RESTRICTED — TIER III — DO NOT DISTRIBUTE
TASHUNKA ENGAGEMENT PROBABILITY INDEX TEPI v4.2.1 · PREDICTIVE OBSERVATION MODEL FOR THE T. THURSFIELD SRI LANKA ITINERARY PORTAL (STIP-1)
TELEMETRY LIVE--:--:--AEST/UTC+10
SUBJECT A — DOSSIER PRIMARY TARGET
SUBJ-A · CAP 03:41 · CONF 0.97
TASHUNKAdesignation: THE OBSERVER
Facial enthusiasm amplitude8.7 / 10
Micro-expression receptivity0.614 ↓
Detected affective statePRE-EMPTIVE AGREEABLENESS
Reliability half-life14.2 h
Inferred beverage stateNON-ZERO
Historical follow-through41.7% → 12.1%
Says "yeah legend" rate94.2%
Actually opens it rate8.4%
Uselessness Coefficient (κ)0.00
ANALYST NOTE: Subject presents high surface-level agreeableness against a documented 85.8 pp gap between stated and revealed intent. Engagement decays on a measurable exponential and does not recover. Prior classification of this divergence as "model uncertainty" has been formally withdrawn; it is now modelled as a stable trait.
SUBJECT B — DOSSIER ASSET / AUTHOR
SUBJ-B · CAP 13:08 · CONF 0.93
T. THURSFIELDdesignation: THE AUTHOR
Enthusiasm reservoir94%
Est. hours spent building site31.4 h
Beverage-assisted confidenceELEVATED
Composure under non-responseDEGRADING
Links deployed (11 d)4
Emotional stake in outcomeCRITICAL
Self-reported "it's not a big deal"FALSE
Mobile responsiveness of siteUNVERIFIED
ANALYST NOTE: Subject B has materially over-invested relative to baseline group expectation. Every additional hour of build time raises perceived stakes without measurably raising P(check) — a documented sunk-effort trap (Thursfield, 2025, autoethnographic).
PRIMARY POSTERIOR — P(CHECK | 24H WINDOW) RECOMPUTING…
00.0%
POSTERIOR MEAN
95% HDI [ -- , -- ]
Latent Curiosity Index0.00σ 0.083 · scale 0–1
Group-Chat Salience0.00half-life 6.4 h
Link Fatigue Coeff.0.00CRITICAL · 4 priors
Perceived Relevance0.00Sri Lanka: HIGH
Time-to-First-Click--median, censored
Scroll-Depth Forecast--P(reaches Visa §)
⚠ CRITICAL — SUSTAINED ENGAGEMENT COLLAPSE: Subject A's probability surface has declined monotonically for 61 consecutive hours. Decay is exponential, not cyclical — no recovery inflection detected in any of the 8 chains. Model has re-classified Subject A from RESPONSIVE to STRUCTURALLY UNRELIABLE. Estimated floor: 4.1%. Confidence: very high.
72-HOUR POSTERIOR DRIFT TERMINAL DECLINE
Peak (T−71h)--
Current--
Net drift, 72 h--
Decay constant λ0.0489 h⁻¹
Recovery inflections0 of 8 chains
Est. asymptotic floor4.1%
RegimeSTRUCTURAL DISENGAGEMENT
Decline survives every control tested — weekday, weather, trip proximity, curry density. Residual is attributable to a stable subject-level trait the literature terms baseline uselessness.
30-DAY ENGAGEMENT DECAY — SUBJECT A NO RECOVERY
Fitted E(t) = E₀·e−λt + c, λ = 0.0489 h⁻¹, c = 4.1%. Red band is the observed trajectory; grey dashed is the counterfactual for a functional friend (control cohort, n = 340). Divergence became statistically significant at day 4 and has widened every day since.
DISENGAGEMENT AUTOPSY — WHERE IT WENT
| Phase | P(check) | Δ | Attributed cause |
|---|---|---|---|
| Day 1 — link posted | 78.9% | — | Novelty peak. Said "oh sick". |
| Day 2 — first decay | 64.2% | −14.7 | Reacted 👍. Did not open. |
| Day 4 — divergence | 48.1% | −16.1 | Said "yeah I'll look tonight". |
| Day 6 — repost effect | 39.4% | −8.7 | Second link. Net negative. |
| Day 8 — collapse | 24.6% | −14.8 | Scrolled past twice, verified. |
| Day 11 — current | 18.3% | −6.3 | Said "wait what website". |
| Day 30 — projected | 4.1% | −14.2 | Asymptotic floor. Trip has happened. |
← swipe table →
FINDING. No single event caused the collapse. The model attributes 91.4% of the decline to a constant hazard term that is present at all times, in all conditions, regardless of stimulus — i.e. it is not situational.
