TC
← All Research
PRD: Power Law Engine (PLE)
PRDVerdict/VAKFeatured

PRD: Power Law Engine (PLE)

**PRD: Power Law Engine (PLE)**

2026-04-0814 min read2,538 words
PRD: Power Law Engine (PLE) File: docs/PRDs/PRD_Power_Law_Engine_PLE.md Owner: Architect + PLMS Lead Status: Draft v0

Trent Carter

1. Purpose & Concept

The Power Law Engine (PLE) is a cross-cutting analytics module that:

  • Pre-execution (Blueprint / PLMS):
  • • Models fat-tailed risk (time, tokens, cost) from historical runs.

    • Adjusts estimates & confidence intervals to reflect avalanche / outlier behavior.

  • During execution (PAS Runtime):
  • • Monitors cascading failures (“avalanches”) and outlier events.

    • Raises early warnings and suggests load shedding / controlled burns.

  • Post-execution:
  • • Updates lane-level and system-level tail statistics (exponents, thresholds).

    • Feeds back into Blueprint, pricing, and PLE-driven dashboards.

    2. Goals / Non-Goals Goals

    • Capture heavy-tailed behavior in PAS runs (cost, time, retries, cascades).

    • Provide tail-aware estimates to Blueprint / PLMS (90–99% CI, not just averages).

    • Detect avalanches (large cascades of tasks/actions) in real time.

    • Provide lane-level risk scores (tail-risk) consumable by:

    • PLMS Blueprint

    • PAS Resource Manager

    • HMI dashboards

    Non-Goals

    • PLE does not change core PAS orchestration logic directly.

    • PLE does not define product pricing by itself (it informs pricing).

    • PLE does not perform ML training; it ingests telemetry and outputs analytics.

    3. High-Level Architecture 3.1 Placement in Verdict System

    flowchart LR

      subgraph User/HMI

        HMI[Verdict HMI\nDashboards & Blueprint UI]

      end

      subgraph Planning

        BP[Blueprint / PLMS\nEstimation & Scoping]

      end

      subgraph Runtime

        PAS[ PAS Root & Agents\n(Dir/Mgr/Prog) ]

        RM[Resource Manager]

      end

      subgraph Analytics

        LOGS[(Action / Run Logs)]

        PLE[Power Law Engine\n(Analytics Service)]

      end

      PAS --> LOGS

      RM  --> LOGS

      LOGS --> PLE

      PLE --> BP

      PLE --> RM

      PLE --> HMI

    Type: Separate internal service (FastAPI / gRPC style), stateless, reading from logs/warehouse, writing back summaries/flags.

    4. Core Concepts & Data 4.1 Event & Avalanche Event = a root-cause–grouped unit of work, typically:

    • One Blueprint job or PAS project_run.

    • Can contain 1..N actions / job cards across agents.

    Avalanche = a large event where:

    • A single triggering failure / mis-estimate leads to:

    • Many retries

    • Rollbacks / re-plans

    • Test re-runs

    • Large token/time overrun

    Key metrics per event:

    • event_id

    • root_cause_id (if available)

    • lane (Code, Docs, Data, QA, etc.)

    • num_actions (job cards / tool calls / steps)

    • tokens_actual, tokens_expected, tokens_delta

    • time_actual, time_expected, time_delta

    • cost_actual, cost_expected, cost_delta

    • num_rollbacks, num_retries

    PLE uses distributions of these to detect power-law tails.

    5. Functional Requirements 5.1 Pre-Execution: Blueprint / PLMS Integration Inputs:

    • Historical event summaries (per lane, per project type).

    • Draft Blueprint estimate: (time_est, cost_est, token_est, lane_mix, risk_profile).

    Behavior:
  • Tail Characterization
  • • For each lane and project type, PLE maintains:

    • Approximate tail exponent (α) for cost/time/overrun.

    • Thresholds for “normal” vs “extreme” events per metric.

  • Tail-Aware Estimate Adjustment
  • • Given a new Blueprint:

    • Adjust cost / time estimates with tail-aware confidence bands (e.g. P50, P90, P99).

    • Produce a Tail Risk Score (0–100) for the project.

  • Output to Blueprint UI / PLMS
  • • Display:

    “Expected” vs “worst-case (P90 / P95)” ranges.

    • Lane contributions to tail risk (which lanes dominate the tail).

    • Optionally: Recommend rehearsal intensity or sandboxing level for high tail-risk jobs.

    5.2 During Execution: PAS Runtime Monitoring Inputs:

    • Streaming (or batched) action_logs from PAS and Resource Manager:

    • Action type, lane, timestamps.

