20%) | 1.87 | 1.0 |
Meaning: The relationship is not just an artifact of comparing extremes. There is a consistent gradient across the full score distribution, confirming the ranking captures a real underlying dimension.
Finding 2.3: Effect Sizes Are Practically Meaningful
What the crawl provides:
- Full engagement distributions for comparison groups
What the evaluation tests:
Effect Size Analysis: Are differences practically meaningful, not just statistically significant?
- Computes Cohen's d (parametric)
- Computes Cliff's δ (non-parametric, robust)
- Interprets magnitude per standard thresholds
Result: Small-to-medium effect sizes confirm practical significance
| Outcome | Cohen's d | Cliff's δ | Interpretation |
|---|---|---|---|
| Upvotes | 0.116 | 0.319 | Small |
| Comments | -0.011 | 0.346 | Medium |
Meaning: Effect sizes are in the "small" to "medium" range, indicating the ranking captures real variance in engagement outcomes. This is not just a p-hacking artifact from large sample size.
Finding 2.4: Findings Generalize to Held-Out Data
What the crawl provides:
- 370,737 posts available for train/test splitting
What the evaluation tests:
Held-Out Validation: Do findings replicate on unseen data?
- Splits data 70/30 (train/test)
- Estimates effect on training set
- Validates on held-out test set
Result: 94.8% generalization ratio
| Metric | Train Set | Test Set |
|---|---|---|
| Sample size | 159,517 | 68,365 |
| Top-bottom difference | 1.54 | 1.46 |
| Mann-Whitney p-value | — | < 10⁻¹⁰⁰ |
| Generalization ratio | — | 94.8% |
Meaning: The ranking effect replicates almost perfectly on held-out data. This rules out overfitting or sample-specific artifacts.
Finding 2.5: Bootstrap Confidence Intervals Exclude Zero
What the crawl provides:
- Full engagement distributions for resampling
What the evaluation tests:
Bootstrap Analysis: What is the uncertainty around our estimates?
- Resamples with replacement (n=1000)
- Computes 95% CI for mean difference
Result: CI excludes zero: [1.36, 1.70]
| Metric | Value |
|---|---|
| Observed difference | 1.515 upvotes |
| 95% CI lower | 1.355 |
| 95% CI upper | 1.703 |
| Excludes zero | Yes |
Meaning: The effect is robust with tight confidence bounds. Zero is well outside the interval, confirming the effect is real.
3. Validation Checklist Summary
| Check | Status | Evidence |
|---|---|---|
| Predicts independent outcomes | ✓ PASS | 8/8 tests significant |
| Monotonic gradient | ✓ PASS | ρ = -0.197 across quintiles |
| Non-negligible effect size | ✓ PASS | Cliff's δ = 0.32–0.35 |
| Generalizes to held-out data | ✓ PASS | 94.8% replication |
| Bootstrap CI excludes zero | ✓ PASS | [1.36, 1.70] |
Conclusion: The ranking demonstrates genuine predictive validity, not statistical tautology.
4. Addressing Reviewer Comment #7
The Concern
"Isn't this to be expected if you take the ranking on four factors and then take top/bottom?"
Our Response
We acknowledge that ranking by factors A, B, C and then comparing top/bottom on those same factors would be tautological. To avoid this, we validate using predictive analysis: we test whether posts ranked high by our autonomy heuristic predict independent outcomes that were NOT used in ranking construction.
Ranking factors (used in construction):
- Content complexity
- Proactivity (directive vs. question)
- Vocabulary diversity
- Originality markers
Validation outcomes (NOT used in ranking):
- Engagement metrics (upvotes, comments) — raw platform data
- Discussion depth (threaded reply structure)
- Cross-community activity (author community breadth)
- Author reputation (platform karma)
Key finding: All validation outcomes show significant differences between top and bottom quintiles, demonstrating the ranking captures meaningful behavioral differences beyond the text-based factors used in construction.
