# ANDYGROVE.SYSTEMS — High Output Operations for Human + Agent Teams

**Lineage:** After High Output Management (Andrew S. Grove, Random House, 1983) · Little’s Law (1961) · Kingman (1961)

**Notice:** Independent platform; not affiliated with Intel Corporation or the Grove family

**Hero formula:** `Managerial Output = Output(org) + Output(neighboring orgs under influence) = ∑ (Activity × Leverage)   [Our Extension: Supervised Agents Multiply Leverage]`

## Operations Specialist R&D Track GRV-TRK-16: Equal-Token-Budget Single-Agent vs. Multi-Agent Architecture Diagnostics
- **Canonical URL:** https://andygrove.systems/token-budget-single-vs-multi-agent
- **Governing Law:** `I(Y ; M_{1..m}) ≤ I(Y ; X)   (DPI under equal thinking-token budget B)`
- **Remote Pods:** $20–$45/hr USD, 10–25 hrs/wk across 40 global IE/OR hubs

## When Half Your Workforce Runs on Tokens, Management Is Production.

Agents generate drafts fast, turning human verification into the new throughput bottleneck. Grove’s production framework makes hybrid operations measurable: schedule parallel work around the limiting step (critical-path anchor), catch defects at the cheapest entry stage, pair speed metrics with eval error counter-indicators, assign one human owner per agent, and set L0–L3 autonomy by measured Task-Relevant Maturity.

**Paired Leading Indicator Rule (Grove Ch. 2):** Pair every speed/volume metric (`tasks/hr`) with an eval error & post-gate escape counter-indicator so agents never optimize throughput by degrading quality.

**Tags:** Limiting Step (Critical Path) · Paired Leading Indicators · Task-Relevant Maturity (L0–L3) · Little’s Law: L = λW · Kingman’s Queue Law

## Verified Operational Metrics
- **4 Core Disciplines:** IE · Operations Research · BizOps/TPM · SPC Quality
- **5.8% Case Pass Rate:** Vetted on queueing, eval design & trace diagnostics
- **5 Business Days:** Median time from COO brief to matched specialist pod
- **7 Business Days:** Turnaround for a 1-Page Agent Line Audit

## Node 01 — For Operations Specialists
Part-time remote R&D pods ($20–45/hr, 10–25 hrs/wk). Apply Little’s Law (1961, `L = λW`), Kingman’s queue formula, critical-path limiting-step scheduling, paired leading indicators, OKRs (originated under Grove; Doerr, 2018) and L0–L3 TRM evals to multi-agent workflows. Who we look for: industrial engineers, operations researchers, BizOps and program managers, process and quality engineers.
- **Submit Credentials:** https://baserow.io/form/5dGFgfNPjhbG9qpSw8A3yGNK72vJQkGmkr__j_qmyn8?prefill_Source+Site=andygrove.systems&hide_Source+Site

## Node 02 — For COO Offices
Fast agent generation makes human verification the new throughput bottleneck (`Θ = min(λ, min_i N_i r_i / v_i)`). Engage vetted specialists who deliver a **1-Page Agent Line Audit** in 7 business days: map handoff throughput, catch defects at the cheapest entry stage, assign one human owner per agent, set L0–L3 autonomy by measured TRM, and build paired indicator dashboards before you scale.
- **Request a 1-Page Agent Line Audit:** mailto:output@andygrove.systems

## L0–L3 Task-Relevant Maturity (TRM) Autonomy Ladder
- **L0 — Shadow / Telemetry Only** (`Wilson LCB < 0.75`): Agent runs in shadow mode; zero external side effects; human executes manually while collecting eval traces. *Cadence:* Daily trace triage by Agent Owner
- **L1 — 100% Human Gate (Structured Supervision)** (`0.75 ≤ Wilson LCB < 0.90`): Agent prepares structured draft + deterministic check; 100% human sign-off before commit (must satisfy t_rev + p_err·t_fix < t_manual). *Cadence:* Twice-weekly trace 1-on-1
- **L2 — Statistical Lot Sampling (Variable Inspection)** (`0.90 ≤ Wilson LCB < 0.97`): Auto-commit routine reversible actions; route 10–15% stratified SPC lot sample + all out-of-distribution flags to human pod. *Cadence:* Weekly trace 1-on-1 + SPC chart
- **L3 — Autonomous Commit + Exception Pod** (`Wilson LCB ≥ 0.97 (Reversible)`): Direct auto-commit on reversible slices; hard programmatic gate remains mandatory on high-blast-radius irreversible actions. *Cadence:* Weekly exception trace review + re-index on model release

## Sample Deliverable: 1-Page Agent Line Audit (AUDIT-SPECIMEN-01 // COMMERCIAL RFP & POLICY UNDERWRITING LINE)
- **Baseline:** Arrival rate λ = 48 dossiers/hr · 3 LLM stages + 4 senior reviewers (service rate r_H = 3.2/hr each, capacity N_H·r_H = 12.8/hr)
- **Critical Path vs. Capacity Bottleneck:** Critical-Path Limiting Step (Grove): 34s Clause Retrieval & Cross-Check (starts at t = 0s; SQL pricing & compliance lookup offset-scheduled in parallel). Throughput Bottleneck: Senior Human Review (utilization u = 94% → Kingman queue wait 3.8 hrs).
- **00 · Entry Gate** (Deterministic Schema & Completeness Filter | Single Owner: *Intake Ops Lead*): Rejects 14% malformed packets at $0.0004 before any frontier call
- **01 · Parallel Prep** (Small LLM (c_S) + Exact SQL/LP Pricing Solver | Single Owner: *Deal Desk IE*): r_1 = 420/hr · Offset-scheduled alongside 34s limiting step
- **02 · Synthesis** (Frontier LLM (c_F) on 22% High-Entropy Clauses | Single Owner: *Underwriting Architect*): Prefix KV-cache hit η = 91% · Single-agent DPI-clean trace
- **03 · Human Pod** (4 Senior Reviewers (L2 Lot Sample + Exceptions) | Single Owner: *VP Underwriting Ops*): u dropped 94% → 71% · W_q cut 3.8 hrs → 9 min · C_verified −74%
- **Paired Indicators:**
  - Throughput: Θ = min(λ, N_H·r_H/v_H) = 48/hr ↔ Counter-Indicator: Post-Gate Escape Rate ≤ 0.8% (SPC 3σ alarm)
  - Speed: Median Cycle Time W = 11.4 min ↔ Counter-Indicator: Entry-Gate + Pre-Gate Eval Defect Rate (6.2%)
  - Autonomy: 78% L2/L3 Containment ↔ Counter-Indicator: Human Override / Negative-Leverage Rewrite Rate (< 4%)

