LeafSleuth Statistical Methodology & Empirical Calibration Results¶
This document records the statistical methodology, hypothesis-testing architecture, failure-mode countermeasures, and empirical calibration results for LeafSleuth (plugins/LeafSleuth). Every simulation, real-human trace evaluation, and large-population scaling benchmark cited below is committed in sleuth-common and sleuth-paper and can be reproduced deterministically from source.
- Governing ADRs:
leafsleuth-ADR-001,leafsleuth-ADR-002,leafsleuth-ADR-003,leafsleuth-ADR-004,leafsleuth-ADR-005,ADR-129 - Verification Suites:
LagAttributionCalibrationTest,MechanismCausalLinkerCalibrationTest,CausalAutoTunerTest,EmpiricalFarmBlackBoxTest,BehavioralIdentityCalibrationTest,ConservationOfMassCalibrationTest,SupplyChainRingCalibrationTest,ResourceExhaustionCalibrationTest,InteractionAndProbeCalibrationTest,SleuthPaperCollectorAndS005Test
1. Asymmetric Statistical Operating Point & Summary Calibration Matrix¶
In a live Minecraft server, false positives and false negatives have asymmetric operational costs:
1. Hard Zero-False-Positive (0.000% FPR) Actuation & Conviction Floor: Throttling an innocent player's comparator item sorter, duplicating items in a virtualized farm, or falsely accusing shared-IP college roommates / modal survival players of ban evasion is unacceptable. All automated governor commitments and LIKELY_SAME_PERSON identity verdicts enforce strict lower-95% confidence/credible bounds ($L_{95}$) and multi-domain corroboration gates.
2. Active Probing & Tiered Sequential Recall to Collapse False Negatives: Passive heuristics alone suffer from collinearity, zero-variance steady states, boundary dilution, and short-session shrinkage. LeafSleuth recovers high statistical power ($1 - \beta$, True Positive Rate / Hit Rate) by pairing passive estimators with active counterfactual micro-dithering (DIAGNOSTIC_MICRO_DITHER), spatial/causal cluster aggregation, bounded taint-debt amortization, and tiered sequential hypothesis testing (Wald's SPRT, MODERATE_INVESTIGATE_LEAD, EMERGING_WATCHLIST_LEAD).
Summary Empirical Calibration Matrix¶
| Subsystem / Statistical Task | Empirical Hit Rate (TPR / Recall) | Empirical False Positive Rate (FPR) | Secondary / Watchlist Rate & Latency | Calibration Dataset & Sample Size |
|---|---|---|---|---|
Marginal Lag Attribution (LagAttributionEngine) |
100.0% Top-1 Hit Rate (40 / 40 seeded trials)92.5%+ CI95 Coverage of true 6.0 ms/tick culprit load |
0.000% Bystander Misattribution (0 / 40 trials; 100% cover 0.0 ms/tick & flag bystanderOnly) |
GC pause false attribution: 0.000% (posterior >= 0.50 subtracted) |
20-player, 50-chunk Monte Carlo simulation (40 random seeds x 200 ticks = 8,000 ticks) |
Periodic Redstone Clock Detection (LagPatternRecognizer) |
100.0% Period Recall (25 / 25 property trials across periods $k \in [3, 25]$ ticks + 8t/20t under Poisson mob noise) |
0.000% False Clock Classification on aperiodic / Poisson entity noise | Exact lag-$k$ period recovery within +- 0 ticks |
jqwik property-based suite + multi-frequency Poisson noise benchmarks |
Closed-Loop Pareto Governor (CausalAutoTuner) |
100.0% Commit Recall across single-chunk, 2x2 boundary-split (4 x 1.1 ms), collinear clocks, and zero-variance steady-state culprits |
0.000% Unwarranted Throttling on GC spikes, idle bystanders, composter-capped hoppers, comparator sorters, pistons, and low-slope novel designs | Live Paper 26.2-56 (bamboo): culprit marginal MSPT reduced 9.24 -> 2.90 ms/tick (-68.6%) |
Closed-loop plant simulations + live Paper 26.2-56 schematic harness |
Causal Farm Virtualization (EmpiricalFarmBlackBox) |
100.0% Virtualization Recall on pure generative DAGs & post-taint-flush webs (after $T_{\text{clean}} = 35$ clean ticks) | 0.000% Dupe / Taint Leakage (0 virtual items minted during player taint or burst-then-unload spikes) |
0 Schmitt-trigger state oscillations across 28m..40m boundary |
8-hop mob/crop DAG + single-item taint debt recovery + .litematic sugarcane benchmarks |
Kinematic Bot & Macro Detection (MotorKinematicsProfiler) |
100.0% Recall (4 / 4 archetypes): headless Baritone, 25%/75% toggle, GCD-smooth Baritone, and Gaussian/Uniform clickers |
0.000% (0 / 18) across real human gameplay traces (botLikelihoodScore < 0.45, HUMAN_QUANTIZED_GCD) |
Exact inverse-cubic sensitivity slider recovery within +- 0.5% |
18 real human traces (10 OpenAI VPT survival contractors + 8 Siddiqui same-PC mouse traces) |
X-Ray & Mining Trajectory Forensics (GoalMarkovProfiler) |
100.0% Surfacing on blind-redirect and cave-padded X-rayers (25.0% unexposed stone redirect yield) |
0.000% False Accusations on high-volume poke-hole branch miners (0.83% yield -> EXONERATED_LUCKY_MINER) |