    • Tokens, retries, rollbacks, errors.

    Behavior:
  • Event Aggregation
  • • Group actions into ongoing events (by project_run, run_id, or root_cause_id).

  • Real-time Avalanche Detection
  • • Monitor for:

    • num_actions > k × median for similar jobs.

    • tokens_delta, time_delta, num_retries, num_rollbacks exceeding lane-specific thresholds.

    • When crossing thresholds:

    • Emit “Avalanche Warning” to:

    • Resource Manager (for load shedding / pause / replan).

    • HMI (visual alert and context).

  • Lane Load & Sandpile Height
  • • Maintain a per-lane load metric (“pile height”):

    • queued complexity, active events, historical risk.

    • When load exceeds safe band:

    • Suggest slowing new intake to that lane.

    • Optionally mark lane as “critical” in scheduling.

    Outputs:

    • avalanche_alerts (event_id, level, lane, suggested action).

    • lane_load_state (per-lane load index, risk flag).

    • run_risk_summary (for current project).

    6. Non-Functional Requirements

    Low overhead:

    • PLE must not block PAS execution.

    • Operates on logs / async streams; outputs are advisory.

    Configurable sensitivity:

    • Tail thresholds & alert levels tunable per lane.

    Explainable:

    • HMI must show _why_ a job is high risk (e.g. “overrun 5× lane median for similar complexity”).

    7. User Experience (HMI)

    Two primary UI surfaces:

  • Blueprint / Estimation Panel
  • • Shows:

    • “Average vs Tail” bars for cost/time.

    • Tail Risk Score badge (e.g., Low / Medium / High).

    • Tooltips: short explanations like:

    “Historically, 5% of similar Code+QA projects cost ≥4× the median. We’ve widened your budget band accordingly.”

  • Runtime Risk Dashboard
  • • Per-project view:

    • Current event size vs typical.

    • Avalanche alerts (timeline).

    • Lane view:

    • Load gauge (pile height).

    • Tail-risk indicator (how often this lane produces big avalanches).

    8. Data Flow (Simplified)

    flowchart TD

      subgraph Logs

        A[Action Logs\n(tokens, time, errors,...)]

        E[Event Summaries\n(historical)]

      end

      subgraph PLE

        B[PLE Batch\nTail Modeling]

        C[PLE Runtime\nAvalanche Monitor]

      end

      subgraph Consumers

        PLMS[PLMS / Blueprint]

        RM[Resource Manager]

        HMI[HMI UI]

      end

      A --> C

      A --> B

      E --> B

      B --> PLMS

      B --> HMI

      C --> RM

      C --> HMI

    9. Three Additional Power-Law Features for Verdict

    These are extra PLE-driven features that exploit power-law dynamics beyond basic risk modeling:

    9.1 Customer Value & “Whale Radar” Observation: Customer revenue / usage typically follows a power law: a few whales contribute most revenue / usage, long tail contributes the rest. Feature:

    • PLE ingests:

    • Per-account usage (tokens, projects, active seats).

    • Revenue / plan tier.

    • Models the tail of account value distribution.

    • Outputs to:

    Growth / Success dashboards:

    • Identify whales early.

    • Warn when whales are doing high-tail-risk projects (combine PLE project risk × account value).

    Blueprint:

    • For high-value customers, suggest more conservative risk bands or extra rehearsal.

    9.2 Blueprint Component / Template “Hit Map” Observation: Use frequency of Blueprint components, templates, and code recipes is heavy-tailed: a small set will power most usage. Feature:

    • PLE tracks:

    • Which Blueprint components are used in which projects.

    • Downstream outcomes: overruns, bugs, SLO violations.

    • Identifies:

    Power components: the 20% components used in 80% of successful projects.

    Toxic components: components that disproportionately appear in tail failures.

    • Uses:

    • HMI marks power components with a “Golden Path” badge.

    • PLMS preferentially suggests power components in auto-generated plans.

    • PAS flags toxic components for refactor / replacement work (controlled burns).

    9.3 Incident Severity & Controlled Burn Planner Observation: Incident severity (downtime, bug impact, hotfix cost) is typically power-law: a few incidents dominate total pain. Feature:

    • PLE ingests:

    • Incident logs (severity, impacted users, time-to-resolve, cost).

    • Models:

    • Severity distribution tail (α).

    • Which lanes, components, and providers dominate catastrophic incidents.

    • Outputs to:

    Controlled Burn Planner:

    • Suggests specific maintenance / refactor projects with maximal tail-risk reduction per unit effort.