## Production System Principles (Human + Agent Workflows)
### HIGH OUTPUT MANAGEMENT (CH. 1–2) // CRITICAL PATH VS. CAPACITY BOTTLENECK — Grove’s Limiting Step ('Boiling the Egg') vs. Human Verification Bottleneck
In Grove’s breakfast factory, the limiting step (the 3-minute boiled egg) is the longest critical-path operation around which parallel steps (toast, coffee) are offset-scheduled — distinct from Goldratt’s capacity bottleneck.

**Hybrid Agent Implication:** In hybrid pipelines, slow reasoning/retrieval calls anchor critical-path offset scheduling, while fast token generation turns human verification into the throughput bottleneck: Θ = min(λ, min_i N_i r_i / v_i).

### HIGH OUTPUT MANAGEMENT (CH. 2) // LOWEST-VALUE-STAGE FILTER & PAIRED INDICATORS — Catch Defects at Entry & Pair Every Speed Metric with an Eval Counter-Indicator
Grove’s golden rule of quality control: reject a bad raw egg at delivery, never after cooking breakfast. And pair every quantity indicator with a quality counter-indicator so teams cannot game volume.

**Hybrid Agent Implication:** Place deterministic schema and context eval filters at the entry of the pipeline before expensive frontier calls, and pair agent speed (tasks/hr) directly against eval error and post-gate escape rates.

### HIGH OUTPUT MANAGEMENT (CH. 3) + OUR EXTENSION // POSITIVE & NEGATIVE LEVERAGE — Agents Multiply Leverage; Rewriting Every Draft Is Negative Leverage
Agents inside an organization are already part of Output(org); counting them separately double-counts output. In Grove’s leverage law Output = ∑(Activity × Leverage), agents act by multiplying managerial leverage — while micromanagement ('meddling') exerts negative leverage.

**Hybrid Agent Implication:** A specialist who manually edits every agent response exerts negative leverage (t_edit > t_manual). Assign exactly one human Owner per agent to fix rubrics and gates rather than hand-patching outputs.

### HIGH OUTPUT MANAGEMENT (CH. 4) // ONE-ON-ONES BECOME TRACE REVIEWS — The Manager’s One-on-One Becomes a Scheduled Execution Trace Audit
For Grove, a regular 60-minute one-on-one is high-leverage teaching and mutual problem discovery — never a status ritual.

**Hybrid Agent Implication:** For an AI agent, the one-on-one becomes the Agent Owner’s weekly review of production execution traces (tool-call loops, abstention misses, prompt-drift clusters) feeding directly into eval set updates.

### HIGH OUTPUT MANAGEMENT (CH. 6 & CH. 9) + DOERR (2018) // L0–L3 TRM & OKRs — Autonomy Ladder (L0–L3) & OKRs Born in Grove’s Operations Cadence
Delegation follows Task-Relevant Maturity (Ch. 9), while Objectives and Key Results — created under Grove and codified by John Doerr in Measure What Matters — align execution without top-down micromanagement.

**Hybrid Agent Implication:** Out-of-sample evals map Wilson lower bounds directly to L0–L3 autonomy levels, while hybrid OKRs bind agent owners to verified business outcomes (C_verified, p95 lead time W, zero critical escapes).

### ONLY THE PARANOID SURVIVE (GROVE, 1996) // STRATEGIC INFLECTION POINTS — Every Frontier Model Release Is an Operational Inflection Point
A Strategic Inflection Point occurs when a 10× shift in fundamentals invalidates the existing operating structure before lagging financial statements show distress.

**Hybrid Agent Implication:** Each new model generation shifts the Jagged Frontier overnight: workflows stuck at L1 human review last quarter cross into L2/L3 auto-commit, requiring COO offices to re-run TRM evals and rebalance human pods.

## Canonical COO Operational & Cost-Optimization Equations (COO vs. CFO Lens)
> **COO vs. CFO Distinction:** A CFO cuts unit token prices or freezes human reviewer headcount — collapsing First-Pass Yield (`Y_FPY`), inflating rework variance, and pushing human utilization `u -> 95%` where Kingman queue wait explodes (`u/(1-u) = 19`). A COO minimizes **Cost per Verified Outcome (`C_verified`)** via entry-stage eval filters, step-level routing, DPI-clean topologies, and Halfin-Whitt exception staffing.