Separates cave-padding decoys via per-segment blind stone redirect LLR | Branch-mining vs. cave-padded X-ray trajectory simulations |
Modal Population & Doppelganger Identity Matrix (BayesianIdentityLinker) |
100.0% Evader Lead Recall (12 / 12):- 9 / 9 (100%) non-silent evaders at LIKELY_SAME_PERSON- 3 / 3 (100%) silent VPN + DPI-rescaled evaders at MODERATE_INVESTIGATE_LEAD |
0.000% (0 / 3,148) false LIKELY_SAME_PERSON convictions:- 0 / 4 distinct-IP doppelgangers- 0 / 4 shared-IP disjoint-base doppelgangers- 0 / 4 active roommate pairs |
Soft MODERATE_INVESTIGATE_LEAD birthday overlap: 1.302% (41 / 3,148)(Down from 3.875% / 122 / 3,148 hard false convictions before modal base-rate discounting) |
$N = 80$ accounts ($\binom{80}{2} = 3,160$ pairs) sampled from modal survival priors (seed = 20261029L) |
Large-Population Multi-Epoch Scaling (RelativeCoordinateSpace + Wald's SPRT) |
100.0% Top-1 Banned Account Recovery (topBanMatchPosterior = 0.993, ACCEPT_H1_SAME_PERSON on multi-context evader) |
0.000% Drift Misattribution (ACCEPT_H0_DISTINCT_PEOPLE on veteran after 4 epochs of hardware/playstyle drift) |
Initial 10K sync: 1,311 msSingle-player IVM update: 8 msBan-evasion query: 9 ms |
$N = 10,000$ accounts ($\binom{10,000}{2} = 49,995,000$ possible pairs) across 4 decay epochs (seed = 20261030L) |
Economy Conservation-of-Mass & Zero-Day Dupe Audit (ConservationOfMassAuditor) |
100.0% Dupe Recall (30 / 30) across chunk-unload/reload, portal, & packet-desync exploits |
0.000% (0 / 90) across 99.9th-percentile Fortune III streaks, 27-slot shulker unpacks (43,200 items), & clean farm black-box yields |
$L_{95}(\Delta M) > \tau_{\text{mass}}$ + Wald SPRT $\Lambda_T \ge 4.60\text{ nats}$ ($\alpha \le 0.01$) | 120 seeded mass-balance trials (ConservationOfMassCalibrationTest) |
Supply-Chain Laundering & Pre-Grief Recon Rings (SupplyChainRingForensics) |
100.0% Ring & Scout Recall (3 / 3 archetypes) (3-hop dead-drop chain $P \ge 0.90$, pre-grief scout pair $P \ge 0.85$) |
0.000% (0 / 30) across public spawn charity chests ($H_{\text{visitors}} \ge 1.8\text{ nats}$), bilateral shop districts, & shared faction bases |
Time-decayed directed custody graph + 64-hash MinHash trajectory lead-lag cross-correlation |
Multi-hop muling & pre-grief recon suite (SupplyChainRingCalibrationTest) |
Resource Exhaustion, NBT Bomb & OldGen Leak Forensics (ResourceExhaustionForensics) |
100.0% Top-1 Culprit Recall (60 / 60) across spiral chunk-gen churn (20/20), nested NBT bombs (20/20), & S-002 OldGen heap ratchets (20/20) |
0.000% (0 / 60) on fast linear Nether-roof highway travelers & YoungGen GC sawtooth bystanders |
EwmaLinearFit marginal allocation (MB/t) & MSPT regression + post-GC OldGen slope (MB/min) |
120 seeded resource & memory trials (ResourceExhaustionCalibrationTest) |
Action-Entropy Macro Ergonomics & Sub-Tick Packet Probing (InteractionAndProbeForensics) |
100.0% Macro & Prober Recall (60 / 60) across Gaussian auto-fishers (15/15), lectern macros (15/15), grinder timers (15/15), & sub-tick packet probers (15/15) |
0.000% (0 / 60) across 1/f pink-noise human fishers, shift-click villager traders, & rhythmic human combat |
Shannon interval entropy + Allan variance $1/f$ slope + Fitts/Hick latency + sub-tick EwmaLinearFit |
120 seeded macro & packet-probe trials (InteractionAndProbeCalibrationTest) |
2. Lag Attribution, Clock Periodicity & Closed-Loop Governor Methodology¶
2.1 Confounder-Resistant Marginal Regression (LagAttributionEngine)¶
Naive event counters blame whichever chunk has the most hoppers or whichever player happens to stand near an active contraption. LagAttributionEngine isolates true marginal tick cost via an exponentially weighted moving-average linear regression (EwmaLinearFit) paired with a fast-adapting 16-tick regime-shift tracker:
$$\Delta Y_t = \max\left(0,\; \text{MSPT}t - \widehat{\text{MSPT}}_t\right)$$},t} - \widehat{\text{GC}
For each active chunk $c$, activity share $S_{c,t} = X_{c,t} / \sum_j X_{j,t}$ is regressed against $\Delta Y_t$ to estimate the marginal contribution $\hat{\beta}_c$ and its 95% confidence interval:
$$\text{CI}_{95}(\hat{\beta}_c) = \left[\max(0,\; \hat{\beta}_c - 1.96 \cdot \text{SE}(\hat{\beta}_c)),\; \hat{\beta}_c + 1.96 \cdot \text{SE}(\hat{\beta}_c)\right]$$
- Bystander Immunity: Player attribution partitions chunk cost strictly by active causal initiation recorded in
WorkAttributionLedgerplus exponentially decaying builder provenance ($\tau = 30\text{ min}$). A passive bystander standing inside a6.0 ms/ticklag chunk with zero causal initiations receives0.00 ms/tick(bystanderOnly = truein40 / 40Monte Carlo trials).