    Blueprint:

    • For critical components (high incident tail), suggests extra testing / redundancy budgets.

    HMI:

    • “Top 10 sources of catastrophic incidents” list.

    10. Open Questions & Next Steps
  • Event grouping semantics
  • • Do we define event_id strictly as project_run, or allow cross-project root causes (e.g., shared infra outage)?

  • Tail modeling approach
  • • Exact method (e.g., Hill estimator + goodness-of-fit checks) can be decided later; PLE PRD only requires:

    • Ability to identify heavy-tailed vs non-heavy-tailed metrics.

    • Approximate exponents & thresholds.

  • Alert routing
  • • Which alerts are auto-surfaced to the user vs only used internally by Resource Manager?

    Yeah, 100% this belongs in Tron-land too. Here’s a clean section you can just bolt onto the bottom of the PRD.

    11. Tron Integration (System-of-Systems Layer) Objective: Expose PLE’s tail-awareness and avalanche signals to Tron, so Tron can make global, cross-project decisions (intake, routing, safety modes, maintenance scheduling). 11.1 Tron’s Role w.r.t. PLE

    At the Tron level, PLE is treated as a global risk oracle:

    • Provides lane- and project-level tail risk.

    • Surfaces avalanche patterns and fuel load (tech-debt / incident tail).

    • Informs policy switches:

    • When to accept/reject / defer incoming work.

    • How aggressively to run rehearsal / sandboxing.

    • When to schedule controlled burns.

    High-level:

    flowchart LR

      subgraph Tron

        T[Tron\nGlobal Policy & Governance]

      end

      subgraph Estimation & Planning

        BP[Blueprint / PLMS]

      end

      subgraph Runtime

        PAS[ PAS Root & Agents ]

        RM[Resource Manager]

      end

      subgraph Analytics

        PLE[Power Law Engine]

      end

      PAS --> PLE

      RM  --> PLE

      BP  --> PLE

      PLE --> BP

      PLE --> RM

      PLE --> T

      T --> BP

      T --> RM

      T --> PAS

    11.2 Tron–PLE Interaction Points 1. Intake & Project Gating

    Input to Tron:

    • Project-level Tail Risk Score from PLE (Low/Med/High).

    • P50 / P90 / P99 cost & time bands.

    Tron policy:

    • For High tail risk + high-value customer:

    • Require rehearsal mode and/or manual approval.

    • For High tail risk + low-value / trial user:

    • Throttle size/scope or defer.

    • For Low tail risk:

    • Auto-approve, minimal guardrails.

    2. Global Lane Pressure & Intake Throttling

    Input to Tron:

    • Per-lane load index (“pile height”) and recent avalanche frequency.

    Tron policy:

    • If a lane’s load/risk exceeds threshold:

    • Reduce intake from Blueprint into that lane.

    • Prefer routing new work to alternative lanes / templates when possible.

    • If a lane is consistently low risk & underutilized:

    • Allow Tron to steer more experimental / high-variance work there.

    3. Controlled Burn Scheduling

    Input to Tron:

    • PLE’s incident tail analysis and fuel load metrics:

    • Components / repos / lanes driving the largest catastrophic incidents.

    Tron policy:

    • Create maintenance / refactor epics and feed them into Blueprint.

    • Reserve a fixed fraction of global capacity (e.g., 10–20%) for Tron-initiated controlled burns targeted at:

    • Top incident sources.

    • Lanes with worst tail exponents.

    11.3 Configuration Ownership

    • PLE implementation & metrics: owned by PLE/Analytics.

    • PLE thresholds, policies, and “what to do when…”:

    • Owned by Tron config:

    • Tail-risk level → required rehearsal level.

    • Lane load index → intake multiplier (e.g., 1.0, 0.5, 0.0).

    • Incident tail severity → maintenance budget rules.

    Tron remains the source of truth for:

    • Global safety posture.

    • How aggressive or conservative Verdict should be, given current tail risk and resource posture.

    11.4 Open Questions (Tron-specific)

    • Does Tron treat PLE as a hard gate (can block work), or as a strong advisory (humans can override)?

    • Where do we surface Tron+PLE decisions in HMI so users understand:

    • “This project was throttled because global tail-risk is elevated.”

    • “This maintenance epic was auto-created by Tron based on PLE incident data.”

    Yeah, that’s exactly the right place to nail it down: Tron owns what happens, PLE just screams into the void. Let’s bolt on a new subsection you can paste into the PRD under the Tron section.