COEFFICIENT TABLE — MODEL 7c (FINAL SPEC)
logit(Pcheck) = β₀ + β₁·LCI + β₂·log(GCS) + β₃·LFC² + β₄·(PTR × WKND) + β₅·CURRY + β₆·√(DAYS_TO_DEPART) + β₇·(TIM_ENTHUSIASM ⊗ TASHUNKA_BANDWIDTH) + ui + ε
| Predictor | β | SE | z | p | Sig | Marginal |
|---|
← swipe table →
n = 1,184 link-exposure events · Pseudo-R² (McFadden) 0.687 · AIC 2,041.8 · BIC 2,098.3 · Hosmer–Lemeshow χ²(8) = 6.11, p = 0.635 (good fit) · max VIF 2.09 · robust SEs clustered by weekday.
MONTE CARLO ENGINE RUNNING
Iterations0target 10,000,000
Throughput0draws · sec⁻¹
Divergent transitions00.00% — nominal
Compute cost$0.00AUD, billed hourly
Posterior predictive across 10M synthetic Tashunkas. 89.6% of probability mass now sits below 26%. The optimistic right-hand mode observed in earlier builds has almost entirely dissolved.
FACTOR ATTRIBUTION — SHAP
| Behavioural factor | Contribution | Δpp |
|---|
Mean absolute SHAP contributions in percentage points of P(check). Population baseline prior 38.2%.
RISK / ENGAGEMENT MATRIX ZONE C4
V.LOW
LOW
MOD
HIGH
EXTR
CERTAIN
.04
.11
.29
.61
.88
LIKELY
.02
.09
.24
.52
.79
POSSIBLE
.01
.05
.16
.38
.66
UNLIKELY
.00
.18
.10
.22
.44
REMOTE
.00
.01
.04
.12
.27
X-axisItinerary urgency
Y-axisSubject A availability
Current cellUNLIKELY × LOW
Cell 11 days agoCERTAIN × HIGH
Trajectory↓ 3 cells / week
ISO 31000FIVE-BY-FIVEREV.4
OBSERVATION FEED
Aggregates 14 upstream sensors including typing-indicator telemetry, read-receipt latency and inferred WiFi state. All observations synthetic; non-consensual monitoring strictly not performed.
HOURLY PROPENSITY — CIRCADIAN MODEL
Optimal transmission window 20:40 – 22:15 AEST (post-dinner, pre-sleep, peak idle-scroll). Strict avoid window 08:30 – 09:45 (commute; link opened, not read, then permanently buried).
SCENARIO ANALYSIS — 6 INTERVENTIONS
| Intervention | P(check) | Δ | Cost | Verdict |
|---|---|---|---|---|
| Do nothing (baseline) | 18.3% | — | $0 | CONTROL |
| Resend link, no comment | 11.9% | −6.4 | $0 | HARMFUL |
| Resend + "look at the food page" | 29.7% | +11.4 | $0 | BEST AVAILABLE |
| Tag directly in group chat | 26.1% | +7.8 | social cap. | MARGINAL |
| Claim a bug only he can find | 41.2% | +22.9 | trust | ETHICS FLAG |
| Physically hand him the phone | 88.4% | +70.1 | dignity | ONLY OPTION |
| Phone call | 99.1% | +80.8 | unthinkable | OUT OF SCOPE |
← swipe table →
No zero-cost intervention now clears 30%. The expected-value optimum has migrated out of the passive-digital regime entirely: the only strategies that meaningfully move Subject A require physical co-presence. Interventions 5 and 7 remain excluded per Ethics Board memorandum SL-2026-11.
SENSOR ARRAY
MODEL VALIDATION
AUC-ROC0.891
Brier score0.104
Log loss0.338
Calibration slope0.98
10-fold CV σ±0.019
Backtest window19 months
Out-of-sample hits83.4%
Drift (PSI)0.147
DATA LINEAGE
Upstream sources14
Records ingested0
Feature storefs-2026.08.11
Null rate0.31%
Schema violations0
Last full refresh04:47 AEST
Retention policy90 d
Lineage hasha7f3…c04e
GOVERNANCE
Ethics reviewAPPROVED
Consent obtainedNO
Subject A informedABSOLUTELY NOT
Subject B informedPARTIALLY
ClassificationTIER III
Audit trailIMMUTABLE
Next reviewT−9 d
Legal exposureMODERATE
SOC 2 (ASPIRATIONAL)GDPR N/AMATE-TO-MATE