2.2 Closing False Negatives in Closed-Loop Governance (CausalAutoTuner)¶
CausalAutoTuner requires $L_{95}(\hat{\beta}_c) \ge \text{minGuaranteedMspt}$ (1.5 ms) before probing a chunk, and requires ComparativeCostGovernor to confirm a statistically significant MSPT reduction ($\ge 0.75\text{ ms}$) during a 20-tick TRIAL_PROBE before committing (COMMITTED_ACTIVE). Two structural false-negative regimes were identified and resolved:
- Low-Variance Steady-State Blind Spot ($\text{Var}(S_{c,t}) \to 0$):
- Failure Mode: When an always-loaded chunk (spawn, chunk-loader, or stationary AFK bot) runs a heavy contraption at constant load ($X_{c,t} \approx \text{const}$), regressor variance $\text{Var}(S_{c,t})$ approaches zero, inflating $\text{SE}(\hat{\beta}c)$ and holding $L$ indefinitely.} < 1.5\text{ ms
- Countermeasure (
DIAGNOSTIC_MICRO_DITHER): When server MSPT is elevated and a chunk exhibits a high point/upper estimate ($\hat{\beta}c \ge 1.5\text{ ms}$ or $U$) with low activity variance (} \ge 3.5\text{ ms}$) but wide uncertainty ($U_{95} - L_{95} > 1.75\text{ msLagAttributionEngine#isLowVarianceSteadyCandidate),CausalAutoTunerentersTunerTrialState.DIAGNOSTIC_MICRO_DITHER. It injects a non-destructive 2-tick cadence deferment window per 20-tick half-cycle (40-tick period), creating controlled counterfactual variance in $S_{c,t}$. - Empirical Result: In
CausalAutoTunerTest#zeroVarianceSteadyStateCulpritConvergesViaDiagnosticMicroDither, an un-dithered constant-rate culprit (4.5 ms/tick) remains stuck at $L_{95} = 0.00\text{ ms}$ (OBSERVE_ONLY), whereasDIAGNOSTIC_MICRO_DITHERcollapses the CI within one 40-tick cycle and commits throttling while leaving an innocent constant-rate chunk untouched. - Multi-Chunk Contraption Dilution (
2x2/3x3Boundary Splitting): - Failure Mode: A single
4.4 ms/tickcontraption built across a2x2chunk intersection splits its load into four1.1 ms/tickper-chunk buckets ($L_{95} \approx 0.8\text{ ms} < 1.5\text{ ms}$ each), evading per-chunk thresholds. - Countermeasure (Connected-Component Spatial/Causal Clustering):
CausalAutoTunergroups adjacent active chunks (Chebyshev distance <= 1or linked viaEmpiricalFarmBlackBox#areChunksCausallyLinked) into connected contraption clusters and computes the variance-summed cluster confidence interval: $$\hat{\beta}{\text{cluster}} = \sum}} \hat{\betac, \qquad \text{SE}$$}} = \sqrt{\sum_{c \in \mathcal{C}} \text{SE}(\hat{\beta}_c)^2 - Empirical Result: Verified in
CausalAutoTunerTest#multiChunkSplitContraptionDetectedViaConnectedClusterAttribution(4 x 1.1 mschunks aggregate to $L_{95}(\hat{\beta}_{\text{cluster}}) > 1.5\text{ ms}$ and commit proportionally across all 4 member chunks).
2.3 Bounded Taint Debt & Provenance Flushing (EmpiricalFarmBlackBox)¶
- Anti-Duplication Invariant: Virtual yield in
GOVERN_BLACKBOXmode must never synthesize player-deposited or unverified items. - False-Negative Countermeasure: Previously, a single accidental player item drop into a water stream permanently poisoned the entire connected container web (
TAINTED_BY_PLAYER_WEB).EmpiricalFarmBlackBoxnow tracks unverified item debt (taintedItemDebt+lastTaintTick) perContainerNodeand connected component: - While
taintedItemDebt > 0or withintaintCleanWindowTicks($T_{\text{clean}} = 35\text{ ticks}$) of the last taint event, virtualization is held inTAINTED_BY_PLAYER_WEB/TAINT_FLUSHINGand virtual reservoirs remain zeroed. - Player-deposited items are excluded from the learned
yieldPerCausalEventnumerator anditemRatesPerMinuteEWMA so a player cannot inflate a farm's virtual minting rate. - As verified
PROVABLE_GENERATIVEarrivals (MOB_NATURAL_SPAWN,CROP_NATURAL_GROWTH,BLOCK_GENERATOR_FORM) flow through the web during the quiet window,taintedItemDebtamortizes down to0, automatically rehabilitating the web toPROVABLE_GENERATIVEandVIRTUALIZED_ACTIVE(EmpiricalFarmBlackBoxTest#singleStrayPlayerDropRecoversAfterTaintDebtFlushesWithoutInflatingYieldRate).
