Kyle Luecke

technical note 04 · infrastructure analytics, survival modeling & field systems

Point-in-Time Sovereign Telemetry, Right-Censored Energization Survival, and Export-First 811 Triage

Eliminating lookahead contamination and complete-case survivorship bias in hyperscale power/civil schedule auditing (N = 31 North American builds, 211-day median physical divergence lead, 112.0-week Kaplan–Meier substation energization median), paired with CSV export-first One-Call locate triage in 811Dog.

Kyle W. Luecke 2026-11-03 10 min read Sources: AI Buildout Whitepaper (PDF) ↗ · 811Dog ↗

1. Two Structural Biases in Infrastructure Delivery Estimation

Public timelines for multi-hundred-megawatt compute campuses, semiconductor fabs, and regional fiber backbones suffer from two systematic estimation errors whenever analysts rely on corporate press releases or naive historical averages:

  1. Retrospective lookahead contamination: Regulatory dockets (air-quality permits, army corps wetlands filings, regional transmission organization interconnection queues, municipal building permits) are mutable databases. If an analyst queries a utility queue today and joins current status fields against a build announced eighteen months ago, post-announcement amendments leak into the feature vector. Auditing whether physical-layer records lead public schedule revisions requires strict point-in-time reconstruction: every permit, substation study, and civil inspection must be evaluated solely at timestamps $t \le T_{\text{eval}}$.
  2. Complete-case survivorship bias under right censoring: High-voltage substation energization queues contain both completed interconnections and active builds still stalled in facilities studies or transformer procurement. Computing the sample median over completed interconnections alone systematically discards the slowest, highest-friction builds still in flight, biasing timeline estimates downward.

2. Model-Free Divergence Lead Time and the Three-Condition Physical Gate

In the Driftbound Labs North American cohort study (The Physical Layer of the AI Infrastructure Buildout; frozen manifest SHA-256 a2202685... covering 31 evaluable builds totaling 2,184 MW of announced capacity across hyperscale data centers, AI clusters, and semiconductor facilities), each project was tracked across sovereign primary records: RTO/ISO interconnection dockets, state environmental air/water permits, county structural/electrical permits, and civil mobilization records.

Before fitting any statistical model, define the purely observational divergence lead time for any project $i$ that eventually disclosed a material schedule slip:

Lead_i = T_public,i − T_physical,i

where $T_{\text{physical},i}$ is the timestamp at which a primary sovereign filing recorded a binding blocker (such as an air-permit suspension, an unexecuted facilities-study restudy trigger, or a foundation/subfab structural hold) and $T_{\text{public},i}$ is the earliest subsequent earnings call, SEC filing, or press update acknowledging the delay.

Across all 9/9 resolved material delay cases in the cohort, primary physical/regulatory divergence preceded public guidance revisions, yielding a median lead time of 211 calendar days (interquartile range 200–225 days; full range 105–339 days — including a 339-day lead on a TCEQ air-permit docket suspension, a 324-day lead on a subfab pier permit hold, and a 211-day lead on an Oncor switchyard interconnection docket).

Utility Interconnection TierBuildsObserved >90 d Slip RateMedian Schedule Slip
Tier 1 · Cleared / Existing Substation140 / 14 (0.0%)0 d
Tier 2 · Active Queue / Facilities Study83 / 8 (37.5%)60 d
Tier 3 · Major Network / Substation Upgrade55 / 5 (100.0%)195 d
Tier 4 · Stalled / Moratorium / Restudy44 / 4 (100.0%)408 d

Applying a conjunctive three-condition physical readiness gate — Utility Pathway Cleared ∧ Structural Permits Issued ∧ Civil Site Mobilization Verified — partitions the 2,184 MW announced headline capacity into 1,128 MW (51.6%) physically cleared on baseline schedule, 680 MW gated on high-voltage power delivery, and 376 MW gated on civil permitting or site stalls.