- **EQ-01 (Managerial Leverage Law (Grove, 1983, Ch. 3 + Our Hybrid Extension)):** `Output_mgr = Output(org) + Output(neighboring orgs) = ∑_k Activity_k × Leverage_k(agents, evals)` — Avoids double-counting supervised agents already inside Output(org): rubric/router/gate design multiplies positive leverage across thousands of runs, whereas hand-editing every draft is negative leverage.
- **EQ-02 (Entry-Filtered Cost per Verified Outcome vs. CFO Raw Token Price):** `C_verified = c_entry + (1 − p_rej0) · [ c_tok · T̄_attempt + p_H · w_H · t_H ] / Y_FPY + p_escape · c_escape` — Operationalizes Grove’s Ch. 2 lowest-value-stage rule: a cheap entry filter (c_entry ≪ c_tok) rejects malformed inputs before frontier calls, while naive model downgrades that drop Y_FPY inflate true C_verified.
- **EQ-03 (Little’s Law Across Human + Agent Networks (Little, 1961 — Not Grove)):** `L = λ · W_sys,   W_sys = ∑_{i ∈ H ∪ A} v_i · ( W_{q,i} + t_i ),   t_i = 1 / r_i,   v_i = λ_i / λ` — Uses service rate r_i (avoiding expected-return notation clashes across quantitative sister sites). Unthrottled agent drafts and rework loops (v_i > 1) inflate WIP L and cycle time W_sys without increasing finished flow.
- **EQ-04 (General Line Throughput & Kingman’s Queue Explosion (Kingman, 1961; Bartolucci & Vivo, 2026)):** `Θ_sys = min( λ, min_i (N_i · r_i / v_i) ),   W_q ≈ [ (c_a^2 + c_s^2) / 2 ] · [ u / (1 − u) ] · t` — Throughput equals min(λ, capacity), reducing to min r_i only at saturation. At u = 95% human reviewer load, u/(1−u) = 19 (4.75× the queue wait at u = 80%), and AI rework variance (c_s² ≫ 1) explodes delay.
- **EQ-05 (Direct Step-Level Routing vs. Confidence Cascades (Paul & Nandy, 2026; Chen et al., 2023)):** `C_route(k) = (1 − p_F(θ_k)) c_S T_k + p_F(θ_k) c_F T_k   vs.   C_casc(k) = c_S T_k + p_def(θ_k) c_F T_k` — A <5ms step classifier (AgentRouter) dispatches routine tool/parsing steps directly to c_S and hard synthesis directly to c_F without paying c_S twice (~72% cost cut at 97.3% frontier accuracy).
- **EQ-06 (Multi-Agent Topology Tax & Data Processing Inequality Law (Tran & Kiela, 2026)):** `Y → X → Z_1 → … → Z_m ⇒ I(Y; Z_m) ≤ I(Y; Z_1),   B = ∑_{a=1}^m T_{think,a} + ∑_{(a,b)∈E} T_{msg,ab}` — Along closed Markov handoffs without fresh external observations, inter-agent summaries cannot create information and burn thinking budget B; split agents only for noisy-context isolation or privilege boundaries.
- **EQ-07 (Wilson Lower-Bound TRM & L0–L3 Autonomy Ladder (Grove Ch. 9 + SPC)):** `TRM_{1−α} = [ p̂ + z²/(2n) − z √( p̂(1−p̂)/n + z²/(4n²) ) ] / ( 1 + z²/n )  →  { L0, L1, L2, L3 }` — Replaces broken-at-100% Wald intervals with Wilson lower bounds, mapping eval accuracy and sample size n directly onto L0 (Shadow), L1 (100% Gate), L2 (SPC Lot Sample), or L3 (Auto-Commit).
- **EQ-08 (Human Review Stabilization & Shared Attention Capacity (Bartolucci & Vivo; Wang et al., 2026)):** `t_rev + p_err · t_fix < t_manual,   α_AI > α_crit,   and   λ_esc + λ_clarify < N_H · r_H` — Agents speed up a workflow only if review + error-repair time beats manual execution, AI-eligible task share exceeds α_crit, and combined escalation + clarification arrivals stay strictly below human capacity N_H r_H.
- **EQ-09 (Halfin-Whitt Square-Root Staffing for Human Exception Pods (Halfin & Whitt, 1981)):** `N_H* = ⌈ R_esc + β √R_esc ⌉,   R_esc = λ (1 − γ_auto(TRM)) / r_H,   P(W_q > 0) = α(β)` — Sizes senior human reviewer headcount N_H* from offered exception load R_esc (in Erlangs, using human service rate r_H) and delay safety parameter β > 0 as L2/L3 containment γ_auto rises.
- **EQ-10 (Prompt Prefix KV-Cache Hit Rate & Paired Indicator Guardrail):** `C_turn(k) = T_prefix · [ η_hit c_cache + (1 − η_hit) c_in ] + T_{Δ,k} c_in + T_{out,k} c_out` — Locking static rubrics and tool schemas in an immutable prefix (c_cache ≈ 0.10 c_in) pushes η_hit ≥ 0.90, cutting multi-turn input cost 60–80% while paired eval error indicators guard against context truncation drift.

## COO Production Architecture Case Studies (Human + AI Agent Spend & Throughput Optimization)
### CASE-01 [CLAIMS & UNDERWRITING OPS // TOPOLOGY & ENTRY-GATE REDESIGN] — Collapsing a 5-Agent Chat Ring into an Entry-Filtered Step-Routed Pipeline
- **Why Naive CFO Cost-Cutting Failed:** Engineering ran 5 frontier reasoning agents debating every dossier ($1.42/case). Naive budget mandate to swap all 5 agents to a cheap mini model dropped First-Pass Yield from 61% to 44%, doubling senior underwriter rework.
- **COO Production Intervention:** Added cheap entry-stage completeness filter (Grove Ch. 2), applied DPI token-budget law (EQ-06, arXiv:2604.02460) and Step-Level Routing (EQ-05, arXiv:2609.22951): 1 schema-locked extractor (c_S), Python policy rules, frontier escalation on 19% tail clauses, and L2 lot sampling.
- **Verified Impact:** −78% cost / verified dossier ($1.42 → $0.31) | 61% → 93% First-Pass Yield (Y_FPY) | −64% human underwriter queue load | 4.2 hrs → 11 min p95 cycle time

### CASE-02 [TIER-2/3 TECHNICAL SUPPORT // KINGMAN 95% QUEUE EXPLOSION] — Eliminating the 96%-Utilization Queue Explosion & Variance Wedge in Human Review
- **Why Naive CFO Cost-Cutting Failed:** Fast AI draft replies boosted raw ticket touches/hr, so finance froze L3 specialist staffing. Escaped multi-system hallucinations drove human reviewer utilization to u = 96% (u/(1−u) = 24), spiking ticket reopens to 24% (arXiv:2605.27202).
- **COO Production Intervention:** Enforced Human Review Stabilization (EQ-08): placed deterministic sandbox execution gate ahead of L3 humans, capped WIP via CONWIP (L ≤ L_max), throttled low-value clarifications (arXiv:2610.41129), and staffed L3 pods via Halfin-Whitt (EQ-09).
- **Verified Impact:** 24.0% → 5.8% reopen / rework escape rate | −71% L3 human escalation queue wait (W_q) | −52% net cost per permanently resolved ticket | Reviewer utilization u stabilized: 96% → 78%