3. Kinematic Biometrics, Hardware GCD Canopy & 2D Fitts's Law Reconstruction¶
3.1 Exact 3D Camera Sensitivity & Asymmetric X/Y Hardware GCD Prior (MotorKinematicsProfiler)¶
Minecraft quantizes mouse camera deltas $(\Delta\text{yaw}, \Delta\text{pitch})$ by the cubic polynomial:
$$f(s) = (0.6 s + 0.2)^3, \qquad \Delta\theta_{\text{GCD}}(s) = 0.15 \cdot f(s)\text{ degrees}$$
MotorKinematicsProfiler extracts both the primary yaw quantization step $\Delta\theta_{\text{yaw,GCD}}$ and the vertical pitch quantization step $\Delta\theta_{\text{pitch,GCD}}$ via pairwise floating-point GCD with harmonic de-aliasing ($m \in {2, 3, 4, 5, 6}$):
- Inverse-Cubic Slider Recovery: Recovers the player's exact client sensitivity percentage $s \in [0\%, 200\%]$ and tracks up to two sensitivity modes to accommodate hardware DPI-clutch toggles (100% <-> 60%).
- Asymmetric Hardware X/Y Aspect-Ratio Canopy (pitchToYawGcdScaleRatio, exactHardwareGcdCanopyKey): Mice configured with asymmetric X/Y DPI stages or custom raw-input aspect ratios exhibit a deterministic ratio $\rho_{p/y} = \Delta\theta_{\text{pitch,GCD}} / \Delta\theta_{\text{yaw,GCD}} \ne 1.0$. BayesianIdentityLinker indexes exactHardwareGcdCanopyKey as an $O(N)$ inverted blocking canopy so non-modal hardware configurations partition the population before running full pairwise comparisons.
3.2 Sensitivity-Invariant 2D Fitts's Law Container Cursor Reconstruction¶
When a ban evader switches to a VPN, stays silent in chat, and alters their 3D camera sensitivity slider, 3D angular speed changes. However, inside container GUIs (InventoryClickEvent), cursor movement between slot indices $s_{k-1} \to s_k$ on the 9-column grid is governed by 2D OS cursor ergonomics and Fitts's Law:
$$D_k = \sqrt{(c_k - c_{k-1})^2 + (r_k - r_{k-1})^2}, \qquad \text{ID}_k = \log_2(1 + D_k)\text{ bits}$$
$$\Delta t_k = a_{\text{react}} + b_{\text{fitts}} \cdot \text{ID}_k$$
MotorKinematicsProfiler fits an online EwmaLinearFit of inter-click duration $\Delta t_k \in [25\text{ ms}, 1200\text{ ms}]$ against the Shannon Index of Difficulty $\text{ID}k$ across non-identical slots ($D_k \ge 0.5$), extracting:
- **Reaction/Click Dwell Intercept ($a$, ms):** Biomechanical click initiation latency.
- }Fitts Index of Performance ($\text{IP} = 1000 / b_{\text{fitts}}$, bits/s): 2D cursor throughput, which remains invariant whether the player performs short adjacent slot clicks or long diagonal chest sweeps (verified in BehavioralIdentityCalibrationTest: recovering $a_{\text{react}} = 65.0 \pm 6.0\text{ ms}$ and $\text{IP} = 22.2 \pm 2.5\text{ bits/s}$ across disjoint slot jump distances).
4. Modal Population Calibration, Behavioral Doppelgangers & The Irreducible Collision Floor¶
4.1 Why Uncalibrated Heuristics Fail on Modal Players¶
In any large server population, a substantial fraction of players cluster around modal defaults:
- ~50% leave Minecraft's mouse sensitivity at the vanilla default (100.0% $\implies \Delta\theta_{\text{GCD}} = 0.1500^\circ$).
- ~55% type casual lowercase chat at 54-70 WPM, use default /home and /msg commands, and talk about everyday survival items (diamonds, villager, iron, redstone, hopper, nether).
- ~55% follow the standard survival state loop (MINING -> CONTAINER_SORT -> CRAFTING -> BUILDING).
To measure this effect rigorously, BehavioralIdentityCalibrationTest#expandedModalPopulationAndNearIdenticalDoppelgangerStatisticalCalibration constructs an $N = 80$ account population ($\binom{80}{2} = 3,160$ total account pairs):
- 40 Innocent Modal Survival Accounts (SyntheticActorPersona#sampleModalSurvivalPersona), including 4 shared-IP household roommate pairs with simultaneous active input overlap ($r \approx 0$).
- 8 Explicit Near-Identical "Behavioral Doppelganger" Pairs (16 distinct accounts) created via SyntheticActorPersona#createNearIdenticalDoppelgangerPair (< 2% kinematic/cadence delta between two distinct humans):
- 4 pairs on distinct IPs/subnets sharing 100% default sensitivity, ~62 WPM lowercase chat, /home & /msg, common survival vocabulary, and identical Markov transition loops at separate bases.
- 4 pairs on shared household IPs playing time-shifted (non-overlapping sessions) with near-identical modal habits at separate bases (spatialOverlapJaccard == 0.0).
- 12 True Same-Person Ban-Evader Pairs (24 accounts) across 4 evasion regimes:
- 3 Shared-IP multibox/sequential alt pairs.
- 3 /24 subnet router-reboot evader pairs.
- 3 VPN-spoofed chatting evader pairs with shared stash transfers / personal shibboleths.
- 3 Silent (0 chat messages), VPN-spoofed (sameSubnet = false), DPI-rescaled evader pairs matched via 2D container ergonomics and early-session action bigrams.