3. Right-Censored Kaplan–Meier Energization Survival and Holdout Calibration

To estimate high-voltage substation energization duration without dropping active multi-year queues, let $t_1 < t_2 < \dots < t_k$ denote the distinct elapsed weeks at which observed energizations occur, $d_j$ the number of substations energized at $t_j$, and $n_j$ the risk set (all tracked substation builds neither energized nor censored prior to $t_j$, retaining active queues censored after $t_j$). The standard product-limit survival estimator is:

Ŝ(t) = ∏_{t_j ≤ t} (1 − d_j / n_j), t_median = inf { t : Ŝ(t) ≤ 0.50 }

Across the 18 substation interconnection transitions in the cohort (12 completed energizations and 6 right-censored active builds still awaiting energization at freeze):

  • Naive complete-case median (dropping the 6 right-censored builds): 100.0 weeks (~23.0 months).
  • Kaplan–Meier product-limit step median (retaining all 18 at-risk transitions): 112.0 weeks (~25.8 months; linearly interpolated median 106.8 weeks; IQR 82.0–142.0 weeks).

Discarding right-censored active queues understates median energization time by three months precisely when queue congestion is worsening. On the pre-specified held-out test split (N = 8 builds), scoring physical-gate probability forecasts against realized status achieved a Brier score of 0.0383 versus 0.2476 for the public-guidance baseline (an 84.5% error reduction, improving 8/8 held-out assets, surviving 8/8 leave-one-program-out refits and 4/4 label-permutation / timestamp-shift placebos). Equally important is the explicit evidentiary boundary stated in the paper: because retrospective point-in-time freezes can still harbor subtle universe-selection effects, zero prospective post-freeze resolutions were claimed at initial publication; live prospective vintages are tracked separately.

4. Field-Scale Ingestion Without Portal Exhaustion: 811Dog

At regional telecommunications construction scale, One-Call (811) utility locate portals expose a related engineering tension: tens of thousands of daily locate tickets cover residential landscaping, fence posts, water service repairs, and sewer laterals, while an outside-plant fiber team cares about directional boring, conduit placement, vault installations, and competing fiber overbuilds along specific corridors.

Fetching full HTML/PDF ticket detail envelopes for every county-wide locate request is slow, brittle, and abusive to public portal infrastructure. 811Dog resolves this with a four-stage local pipeline verified by 335 offline pytest tests:

def _worth_detail_fetch(summary_row: dict, config: ScoringConfig) -> bool:
    # Stage 1: CSV export-first pre-filter before any detail/PDF HTTP fetch.
    work_for = norm_token(summary_row.get("WorkDoneFor", ""))
    excavator = norm_token(summary_row.get("ExcavatorName", ""))
    remarks = norm_token(summary_row.get("Remarks", ""))

    if matches_telecom_owner(work_for, config.telecom_owners):
        return True
    if matches_utility_keyword(remarks, config.infrastructure_terms):
        return True
    # Conjunctive contractor gate: known utility contractors also dig water/gas;
    # require either telecom owner context or non-excluded excavation remarks.
    if matches_contractor(excavator, config.tracked_contractors):
        return not matches_excluded_utility(remarks, config.non_telecom_exclusions)
    return False
  1. Export-first CSV triage: Bulk tabular summaries are downloaded in a single request and filtered in memory via _worth_detail_fetch, shrinking detail/PDF fetches by roughly two orders of magnitude before touching individual ticket endpoints.
  2. Conjunctive two-sided contractor scoring: Multi-utility underground contractors routinely bore both fiber conduit and municipal water/gas mains. Scoring contractor identity alone floods operators with water-main false positives; scoring explicit "FIBER" strings alone misses terse subcontractor tickets labeled simply "BORE 2 INCH CONDUIT". Requiring tracked_contractor ∧ compatible_work_type while emitting an explicit JSON array of relevance_reasons keeps precision high and makes every alert auditable.
  3. Append-only SQLite history and offline rescoring (tn811 rescore): Raw normalized ticket states are stored locally in SQLite so threshold or rule adjustments replay deterministically against historical snapshots without re-scraping the upstream portal.
  4. Deterministic PII redaction boundary (data_boundary.py): Raw locate tickets contain field-caller names, direct cell numbers, and dispatch emails. Before generating static JSON payloads for dashboard viewing, data_boundary.py strips contact PII at the serialization boundary so static views expose corridor, street, utility-type, and scoring provenance without leaking personal contact records.