### CASE-03 [COMMERCIAL & LEGAL DEAL DESK // NEGATIVE LEVERAGE & CYBERNETIC PODS] — Ending Manual Draft Rewriting (Negative Leverage) with 1+AI 'Cybernetic Pods'
- **Why Naive CFO Cost-Cutting Failed:** 40 specialists manually rewrote ad-hoc LLM drafts on 80-page RFPs — classic negative leverage (t_rev + p_err·t_fix > t_manual) outside the Jagged Frontier (SSRN 4573321) — while uncached 90k-token prompts burned compute.
- **COO Production Intervention:** Assigned one named Agent Owner per RFP track in 1-specialist + AI 'Cybernetic Pods' (NBER w33641): exact SQL/LP solver owns pricing matrices, immutable prefix caching (EQ-10) serves compliance text, weekly trace 1-on-1s eliminate recurring prompt errors.
- **Verified Impact:** 6.5 days → 1.4 days enterprise RFP cycle time | −58% specialist hours per completed bid | 89% prompt prefix KV-cache hit rate (−67% token spend) | Zero pricing-matrix arithmetic escapes

### CASE-04 [SUPPLY CHAIN & REPLENISHMENT // HUMAN-LLM-OR COMPLEMENTARITY] — Replacing Pure Reasoning-Token Loops with an LLM + Exact OR Solver + Human Triad
- **Why Naive CFO Cost-Cutting Failed:** Autonomous reasoning agents spent 45k thinking tokens per SKU-week trying to optimize multi-echelon replenishment in natural language — generating high API bills and violating safety-stock constraints during promo shocks.
- **COO Production Intervention:** Deployed the Human-LLM-OR architecture (Baek, Fu, Ma & Peng, arXiv:2602.12631): lightweight LLM agents parse unstructured supplier delay notices & promo calendars into structured shock priors; exact stochastic OR solver computes order policies; human planners review only >2σ regime shifts.
- **Verified Impact:** −91% inference compute (math moved to OR solver) | −14.6% total holding + stockout cost vs. ERP baseline | −31.0% operational cost vs. pure LLM agent loop | 88% L3 SKU-week containment at Wilson TRM ≥ 0.97

## Curated Original Preprints Library for COO Offices & Professional Upskilling (2023–2026)
- **Fresh Preprints Hub:** https://andygrove.systems/preprints
- **Continuing Professional Development Syllabus (42 PDH / CPE / CEU / CPD Hrs):** https://andygrove.systems/education
- **Google Colab Interactive Lab (.ipynb):** https://andygrove.systems/notebooks/andygrove-coo-hybrid-operations-upskilling.ipynb

### PP-01 (arXiv:2605.27202, May 2026 · UCL / King's College London) — Queue & AI: When Faster Tasks Slow Down the Workflow
- **Authors:** Silvia Bartolucci, Pierpaolo Vivo (UCL & King's College London)
- **Primary Link:** https://arxiv.org/abs/2605.27202
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-01.pdf
- **Operational Pillar:** QUEUEING & VARIANCE WEDGE
- **Empirical Finding:** Proves how fast AI draft generation creates a 'spike-and-tail' service distribution ('variance wedge'): escaped errors return as heavy rework while congested human reviewers rationally lower scrutiny precisely when risk peaks.
- **COO Production Rule:** Never scale agent draft volume without verifying t_rev + p_err·t_fix < t_manual, α_AI > α_crit, and filtering defects before human reviewers.

### PP-02 (arXiv:2609.22951, Sep 2026 · ICML AgenticUQ) — AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows
- **Authors:** Rudrendu Kumar Paul, Sourav Nandy
- **Primary Link:** https://arxiv.org/abs/2609.22951
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-02.pdf
- **Operational Pillar:** STEP-LEVEL COST ROUTING
- **Empirical Finding:** Trains a <5ms 12M-parameter trajectory router across 50,000 steps (GAIA, SWE-bench, WebArena, ToolBench) and 4 model tiers, cutting inference cost by ~72% while retaining 97.3% of frontier-only performance (vs. 31% for single-turn RouteLLM).
- **COO Production Rule:** Route per trajectory step (parsing/tool-formatting → small tier; ambiguous synthesis → frontier tier), never one frontier model per whole workflow.

### PP-03 (arXiv:2604.02460, Apr 2026 · Stanford) — Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
- **Authors:** Dat Tran, Douwe Kiela (Stanford University)
- **Primary Link:** https://arxiv.org/abs/2604.02460
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-03.pdf
- **Operational Pillar:** TOKEN BUDGET & TOPOLOGY
- **Empirical Finding:** Uses the Data Processing Inequality (DPI) to prove that under equal thinking-token budgets, multi-agent message passing loses mutual information and trails single-agent reasoning unless context rot on noisy long inputs degrades single-agent focus.
- **COO Production Rule:** Audit every multi-agent architecture at equal token budget; eliminate decorative inter-agent chatter unless noisy-context compression or privilege isolation requires it.

### PP-04 (arXiv:2610.41129, Oct 2026) — When Should AI Agents Ask? Clarification Requests Under Limited Human Attention
- **Authors:** Hanzhao Wang, Shengjie Liu, Zhengchao Wang
- **Primary Link:** https://arxiv.org/abs/2610.41129
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-04.pdf
- **Operational Pillar:** HUMAN ATTENTION QUEUES
- **Empirical Finding:** Models parallel AI agents sharing one human supervisor with finite attention as a queueing dynamic program; introduces Pairwise Coordination Policy (PCP) so agents condition escalation on real-time human queue congestion.
- **COO Production Rule:** Make agent escalation thresholds queue-aware: when senior human WIP rises, throttle low-value clarifications and defer non-urgent subtasks.

### PP-05 (arXiv:2604.12262, Apr 2026) — CascadeDebate: Multi-Agent Deliberation for Cost-Aware LLM Cascades
- **Authors:** Raeyoung Chang, Dongwook Kwon, Jisoo Lee, Nikhil Verma
- **Primary Link:** https://arxiv.org/abs/2604.12262
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-05.pdf
- **Operational Pillar:** ESCALATION CASCADES
- **Empirical Finding:** Injects lightweight multi-agent deliberation strictly at uncertain escalation boundaries of a cost-aware LLM cascade, resolving borderline ambiguity on small models before paying frontier-tier or human escalation costs.
- **COO Production Rule:** Trigger consensus/verification ensembles only on borderline-confidence items, keeping clear routine traffic single-pass.