4.2 Before vs. After Information-Theoretic Base-Rate Discounting¶
In according with Shannon self-information $I(f) = -\ln P(f \mid H_0)$, sharing a feature that 50% of the server shares carries almost zero forensic weight, whereas sharing a rare feature ($P(f \mid H_0) \ll 1\%$) carries high weight. Three mathematical calibrations were applied:
1. Population-Scaled Common Lexicon Virtual Prior (LinguisticStylometryProfiler): Common Minecraft terms (COMMON_MINECRAFT_LEXICON) are seeded with a virtual baseline document frequency $\text{df}_{\text{prior}} = \max(16, N / 7)$ so everyday game words never cross RARE_SHIBBOLETH_MIN_IDF (3.6), and modal lowercase / default /home & /msg / 48-74 WPM habits are discounted by 0.30x-0.35x when uncorroborated by rare shibboleths.
2. Default 100% Sensitivity & Disjoint-Base Markov Discounting (MotorKinematicsProfiler, GoalMarkovProfiler): Matching the vanilla 100% sensitivity bucket (0.1485-0.1515 deg) without corroborating locomotion or 2D container ergonomics is discounted from +0.85 to +0.18, and coarse survival Markov cosine similarity when spatialOverlapJaccard == 0.0 and contraptionCosine == 0.0 is discounted by 0.25x.
3. Forensic Anchor Gate & Household Disjoint-Base Guard (BayesianIdentityLinker):
- Accounts on distinct IPs/subnets with zero high-entropy forensic connectors (hasHighEntropyForensicAnchor == false: no shared rare shibboleth, no shared stash chunk/container Jaccard, no directed item transfer, and no matching 64-bit bot script hash) are capped at MODERATE_INVESTIGATE_LEAD (never convicted as LIKELY_SAME_PERSON).
- Accounts sharing a household IP/subnet that play at completely disjoint bases (spatialOverlapJaccard < 0.02, zero item transfers, zero rare shibboleths, and no negative temporal multibox switching) classify as SHARED_ENV_AMBIGUOUS or MODERATE_INVESTIGATE_LEAD.
Metric Across All $\binom{80}{2} = 3,160$ Pairs (12 Evader Pairs, 3,148 Innocent Pairs) |
Uncalibrated Naive Pillar Sum | Calibrated Base-Rate & Anchor Gate |
|---|---|---|
True Evader Lead Recall (TP_AnyLead) |
12 / 12 (100.0%) |
12 / 12 (100.0%) (9/9 non-silent at LIKELY_SAME_PERSON, 3/3 silent VPN at MODERATE_INVESTIGATE_LEAD) |
Innocent False Convictions (FP_LikelySame) |
122 / 3,148 (3.875%) |
0 / 3,148 (0.000%) |
| Distinct-IP Doppelganger False Convictions | 4 / 4 (100.0%) |
0 / 4 (0.000%) |
| Shared-IP Time-Shifted Doppelganger False Convictions | 4 / 4 (100.0%) |
0 / 4 (0.000%) (classified as SHARED_ENV_AMBIGUOUS / MODERATE_INVESTIGATE_LEAD) |
Soft MODERATE_INVESTIGATE_LEAD Birthday Overlap |
284 / 3,148 (9.022%) |
41 / 3,148 (1.302%) |
4.3 Proof of the Irreducible Statistical Collision Floor¶
In any statistical identity system, there is a theoretical boundary where two distinct humans become indistinguishable from passive telemetry alone. BehavioralIdentityCalibrationTest explicitly proves this boundary (IdenticalTwinA vs IdenticalTwinB):
- When two distinct people share the same household IP, play time-shifted (never simultaneously online, so temporal correlation $r$ cannot separate them), use the same non-default mouse sensitivity (82%), type at the same WPM with the same rare shibboleth ("quasiconnectivity"), and access the same community chest chunks, linker.compare(twinA, twinB) returns LIKELY_SAME_PERSON.
- This is why /sleuth compare and /sleuth cases are designed as explainable decision-support tools for human moderators rather than automated ban triggers.
5. Database-Grade Scaling ($N = 10,000+$ Accounts), Concept Drift & Wald's SPRT¶
On large long-running servers ($N = 10,000$ to $100,000+$ cumulative accounts over multiple seasons), naive pairwise comparison requires $\binom{10,000}{2} = 49,995,000$ evaluations, per-player history sets grow unboundedly, and player/server behavior drifts over months. LeafSleuth applies six techniques from database systems, streaming algorithms, and sequential analysis (BehavioralIdentityCalibrationTest#tenThousandAccountMultiEpochDriftAndDatabaseScalingCalibration):
5.1 Inverted Blocking Indexes, 8-Byte SimHash Canopies & $O(D)$ Incremental View Maintenance¶
- Inverted Equivalence Indexes (
RelativeCoordinateSpace,BayesianIdentityLinker): Mandatory candidate pairs sharing anipHash,subnetHash, orexactHardwareGcdCanopyKeyare emitted via inverted index buckets in $O(\sum_b |B_b|^2) = O(N)$ time rather than scanning an $O(N^2)$ nested loop. - 8-Byte Charikar
SimHashCanopy Sybil Weighting: Sybil clone-cluster down-weighting ($w_i = 1 / m_i$ for dense bot swarms) groups accounts by the 8 bytes of their 64-bitSimHashsignature and only evaluates exact cosine/Hamming distance within matching byte canopies (capping per-bucket probes at64), reducing initial population centroid computation from $O(N^2)$ to $O(N)$. - Algebraic $O(D)$ Incremental View Maintenance (IVM):
RelativeCoordinateSpacemaintains running weighted coordinate sums $(\sum w_i,\; \sum w_i y_{i,d},\; \sum w_i y_{i,d}^2)$ across $D = 16$ axes and $64$ hyperplanes, andBayesianIdentityLinkertracks per-player telemetry activity stamps (lastSyncedTelemetryStamp). When an online player records new telemetry in a 10,000-account server, only that player's feature vector is delta-updated ($O(D)$ subtract-old / add-new), reducing single-player coordinate update latency from754 ms($O(N \cdot D)$ full rescan) to8 ms.