    11.5 Alert vs Action Policies & Tron UI 11.5.1 Alert vs Action: Two Independent Settings Tron treats PLE signals with two separate policy axes, both user-configurable:
  • Alert Level (how noisy the UI is)
  • Dropdown in Settings → “Tron / PLE Notifications”:

    • Silent:

    • No popups.

    • Status bar only (icon + color).

    • Tron View still records everything.

    • Normal (default):

    • Status bar indicator.

    • Inline toast for High-severity events (e.g., PLE avalanche, lane overload).

    • Tron View error list entries.

    • Verbose:

    • All of the above plus extra warnings for Medium-level anomalies.

    • Suitable for debugging / power users.

  • Action Level (how aggressive Tron is)
  • Separate dropdown: “Tron Actions on PLE Events”:

    • Off

    • Tron never blocks or throttles based on PLE.

    • Only logs & visualizes (pure advisory).

    • Advisory Only

    • Tron recommends actions (slow lane, rehearse, split project) but does not enforce.

    • UI shows “Recommended by Tron” badges; user may apply or ignore.

    • Soft Gate

    • For high tail-risk projects or active avalanches:

    • Tron pauses execution at decision points and prompts the user:

    “Tron recommends pausing intake to Code lane. Proceed anyway?”

    • User can override on a per-case basis.

    • Hard Gate

    • Tron blocks risky operations unless explicitly overridden with a strong action (e.g., “Force Run Anyway”).

    • Intended for high-stakes / enterprise / compliance-heavy environments.

    Design note: Both settings are per-user (for UX noise) but can also be workspace-locked by policy in enterprise mode.

    11.5.2 Status Bar Indicators During project execution (in Verdict IDE / HMI main window):

    • A dedicated Tron/PLE status item appears in the status bar with:

    Icon state:

    • ✅ Green: No active Tron warnings.

    • 🟡 Yellow: Warnings (medium tail-risk or lane load).

    • 🔴 Red: Active Tron errors / hard gates.

    Badge count: number of active Tron events affecting current project.

    Behavior:

    • Clicking the Tron/PLE status item opens the Tron View, focused on:

    • The current project_run.

    • The most recent warning/error if any.

    11.5.3 Tron View (Global PLE / Tron Console) The Tron View is a dedicated panel / window that complements Sequence and Tree views:

    Entry points:

    • Status bar Tron icon.

    • Clicking any “Tron error” or “Tron warning” link from:

    • Run logs.

    • Notifications.

    • Blueprint / PLMS risk panel.

    Layout (conceptual):

    flowchart TD

      A[Tron Error Toast / Status Bar] --> B[Tron View\n(Global Control Panel)]

      B --> C[Project Tab\n(Current run-focused)]

      B --> D[Lanes Tab\nLane load & tail risk]

      B --> E[Incidents Tab\nHistorical PLE/Tron events]

    Project Tab (default when opened during a run):

    • Summary strip:

    • Tail Risk Score for the project.

    • Current Action Level (Off / Advisory / Soft / Hard).

    • Current Alert Level (Silent / Normal / Verbose).

    • Event timeline:

    • PLE signals (avalanche warnings, lane overloads, incident risk spikes).

    • Tron decisions (throttled intake, enforced rehearsal, blocked operation).

    • Details panel:

    • Clicking an event shows:

    • Raw PLE metrics (event size, Δtokens/time, lane load).

    • Tron’s decision (what it did / suggested).

    • If applicable, user override logged (“Force run at 12:32 by Trent”).

    Lanes Tab:

    • Table / dashboard:

    • Each lane with:

    • Load index (“pile height”).

    • Tail-risk rating (Low / Med / High).

    • Recent avalanche count.

    • Buttons or links to:

    • Open lane configuration.

    • Schedule maintenance (controlled burns) if PLE suggests.

    Incidents Tab:

    • Historical list of Tron+PLE events:

    • Sortable by severity, date, lane, project, customer.

    • Clicking any row jumps to the detailed incident view (same as Project Tab detail panel, but historical).

    11.5.4 Tron Error Click-Through Behavior Whenever Tron generates an error or hard gate:

    • UI shows a clickable message, e.g.:

    “Tron: Hard gate activated for Project — PLE detected extreme tail-risk in Code lane. [View Details]”

    • Clicking “View Details”:

    • Opens Tron View → Project Tab, focused on:

    • That project.

    • That specific Tron/PLE event highlighted.

    • User can:

    • See full context: metrics, distributions, why it triggered.

    • Change Action Level (e.g., temporarily drop from Hard Gate → Soft Gate).

    • Choose to override (where allowed by policy).

    Related Research