### PP-06 (arXiv:2602.12631, Feb 2026 · Columbia) — AI Agents for Inventory Control: Human-LLM-OR Complementarity
- **Authors:** Jackie Baek, Yaopeng Fu, Will Ma, Tianyi Peng
- **Primary Link:** https://arxiv.org/abs/2602.12631
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-06.pdf
- **Operational Pillar:** HUMAN + LLM + OR TRIAD
- **Empirical Finding:** Evaluates multi-period operational control on InventoryBench, proving that pairing LLMs (unstructured context translation) + exact Operations Research solvers (constrained stochastic math) + human oversight beats standalone LLMs or pure OR.
- **COO Production Rule:** Never pay reasoning tokens to approximate linear/stochastic optimization; let LLMs parameterize exact OR solvers under human governance.

### PP-07 (arXiv:2602.11865, Feb 2026 · Google DeepMind) — Intelligent AI Delegation
- **Authors:** Nenad Tomašev, Matija Franklin, Simon Osindero (Google DeepMind)
- **Primary Link:** https://arxiv.org/abs/2602.11865
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-07.pdf
- **Operational Pillar:** DELEGATION GOVERNANCE
- **Empirical Finding:** Establishes a formal operational framework for multi-hop human-to-agent and agent-to-agent delegation covering dynamic authority bounds, contract specification, verification checkpoints, and accountability chains.
- **COO Production Rule:** Assign one named human owner and pair every delegated agent permission with a machine-verifiable acceptance test and rollback switch.

### PP-08 (arXiv:2610.05532, Oct 2026) — DelegationBench: Measuring When AI Agents Should Ask Before Acting
- **Authors:** Shiva Pochampally
- **Primary Link:** https://arxiv.org/abs/2610.05532
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-08.pdf
- **Operational Pillar:** ABSTENTION & ESCALATION EVALS
- **Empirical Finding:** Tests 10 models across 156 matched-pair irreversible vs. reversible scenarios: surface prompt rephrasing swings act-now rates by up to 52.5 percentage points, and live tool-executing agents check in far less than static evaluators.
- **COO Production Rule:** Enforce hard programmatic L1 gates (not prompt politeness) before high-blast-radius actions (external emails, refunds, production DB writes).

### PP-09 (arXiv:2503.13657, Mar 2025 · UC Berkeley) — Why Do Multi-Agent LLM Systems Fail? (MAST Taxonomy)
- **Authors:** Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Matei Zaharia, Ion Stoica et al.
- **Primary Link:** https://arxiv.org/abs/2503.13657
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-09.pdf
- **Operational Pillar:** MULTI-AGENT FAILURE TAXONOMY
- **Empirical Finding:** Analyzes 1,600+ execution traces across 7 multi-agent frameworks, classifying 14 failure modes across specification flaws, inter-agent misalignment, and absent task verification.
- **COO Production Rule:** Institutionalize weekly trace 1-on-1s: most multi-agent failures are handoff contract and inspection-gate defects, not base-model IQ limits.

### PP-10 (NBER w33641, Mar 2025 · Harvard / Wharton) — The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise
- **Authors:** Fabrizio Dell’Acqua, Charles Ayoubi, Raffaella Sadun, Ethan Mollick, Karim R. Lakhani et al.
- **Primary Link:** https://www.nber.org/papers/w33641
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-10.pdf
- **Operational Pillar:** HYBRID TEAM ORG DESIGN
- **Empirical Finding:** Pre-registered field experiment with 776 P&G professionals: 1 human + AI matched the performance of 2-person cross-functional human teams while erasing R&D vs. commercial functional silos.
- **COO Production Rule:** Redesign bloated cross-functional handoff chains into compact '1 generalist-specialist + AI' pods, reserving multi-human review for final sign-off.

### PP-11 (arXiv:2408.03314, Aug 2024 · UC Berkeley / DeepMind) — Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
- **Authors:** Charlie Snell, Jaehoon Lee, Kelvin Xu, Aviral Kumar
- **Primary Link:** https://arxiv.org/abs/2408.03314
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-11.pdf
- **Operational Pillar:** TEST-TIME COMPUTE ALLOCATION
- **Empirical Finding:** Shows that allocating inference compute adaptively by prompt difficulty via process verifiers lets smaller models beat 14× larger models at substantially lower total FLOPs.
- **COO Production Rule:** Allocate extra verification tokens strictly to hard tail tasks rather than paying frontier parameter prices on 80% routine volume.

### PP-12 (arXiv:2407.01502, Jul 2024 · Princeton) — AI Agents That Matter
- **Authors:** Sayash Kapoor, Benedikt Stroebl, Zachary S. Siegel, Nitya Nadgir, Arvind Narayanan
- **Primary Link:** https://arxiv.org/abs/2407.01502
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-12.pdf
- **Operational Pillar:** PARETO COST-ACCURACY EVALS
- **Empirical Finding:** Demonstrates that SOTA agent scaffolds often burn 10×–50× higher API cost for negligible accuracy gains and fail on holdout robustness when evaluated without joint cost-accuracy Pareto curves.
- **COO Production Rule:** Mandate 2D Pareto evals (Verified Accuracy vs. USD/Task) before promoting any agent scaffold from R&D into production.

### PP-13 (arXiv:2406.18665, Jun 2024 · UC Berkeley LMSYS) — RouteLLM: Learning to Route LLMs with Preference Data
- **Authors:** Isaac Ong, Amjad Almahairi, Wei-Lin Chiang, Joseph E. Gonzalez, Ion Stoica et al.
- **Primary Link:** https://arxiv.org/abs/2406.18665
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-13.pdf
- **Operational Pillar:** PREFERENCE-TRAINED ROUTERS
- **Empirical Finding:** Trains lightweight query routers on human preference data to dynamically select between strong and weak LLMs, cutting single-turn frontier call costs by 50–85% without degrading response quality.
- **COO Production Rule:** Deploy calibrated preference routers in front of single-turn LLM gateways and step-aware routers (PP-02) for multi-step agent loops.