5.2 Bounded Streaming Sketches (MinHash + Space-Saving)¶
To bound per-player memory to a constant footprint (< 8 KB/account) over multi-year horizons without freezing learning after the first 512 events:
- 256-Entry LRU Ring + 64-Hash MinHash Spatial Sketch (GoalMarkovProfiler): visitedChunks and accessedContainers maintain a sliding 256-entry LinkedHashSet LRU ring buffer of recent coordinates alongside a 64-function SplitMix64 bottom-$k$ MinHash signature (long[64]). Even after a veteran explores > 700 chunks and moves to a new secret stash base, spatialSketchJaccard accurately detects shared stash overlap (> 0.02 Jaccard vs 0.00 under a static 512-entry cap).
- Metwally Space-Saving / IDF-Weighted Lexical & Trigram Sketch (LinguisticStylometryProfiler): Per-player token maps are bounded at MAX_LEXICAL_SKETCH_SIZE = 512 (and MAX_TRIGRAM_SKETCH_SIZE = 256), evicting the lowest-utility entry ($\text{count} \times \text{IDF}$) when full so new recurrent personal shibboleths ("chronosynthesis_v4") are retained even after typing > 600 unique filler words.
5.3 Bitemporal Forward Epoch Decay & Point-in-Time Frozen Ban Snapshots¶
To handle macro-drift (server meta/version shifts) and micro-drift (a legitimate player buying a new mouse or changing playstyle over months):
- Active Account Exponential Epoch Decay (BayesianIdentityLinker#advanceBehavioralEpoch): Multiplies active unbanned players' Markov transition matrices, lexical/trigram counts, Welford kinematic sample weights, and global document frequencies by a forward decay factor $\gamma \in (0, 1)$ at epoch boundaries, and prunes stale secondary GCD sensitivity modes whose window support falls below < 20%.
- Frozen Point-in-Time Ban Snapshots (isAccountBanned): When an account is marked banned (markAccountBanned), its 4-pillar behavioral sketch is frozen at the moment of the ban (advanceBehavioralEpoch skips banned accounts). Because ban evaders create replacement alts using their current hardware and muscle memory at the time of the ban, frozen ban snapshots remain sharp (topBanMatchPosterior = 0.993 in Epoch 4) while innocent veterans who drifted away from their Epoch 1 hardware are cleanly exonerated (ACCEPT_H0_DISTINCT_PEOPLE).
5.4 Wald's Sequential Probability Ratio Test (SPRT) Across Independent Contexts¶
Instead of making a static single-window classification decision, BayesianIdentityLinker#evaluateWaldSprt(accountA, accountB, alpha, beta) accumulates shrunk log-likelihood ratios across 5 independent behavioral domains:
1. LINGUISTIC_STYLOMETRY (chat vocabulary, shibboleths, WPM burst regression)
2. CAMERA_3D_KINEMATICS (3D yaw/pitch GCD quantization, X/Y aspect ratio, turn dynamics)
3. INVENTORY_FITTS_ERGONOMICS (2D container cursor Fitts's Law intercept $a_{\text{react}}$ & throughput $\text{IP}$)
4. TEMPORAL_CORRELATION (simultaneous vs alternating active input burst correlation $r$)
5. GOAL_MARKOV_SPATIAL (gameplay state transitions, contraption signatures, MinHash spatial Jaccard)
Given target false-positive ceiling $\alpha$ (default 0.01) and false-negative ceiling $\beta$ (default 0.05), Wald's stopping boundaries are:
$$A = \ln\left(\frac{1 - \beta}{\alpha}\right) \approx +4.55\text{ nats}, \qquad B = \ln\left(\frac{\beta}{1 - \alpha}\right) \approx -2.99\text{ nats}$$
ACCEPT_H1_SAME_PERSON: Requires $\Lambda_T \ge A$ and positive corroboration ($\text{LLR}_c \ge +0.35$) across at least2independent behavioral contexts (or1context + a hard forensic anchor). Transient single-context camera collisions (SingleContextPeer) remain inCONTINUE_SAMPLINGuntil secondary domains corroborate or refute the link.ACCEPT_H0_DISTINCT_PEOPLE: Triggered when $\Lambda_T \le B$ (or when simultaneous independent play $r \approx 0$ refutes single-operator control).