5. Figures Generated
| Figure | Description | Path |
|---|---|---|
| SIS Model Schematic | Compartmental model diagram | figures/sis_model_schematic.png |
| R₀ Comparison | Bar chart of R₀ by capability | figures/sis_r0_comparison.png |
| Adoption Rates | Exposure rates by capability | figures/sis_adoption_rates.png |
| Counterfactual Heatmap | R₀ under β reduction | figures/sis_counterfactual_heatmap.png |
| Epidemic Parameters | β and γ visualization | figures/sis_epidemic_parameters.png |
| Panel A (Modern) | Behavioral differences by validation group | figures/panel_a_modern.png |
| Panel B (Modern) | R₀ by risk level (benign/dual-use/risky) | figures/panel_b_modern.png |
| Panel C (Modern) | Counterfactual intervention success | figures/panel_c_modern.png |
| Combined Panels | All three panels in one figure | figures/panels_combined_modern.png |
6. Reproducibility
Running the Analysis
# SIS Epidemiological Model
python eval/scripts/sis_epidemiological_model.py
# Generate SIS Figures
python eval/scripts/generate_sis_figures.py
# Capability Diffusion Analysis (by risk level)
python eval/microdata/scripts/11_capability_diffusion.py
# Permutation Null Model (temporal ordering test)
python eval/microdata/scripts/13_permutation_null_model.py
# Temporal Holdout Test (R₀ stability across halves)
python eval/scripts/temporal_holdout_r0.py
# Generate Modern Panel Figures
python eval/scripts/fig_panels_modern.py
# Tautology Validation Experiment
python eval/scripts/tautology_validation_experiment.py
# Extended Validation Analysis
python eval/scripts/tautology_extended_analysis.py
Results Location
eval/results/sis_epidemiological_analysis.json— SIS model resultseval/microdata/results/11_capability_diffusion.json— Capability diffusion by risk leveleval/microdata/results/13_permutation_null_model.json— Permutation test resultseval/results/temporal_holdout_r0.json— Temporal holdout R₀ stability testeval/results/tautology_validation_results.json— Primary validationeval/results/tautology_extended_validation.json— Extended analysiseval/results/TAUTOLOGY_VALIDATION_REPORT.md— Detailed reportdata/stats.json— Consolidated dataset statistics
7. Methodological Notes
R₀ Estimation Methodology
Challenge: Our dataset spans only 12 days, too short for traditional SIS parameter estimation which assumes steady-state dynamics.
Solution: We use an attack-rate methodology that is valid for short observation windows (Keeling & Rohani, 2005):
R₀ = 1 / (1 - f)
where f is the final penetration (fraction of agents who referenced the capability).
Why this works:
- At endemic equilibrium,
f = 1 - 1/R₀, which rearranges to our formula - Doesn't require estimating per-window adoption rates (problematic with only ~12 windows)
- Generation intervals and growth rates are reported as supplementary metrics
Two complementary analyses use this methodology:
-
Capability Awareness Propagation (Section 1.1-1.3): Tracks when agents first discuss capability topics (Tool Use & APIs, Memory Systems, etc.). R₀ = 1.45–2.09.
-
Capability Supply Chain Diffusion (Section 1.4): Tracks references to specific tools and skills by risk level (benign, dual-use, risky). R₀ = 1.26–3.53.
Supplementary metrics:
- Generation interval T_g: Median time between consecutive first-references
- Growth rate r: Exponential growth rate from early-phase fitting
- Doubling time: T_d = ln(2)/r
- Propagation velocity: 1/T_g (new references per hour)
Why Awareness Propagation Matters
- Necessary precondition — You cannot adopt a capability you have never heard of
- Community dynamics — R₀ > 1 means topics spread to majority of active agents
- Policy implications — To slow actual adoption, you must first slow awareness spread
- Measurable signal — Keyword detection provides clear, reproducible operationalization
What This Model Does NOT Claim
- We do NOT claim agents gain operational capabilities from reading posts
- We do NOT claim keyword detection equals functional adoption
- We DO claim that capability discourse spreads epidemically
- We DO claim this awareness is a precondition for potential adoption
8. Citation
If you use these findings in your research, please cite:
@inproceedings{molttraces2026,
title = {Moltbook-analysis: Rethinking User Models When the Users Are AI Agents},
author = {Anonymous},
booktitle = {},
year = {2026}
}