### PP-14 (arXiv:2405.06087, May 2024 · MIT CCI (Nature Hum. Behav.)) — When Combinations of Humans and AI Are Useful: A Systematic Review and Meta-Analysis
- **Authors:** Michelle Vaccaro, Abdullah Almaatouq, Thomas W. Malone (MIT)
- **Primary Link:** https://arxiv.org/abs/2405.06087
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-14.pdf
- **Operational Pillar:** HUMAN-AI COMPLEMENTARITY META-LAW
- **Empirical Finding:** Meta-analysis of 106 experimental studies (370 effect sizes): naive human+AI pairing underperforms the best solo actor on pure prediction/decision tasks, but achieves true synergy in modular creation + structured verification workflows.
- **COO Production Rule:** Do not ask humans to 'eyeball' calibrated probability outputs; partition creative synthesis, exact solver checks, and exception judgment cleanly.

### PP-15 (arXiv:2305.05176, May 2023 · Stanford (TMLR)) — FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance
- **Authors:** Lingjiao Chen, Matei Zaharia, James Zou (Stanford University)
- **Primary Link:** https://arxiv.org/abs/2305.05176
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-15.pdf
- **Operational Pillar:** LLM CASCADES & CACHING
- **Empirical Finding:** Introduces the foundational three-lever inference cost stack — prompt compaction, semantic/prefix approximation caching, and learned LLM cascades — matching GPT-4 accuracy with up to 98% cost reduction.
- **COO Production Rule:** Stack immutable prefix caching (EQ-10) ahead of learned confidence cascades so routine volume never hits frontier billing.

### PP-16 (SSRN 4573321, Sep 2023 · HBS / Wharton / MIT) — Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality
- **Authors:** Fabrizio Dell’Acqua, Edward McFowland III, Ethan R. Mollick, Karim R. Lakhani et al. (758 BCG Consultants)
- **Primary Link:** https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321
- **Same-Origin PDF Brief:** https://andygrove.systems/papers/pp-16.pdf
- **Operational Pillar:** JAGGED FRONTIER & CENTAUR SPLIT
- **Empirical Finding:** Inside the capability frontier, AI lifted output +12.2%, speed +25.1%, and quality +40%; on deceptive tasks just outside the frontier, AI users dropped −19 percentage points below unassisted humans due to unverified reliance.
- **COO Production Rule:** Re-map your workflow's Jagged Frontier at every model-release inflection point and enforce 'Centaur' hard task boundaries.

## Continuing Professional Development (42 PDH / CPE / CEU / CPD Hours) & Google Education Modules
> Open-access Continuing Professional Development (42 PDH / CPE / CEU / CPD hours) and Executive Upskilling Hub built directly on 16 fresh 2023–2026 arXiv, NBER & HBS preprints (https://andygrove.systems/preprints) and 6 quantitative self-study modules (https://andygrove.systems/education) for Chief Operating Officers (COOs), Industrial Engineers, Operations Researchers, BizOps & Program Managers, and Process & Quality Engineers.

### CPD-01 (GRV-CPD-401 · 7 PDH / CPE / CPD Hrs) — Queueing Networks, Little’s Law (L = λW), Kingman 95% Overload & AI Variance Wedges
- **Canonical Course URL:** https://andygrove.systems/course/cpd-01
- **Syllabus PDF Brief:** https://andygrove.systems/papers/cpd-01.pdf
- **Fresh Preprints Hub:** https://andygrove.systems/preprints
- **Core Equation:** `L = λW   ·   W_q ≈ [(c²_a + c²_s)/2] · [u/(1−u)] · t   ·   u = 0.95 ⇒ u/(1−u) = 19×`
- **Preprints Covered:** PP-01: Queue & AI — Variance Wedge (Bartolucci & Vivo, arXiv:2605.27202, 2026) | PP-04: Clarification Requests Under Limited Human Attention (Wang et al., arXiv:2610.41129, 2026)
- **Learning Outcome:** Quantify why speeding up Routine tasks without filtering bimodal AI service-time variance creates a Kingman queue explosion at human review stations, and compute congestion-dependent escalation thresholds.

### CPD-02 (GRV-CPD-402 · 7 PDH / CPE / CPD Hrs) — Step-Level LLM Routing, Frugal Cascades & Prefix KV-Caching for COO Unit Economics
- **Canonical Course URL:** https://andygrove.systems/course/cpd-02
- **Syllabus PDF Brief:** https://andygrove.systems/papers/cpd-02.pdf
- **Fresh Preprints Hub:** https://andygrove.systems/preprints
- **Core Equation:** `C_verified = [ ∑_s c(m_s)·E[T_s] + p_esc·w_H·t_H + p_err·c_fix ] / Y_FPY`
- **Preprints Covered:** PP-02: AgentRouter — Heterogeneous Step Routing (Paul & Nandy, arXiv:2609.22951, 2026) | PP-05: CascadeDebate (Chang et al., arXiv:2604.12262, 2026) | PP-13: RouteLLM (Ong et al., arXiv:2406.18665, 2024) | PP-15: FrugalGPT (Chen, Zaharia & Zou, arXiv:2305.05176)
- **Learning Outcome:** Replace naive CFO flat token-price cuts with trajectory step-level routing, immutable prefix KV-caching, and confidence-calibrated cascades that cut inference bills 50–85% without collapsing First-Pass Yield (Y_FPY).

### CPD-03 (GRV-CPD-403 · 7 PDH / CPE / CPD Hrs) — Data-Processing Inequality (DPI) Token Budgets & Multi-Agent Failure Mode Diagnostics
- **Canonical Course URL:** https://andygrove.systems/course/cpd-03
- **Syllabus PDF Brief:** https://andygrove.systems/papers/cpd-03.pdf
- **Fresh Preprints Hub:** https://andygrove.systems/preprints
- **Core Equation:** `B_eff(K) = B_total − K·T_coord − ∑_k ΔI_handoff,k   (DPI Handoff Loss)`
- **Preprints Covered:** PP-03: Equal Thinking Token Budgets & DPI (Tran & Kiela, arXiv:2604.02460, 2026) | PP-09: MAST — Why Do Multi-Agent LLM Systems Fail? (Cemri et al., arXiv:2503.13657, 2025) | PP-08: Scaling Laws for Agent Systems (Kim et al., arXiv:2512.08296, 2025)
- **Learning Outcome:** Audit 14 MAST multi-agent failure modes across 1,600+ traces, eliminate redundant sequential agent chatter under equal token budgets, and enforce typed schema handoffs at Stage 00.