6. Head-to-Head Empirical Comparison Against Conventional Detection Mechanisms¶
To demonstrate why LeafSleuth's statistical architecture outperforms conventional server tools, LagAttributionCalibrationTest#headToHeadComparisonAgainstConventionalLagDetectors and BehavioralIdentityCalibrationTest#headToHeadComparisonAgainstConventionalBehavioralAndCheatDetectors evaluate standard industry baseline mechanisms on the exact same benchmark datasets:
| Problem Domain | Detection Mechanism | Empirical Hit Rate (TPR / Recall) |
Empirical False Positive Rate (FPR) |
Structural Failure Mode of Baseline |
|---|---|---|---|---|
Chunk Lag Attribution (40 Monte Carlo trials, 80-event capped base vs 35-event entity/rail culprit) |
Raw Event / Tile-Entity Counter (e.g., ClearLag / entity-count scanners) | 0.0% (0 / 40 Top-1 hits) |
100.0% (40 / 40) false positives on optimized composter-capped bases |
Counts event volume ($X_{c,t}$) instead of marginal $\Delta\text{MSPT}$ per event |
Chunk Lag Attribution (40 trials with 22 ms JVM GC pauses) |
Unadjusted Peak-Tick Spike Detector (no GC posterior subtraction) | 0.0% (0 / 40 on GC ticks) |
100.0% (40 / 40) false attribution to innocent chunks active during GC |
Blames whatever chunk happens to fire during a Stop-The-World GC pause |
Player Lag Attribution (40 trials with passive visitor in hotspot chunk) |
Proximity / Chunk-Loader Blame (equal credit to chunk residents) | 100.0% (includes owner) |
100.0% (40 / 40) innocent bystander misattribution (>= 3.0 ms/t) |
Credits chunk lag by spatial proximity rather than causal initiation |
| Chunk & Player Lag Attribution | LeafSleuth (LagAttributionEngine + DIAGNOSTIC_MICRO_DITHER + Cluster CI95) |
100.0% (40 / 40 Top-1) |
0.000% (0 / 40) on capped bases, bystanders, and GC pauses |
Regresses marginal $\Delta\text{MSPT}_t$ after GC subtraction with active micro-dither probing |
Ban-Evasion / Alt Linking (N = 80, 3,160 pairs: 12 evaders, 8 shared-IP household pairs) |
Exact IP Hash Match Only (standard /alts IP check) |
25.0% (3 / 12 evaders) |
100.0% (8 / 8) false convictions on shared-IP roommates & households |
Misses /24 router reboots & VPNs; convicts every sibling/college roommate |
Ban-Evasion / Alt Linking (N = 80, 3,160 pairs) |
/24 Subnet Prefix Match Only |
50.0% (6 / 12 evaders) |
100.0% (8 / 8) false convictions on shared-IP households + ISP neighbors |
Misses all VPN/cellular evaders; flags entire dorm/ISP /24 blocks |
Ban-Evasion / Alt Linking (N = 80, 3,160 pairs) |
Uncalibrated Behavioral Sum (no modal base-rate discounting) | 100.0% (12 / 12 evaders) |
3.875% (122 / 3,148) hard false convictions (9.022% soft overlap) |
Collides on modal defaults (100% vanilla sensitivity, /home, common survival loops) |
Ban-Evasion / Alt Linking (N = 80, 3,160 pairs) |
LeafSleuth (BayesianIdentityLinker + Modal Discounting + 2D Fitts + Wald's SPRT) |
100.0% (12 / 12 leads) (9/9 non-silent LIKELY_SAME, 3/3 silent VPN LEAD) |
0.000% (0 / 3,148) false LIKELY_SAME_PERSON convictions (0 / 8 households) |
Discounts low-information modal defaults; bounds unanchored birthday overlap to 1.302% with explicit confidence reporting |
Bot & Macro Detection (4 bot archetypes vs 18 real human traces) |
Static CPS Cap / Zero-Variance Check ($\text{CPS} > 15$ or $\text{CV} < 0.03$) | 25.0% (1 / 4 archetypes) |
0.0% on normal CPS; flags legitimate drag/butterfly bursts |
Misses Gaussian/Uniform randomized ghost clickers (11-13 CPS) & Baritone |
Bot & Macro Detection (4 bot archetypes vs 18 real human traces) |
Session-Average Yaw/Pitch GCD Only | 25.0% (1 / 4 archetypes) |
Falsely flags human hardware DPI-clutch toggles (100% <-> 60%) |
Misses GCD-quantized Baritone and 25% bot / 75% human regime-toggle play |
Bot & Macro Detection (4 bot archetypes vs 18 real human traces) |
LeafSleuth (MotorKinematicsProfiler bimodal GCD + windowed regime toggle) |
100.0% (4 / 4 archetypes) |
0.000% (0 / 18) across OpenAI VPT & Siddiqui real human traces |
Tracks bimodal DPI clutches, 24-sample sliding windows, and interval kurtosis |
| X-Ray Forensics (blind-redirect & cave-padded X-ray vs branch & cave miners) | Raw Diamond-to-Stone Ratio ($\text{ores} / \text{stone} > \tau$) | 50.0% (misses cave-padded X-rayers) |
Falsely accuses lucky open-cave explorers who break little solid stone | Defeated by mining decoy stone in caves to dilute total ore ratio |
| X-Ray Forensics | LeafSleuth (GoalMarkovProfiler air-exposure + blind redirect LLR) |
100.0% (2 / 2 archetypes) |
0.000% (EXONERATED_LUCKY_MINER on branch & cave miners) |
Scores per-segment blind stone heading redirects into unexposed veins |
7. Operator Confidence-Level Reporting & The LeafConfig 4-Pillar Identity Forensics Panel¶