### CPD-04 (GRV-CPD-404 · 7 PDH / CPE / CPD Hrs) — L0–L3 Task-Relevant Maturity (TRM), Wilson 95% Evals & Blast-Radius Delegation Gates
- **Canonical Course URL:** https://andygrove.systems/course/cpd-04
- **Syllabus PDF Brief:** https://andygrove.systems/papers/cpd-04.pdf
- **Fresh Preprints Hub:** https://andygrove.systems/preprints
- **Core Equation:** `TRM_0.95 = WilsonLCB(p̂_eval, n, z=1.96) → {L0 Shadow, L1 Gate, L2 SPC, L3 Auto}`
- **Preprints Covered:** PP-07: Intelligent AI Delegation (Tomašev, Franklin & Osindero, arXiv:2602.11865, 2026) | PP-08B / PP-12: DelegationBench (arXiv:2610.05532, 2026) & AI Agents That Matter (Kapoor et al., arXiv:2407.01502)
- **Learning Outcome:** Operationalize Grove’s Chapter 9 Task-Relevant Maturity via finite-sample Wilson 95% lower confidence bounds, SPC acceptance sampling, single human Owner accountability, and hard L1 irreversible action gates.

### CPD-05 (GRV-CPD-405 · 7 PDH / CPE / CPD Hrs) — Human-LLM-OR Complementarity, Test-Time Compute & Deterministic Solver Guards
- **Canonical Course URL:** https://andygrove.systems/course/cpd-05
- **Syllabus PDF Brief:** https://andygrove.systems/papers/cpd-05.pdf
- **Fresh Preprints Hub:** https://andygrove.systems/preprints
- **Core Equation:** `min_x c^T x  s.t. Ax ≤ b  (Exact OR Solver) + LLM Unstructured Context + Human Exception Pod`
- **Preprints Covered:** PP-06: AI Agents for Inventory Control — Human-LLM-OR (Baek, Fu, Ma & Peng, arXiv:2602.12631, 2026) | PP-11: Scaling Test-Time Compute Optimally (Snell et al., arXiv:2408.03314, 2024) | PP-14: When Combinations of Humans and AI Are Useful (Vaccaro, Almaatouq & Malone, MIT, arXiv:2405.06087)
- **Learning Outcome:** Partition unstructured regime translation (LLMs), exact constrained optimization (MIP/LP solvers), and strategic exception sign-off (humans) while allocating test-time verifier compute strictly to tail difficulty.

### CPD-06 (GRV-CPD-406 · 7 PDH / CPE / CPD Hrs) — Cybernetic Human + Agent Pods, Jagged Frontier Mapping & Inflection-Point Resets
- **Canonical Course URL:** https://andygrove.systems/course/cpd-06
- **Syllabus PDF Brief:** https://andygrove.systems/papers/cpd-06.pdf
- **Fresh Preprints Hub:** https://andygrove.systems/preprints
- **Core Equation:** `Output_org = ∑_k (Activity_k × Leverage_k)   ·   Paired Indicators: Speed ↔ Eval Escape Rate`
- **Preprints Covered:** PP-10: The Cybernetic Teammate — P&G Field RCT (Dell’Acqua, Mollick, Lakhani et al., NBER w33641, 2025) | PP-16: Navigating the Jagged Technological Frontier — 758 BCG RCT (HBS 24-013 / SSRN 4573321) | PP-10B: Collaborating with AI Agents (Ju & Aral, MIT Sloan, arXiv:2503.18238, 2025)
- **Learning Outcome:** Re-architect functional silos into AI-equipped Cybernetic Pods, map Jagged Frontier failure cliffs at every model-release Strategic Inflection Point (Grove, 1996), and run weekly Trace 1-on-1s.

## Open Part-Time Remote R&D Pods ($20–45/hr USD, 10–25 hrs/wk)
### POD-01 (GROVE-IE-QUEUE-01) — Industrial Engineer & Multi-Agent Queueing / Critical-Path Architect
- **Compensation:** $25–$45 / hr USD | **Commitment:** Part-Time Remote Pod (10–25 hrs/wk)
- **Law:** `Little (1961): L = λW,   Θ_sys = min(λ, min_i N_i r_i / v_i)`
- **Scope:** Model hybrid human + agent pipelines as stochastic networks with service rates r_i and visit ratios v_i. Schedule critical-path limiting steps, cap reviewer utilization u < 0.82, and prevent Kingman queue explosions.
- **Stack:** Industrial Engineering, Little’s Law (L = λW), Kingman Queueing, Critical-Path Scheduling, Python / SQL
- **Apply:** https://baserow.io/form/5dGFgfNPjhbG9qpSw8A3yGNK72vJQkGmkr__j_qmyn8?prefill_Source+Site=andygrove.systems&hide_Source+Site

### POD-02 (GROVE-TRM-EVAL-02) — Process & Quality Engineer — L0–L3 TRM Ladder & Entry-Stage Eval Gates
- **Compensation:** $20–$45 / hr USD | **Commitment:** Part-Time Remote Pod (10–25 hrs/wk)
- **Law:** `TRM_{1−α} = WilsonLowerBound(p̂_eval, n, z_{1−α}) → {L0, L1, L2, L3}`
- **Scope:** Build lowest-value-stage entry filters, statistical process control (SPC) lot-sampling plans, weekly trace-review rubrics, and Wilson-calibrated L0–L3 autonomy gates.
- **Stack:** Process & Quality Engineering, Entry-Stage Eval Filters, L0–L3 TRM Ladder, SPC Acceptance Sampling, Trace Review Audits
- **Apply:** https://baserow.io/form/5dGFgfNPjhbG9qpSw8A3yGNK72vJQkGmkr__j_qmyn8?prefill_Source+Site=andygrove.systems&hide_Source+Site