Because passive behavioral telemetry across thousands of accounts inevitably exhibits a small birthday-paradox overlap on unanchored leads (1.302% at MODERATE_INVESTIGATE_LEAD vs 0.000% at LIKELY_SAME_PERSON), LeafSleuth surfaces the calibrated confidence level, 95% credible interval, and both supporting and exonerating evidence on every operator interface:
- In-Game Book & Chat Confidence Guidance (
SleuthMenuBuilder#formatConfidenceGuidance,SleuthCommandRegistrar,ModeratorInvestigationCase#formatTriageLine): - Every
/sleuth report,/sleuth compare,/sleuth evaders, and/sleuth casesentry displaysP(Same) = XX.X% +- YY.Y% [95% CI: L95%..U95%], the natural-log Bayes factor (Log-BF), and a tier-specific confidence note:LIKELY_SAME_PERSON: "Corroborated multi-domain match (0.00% calibrated modal FPR)"MODERATE_INVESTIGATE_LEAD: "Advisory lead (~1.3% modal birthday overlap; verify in spectator)"EMERGING_WATCHLIST_LEAD: "Emerging watchlist lead (Needs ~N more motor/inventory samples to reach 95% CI link threshold)"SHARED_ENV_AMBIGUOUS: "Shared IP/subnet with disjoint bases (household / roommate candidate)"LIKELY_DISTINCT_PEOPLE: "Exonerated distinct operators (simultaneous independent play)"
LeafConfigWeb Workspace 4-Pillar Identity Forensics Panel (ConfigRemediationBridge#toWebIdentityComparisonCard,LeafConfigRuntime.LeafSleuthIdentityComparisonEntry,#sleuth-identity-cards-grid):ConfigRemediationBridge.toWebIdentityComparisonCard(IdentityComparisonReport)andLeafConfigRuntime.LeafSleuthMapContributionexport structuredidentityComparisonscards directly into theLeafConfigDiagnostics & Radar panel (plugins/LeafConfig/docs/editor/index.html).- Each card renders the pair's
P(Same Operator) +- halfWidth [95% CI: L95%..U95%], network origin badge (Shared IP / HouseholdvsDistinct IP / VPN), verdict badge, and confidence-level guidance banner above two side-by-side forensic columns:- Why Same Operator (
whySameTells,LLR > 0): Lists the positive supporting indicators (e.g., matching 2D Fitts container cursor throughput $\text{IP}$ and intercept $a_{\text{react}}$, asymmetric X/Y hardware GCD ratio, shared rare shibboleth IDF, alternating multibox input bursts, or shared secret stashMinHashJaccard). - Why Distinct People (
whyDistinctTells,LLR < 0/ Base-Rate Caveats): Lists the exonerating evidence or counter-hypothesis caveats (e.g., simultaneous independent active input $r \approx 0$ proving two distinct hands on a shared household router, divergent wrist-vs-arm turn kinematics, disjoint survival bases with zero item transfers, or the~1.3%modal birthday-overlap caveat when no shared IP or stash anchor is present).
- Why Same Operator (
- Unified Causal Server Forensics Engines & Independent
forensics.*Toggles (leafsleuth-ADR-005,/sleuth audit <economy|rings|resources|probes>): ModeratorInvestigationCase,ConfigRemediationBridge#toWebForensicCaseCard,SleuthMenuBuilder#buildForensicCasesPage, andLeafConfigRuntime.LeafSleuthForensicCaseEntrysurface all four causal server forensic engines (ConservationOfMassAuditor,SupplyChainRingForensics,ResourceExhaustionForensics,InteractionAndProbeForensics) withCI95bounds and side-by-sidesupportingTellsvs.exoneratingTells.- Each engine is gated by an independent toggle in
plugins/LeafSleuth/config.yml(forensics.economyDupeAuditEnabled,forensics.supplyChainRingEnabled,forensics.resourceExhaustionEnabled,forensics.macroAndProbeEnabled).
8. Reproducing the Calibration Benchmarks from Source¶
Run the automated calibration suites directly via the Gradle wrapper:
# Run all lag attribution, periodicity, governor, farm black-box, behavioral, and forensic engine calibration suites
./gradlew :plugins:LeafSleuth:sleuth-common:test :plugins:LeafSleuth:sleuth-paper:test :plugins:LeafConfig:leafconfig-common:test --info
# Run the four unified causal server forensics calibration suites (leafsleuth-ADR-005)
./gradlew :plugins:LeafSleuth:sleuth-common:test --tests "*ConservationOfMassCalibrationTest*" --tests "*SupplyChainRingCalibrationTest*" --tests "*ResourceExhaustionCalibrationTest*" --tests "*InteractionAndProbeCalibrationTest*" --info
# Run the head-to-head lag & behavioral comparisons against conventional detection mechanisms
./gradlew :plugins:LeafSleuth:sleuth-common:test --tests "*LagAttributionCalibrationTest.headToHeadComparisonAgainstConventionalLagDetectors*" --tests "*BehavioralIdentityCalibrationTest.headToHeadComparisonAgainstConventionalBehavioralAndCheatDetectors*" --info
# Run the N = 80 (3,160-pair) modal population & near-identical doppelganger calibration matrix
./gradlew :plugins:LeafSleuth:sleuth-common:test --tests "*BehavioralIdentityCalibrationTest.expandedModalPopulationAndNearIdenticalDoppelgangerStatisticalCalibration*" --info
# Run the N = 10,000 (49,995,000-pair) multi-epoch concept drift, MinHash/Space-Saving, Fitts's Law & Wald SPRT benchmark
./gradlew :plugins:LeafSleuth:sleuth-common:test --tests "*BehavioralIdentityCalibrationTest.tenThousandAccountMultiEpochDriftAndDatabaseScalingCalibration*" --info