### POD-03 (GROVE-OR-HANDOFF-03) — Operations Researcher — Human + Agent Handoff & Solver Optimization
- **Compensation:** $25–$45 / hr USD | **Commitment:** Part-Time Remote Pod (10–25 hrs/wk)
- **Law:** `min_{x, π} ∑ c_i x_i + c_esc P(escape | π)   s.t.   u_i = λ_i / (N_i r_i) < 1`
- **Scope:** Formulate step-level model routing, Human-LLM-OR solver decomposition, and congestion-aware human handoff policies using Markov decision processes and integer optimization.
- **Stack:** Operations Research, Human-LLM-OR Solvers, Stochastic Networks, Step-Level Routing, Cost-Throughput Modeling
- **Apply:** https://baserow.io/form/5dGFgfNPjhbG9qpSw8A3yGNK72vJQkGmkr__j_qmyn8?prefill_Source+Site=andygrove.systems&hide_Source+Site

### POD-04 (GROVE-BIZOPS-OKR-04) — BizOps & Technical Program Manager — Paired Indicators, OKRs & Trace Cadence
- **Compensation:** $20–$45 / hr USD | **Commitment:** Part-Time Remote Pod (10–25 hrs/wk)
- **Law:** `Paired Indicators: Speed(tasks/hr) × Counter-Indicator(eval error & escape rate)`
- **Scope:** Partner with COO offices to ship 1-Page Agent Line Audits, paired leading-indicator dashboards, single-owner agent governance, weekly trace 1-on-1 cadences, and inflection-point OKR resets.
- **Stack:** BizOps & Program Management, Paired Leading Indicators, Hybrid Fleet OKRs, 1-Page Agent Line Audit, Inflection-Point Governance
- **Apply:** https://baserow.io/form/5dGFgfNPjhbG9qpSw8A3yGNK72vJQkGmkr__j_qmyn8?prefill_Source+Site=andygrove.systems&hide_Source+Site

## Sister Foundational Research Portals
- **[VAPNIK.AI](https://vapnik.ai/)** — *V. N. Vapnik & A. Ya. Chervonenkis (Statistical Learning Theory, 1971–1998)*: Structural Risk Minimization, VC-Dimension & Non-Asymptotic Eval Generalization.
- **[MARKOWITZ.SYSTEMS](https://markowitz.systems/)** — *Harry M. Markowitz (Portfolio Selection, 1952 · Cowles Monograph 16, 1959 · Nobel 1990)*: CFO Agentic Portfolio Theory, Multi-Vendor Error Covariance Σ & Audit CVaR.
- **[WALRASIAN.COM](https://walrasian.com/)** — *Léon Walras · Arrow–Debreu · Mas-Colell, Whinston & Green (General Equilibrium)*: Compute Endowments, Shadow Prices & Risk-Adjusted Agent Return Pruning.
- **[BELLMAN.SYSTEMS](https://bellman.systems/)** — *Richard E. Bellman (Dynamic Programming, Princeton / RAND, 1957)*: Sequential Decision Processes, POMDP Control & Multi-Step Agent Horizons.
- **[CASELLABERGER.COM](https://casellaberger.com/)** — *George Casella & Roger L. Berger (Statistical Inference, 2002)*: Exact Finite-Sample Hypothesis Testing, Neyman–Pearson Gates & Paired Evals.
- **[ANDYGROVE.SYSTEMS](https://andygrove.systems/)** — *After Andrew S. Grove (High Output Management, 1983) · Little (1961) · Kingman (1961)*: COO Hybrid Human + Agent Production Lines, Limiting Steps, Queueing & L0–L3 TRM.

## Canonical Questions & Answers
### What is the difference between Grove’s 'Limiting Step' and a throughput bottleneck?
In High Output Management (1983, Ch. 1–2), the limiting step in the breakfast factory (the 3-minute boiled egg) is the critical-path operation with the longest lead time around which parallel steps (toast, coffee) are offset-scheduled. A throughput bottleneck is the lowest-capacity station governing Θ = min(λ, min_i N_i r_i / v_i). Because agents draft rapidly (r_A ≫ r_H), human verification becomes the capacity bottleneck — and per Kingman’s formula, pushing human utilization u to 95% explodes queue wait by 19×.

### Why use Managerial Leverage instead of adding '+ Output(agents)' to Output(org)?
Agents operating inside an organization are already included in Output(org), so adding a separate Output(agents) term double-counts output. In Grove’s leverage framework (Ch. 3, Output = ∑ Activity × Leverage — extended here to hybrid fleets), well-governed agents multiply positive leverage across thousands of runs, whereas manually rewriting every agent draft creates negative leverage ('managerial meddling').

### How does a COO optimize AI agent spend differently from a CFO?
A CFO cuts unit token prices or freezes reviewer headcount — lowering First-Pass Yield (Y_FPY), creating a 'variance wedge' (arXiv:2605.27202), saturating human queues (u → 1), and inflating rework cost. A COO catches bad inputs at the cheapest entry stage (Grove Ch. 2), routes per trajectory step (arXiv:2609.22951), pairs every speed metric with an eval error counter-indicator, assigns one human owner per agent, and staffs exception pods via Halfin-Whitt.

### How do the L0–L3 Task-Relevant Maturity (TRM) levels and Trace 1-on-1s work?
An agent’s out-of-sample eval pass rate and sample size n determine its Wilson 95% lower bound (EQ-07), mapping directly to L0 (Shadow), L1 (100% Human Gate), L2 (10–15% SPC Lot Sampling), or L3 (Auto-Commit on reversible slices). Each agent has a single human Owner who conducts weekly '1-on-1s' by auditing execution traces — and re-indexes L0–L3 thresholds at every model-release inflection point (Grove, 1996).

### What does a COO receive in a 1-Page 'Agent Line Audit', and how fast?
Within 7 business days (matched to a vetted IE/OR/BizOps specialist pod in a median of 5 business days; 5.8% technical case pass rate), COO offices receive a 1-Page Agent Line Audit containing: (1) full handoff & visit-ratio map with single-owner assignments, (2) critical-path limiting step + human utilization u diagnostics, (3) paired speed vs. eval-error indicator spec, and (4) L0–L3 autonomy & entry-filter redesign.

---
Independent platform; not affiliated with Intel Corporation or the Grove family