Operational hypotheses derived from research; not independently validated.

not validatedEnterprise extrapolation is modeled, not measured. Ranges such as 80% outcomes, 4–6 week baselines, and maturity-ladder timelines are directional hypotheses from the 249-session historical precursor — they have not been validated at enterprise scale.
Applications / Future Work

Extending the measurement instrument to operational scenarios.

These are working hypotheses for how the measured cost and correctness patterns might apply at enterprise scale. They have not been independently validated and should be treated as directional.

249
Historical Experiment Sessions
Across 8 models, 3 architectures
10
Decision Rules
Rules 1–5 measured · Rules 6–10 modeled
69×
Measured Cost Gap
Identical output. Different cost.
$309.17
Total Research Cost
Every session recorded and inspectable

What is the Accelerator?

The Accelerator is an implementation guide and measurement tooling — open-source Python tooling that instruments your AI sessions to measure cost-per-outcome across models. It is not a SaaS product. It runs locally against your sessions.

The Problem

You've deployed AI. You don't have acceleration. Here's why.

01

You bought AI tools. Individual output went up. Team velocity stayed flat.

Every developer has Copilot. Every quarterly report shows more AI usage. Deployment frequency hasn't moved. 80% of AI initiatives never deliver measurable business value. Individual execution solves the wrong bottleneck.

02

You set token budgets. Costs went down. AI effectiveness went with it.

Finance introduced sensible controls. Reasoning models got restricted. Context windows got shortened. Conversation history got truncated. You asked your best architect to design a system after reading only a fraction of the documentation — and got what you paid for.

03

You adopted reasoning models. Quality improved. Costs became completely unpredictable.

One architecture redesign costs $0.015. The same task on another model costs $1.01. Your budget spreadsheet has no column for "model variance" — and without calibrated pass-rate data, every model routing decision is a bet.

04

Your board wants 3-year AI projections. Nobody can model what they can't measure.

Codebase growth compounds quadratically. The β·v·N² term dominates long-run cost — model the curve before architectural commitments.

The problem isn't your team. It's the invisible gap between AI investment and business outcome. That's what the Accelerator measures. That's what it closes.

What the Accelerator delivers

Not a cost tracker. A business acceleration engine — powered by Agentic Dynamics measurement and precision FinOps, backed by 249 historical experiment sessions.

Precision FinOps Measurement

Not cost-per-prompt. Cost-per-outcome. Baseline WOC ratio, escalation rate, model utilization profile — captured by a calibrated instrument that knows the difference between cheap and efficient. Every token, every route, every recovery event tracked and inspectable.

Systemic Efficiency

Not cheaper models. Smarter measurement. Track cost-per-outcome by task type, model, and team. Every session's cost is inspectable. Every routing decision is backed by calibrated pass-rate data.

The AI Maturity Ladder

Engage at any level. Individual worker tool selection. Team budgeting and cost controls. Governance and compliance. 3-year projections and CapEx-to-OpEx modeling. Full augmented workforce transformation (hypothetical — not tested at enterprise scale). Start where you are. Scale to where you need to be.

What This Means for Your Enterprise

Every enterprise AI concern maps to a measured dimension of the instrument. Here's the translation layer.

If you're worried about...Our measurement is...What you can do
Paying for models to re-learn your architecture every sessionCache Strategy + Snowball RuleSwitch to cache-read at $0.14/Mtok. Measure N² curve. If cache hit rate under 50%, don't write.
Models that break when instructions are incomplete or contradictoryGrit — Flail Rate + Recovery CostDon't deploy models with >15% flail rate. High-Grit models (DeepSeek 8%) self-correct. Low-Grit models (nano 14%) cannot be trusted unsupervised.
Budget overruns from models that narrate every decisionExplanation Tax εRoute routine work to models with low flail (DeepSeek 8.4%). Reserve high-flail models (Claude 11.4%) for auditable work only. Measured from 227 historical experiments.
Autonomous systems silently failing at scaleFirst-Pass Rule — WOC RatioTrack WOC by model, task type, and time of day. Escalate if below 0.85. Design cascades so low-Grit models never enter unsupervised pipelines.
3-year cost projections that keep surprising financeEPM Horizon + Snowball RuleModel the N² curve before committing. Hedge energy price exposure. A 2× EPM spike during a 5-year horizon doubles cumulative costs.
Every AI session billed at premium rates by defaultCache Strategy + Model SelectionRoute routine work to $0.016/session models. Only 11.5% of sessions need escalation. Cache-read at $0.14/Mtok instead of cache-write at $3.75/Mtok.
The bottom line — the organizations that win won't have the largest AI budgets. They'll have the most disciplined measurement — routing work to the cheapest capable model for every task, tracking cost-per-outcome, and never paying for more reasoning than the job requires.

How it works

From invisible spend to precision acceleration. Engineered methodology, not a black box.

Weeks 1–2

Baseline & Diagnose

Capture your current cost-to-serve per task type. Establish WOC ratio, escalation rate, model utilization profile. Identify the 20% of workflows driving 80% of AI spend. Every organization has a different bottleneck — we find yours with measurement, not assumptions.

Weeks 3–4

Tune & Enforce

Apply the 10 decision rules. Implement cascade routing, cache layering, and budget enforcement. Route 78.6% of tasks to the cheapest tier. Escalate only what needs reasoning. Track cost-per-outcome in real time, not at the end of the quarter.

Week 5+

Monitor & Compound

Dashboard cost-per-outcome, throughput, and budget runway in real time. Every session feeds back into your calibrated cost model — routing gets smarter, projections get tighter, and the instrument adapts to your actual workloads.

Directional
Cost Modeling
Per-task cost differentials measured across 249 historical experiment sessions. Enterprise extrapolations are modeled, not validated.
69×
Measured Cost Gap
Identical task. Cheapest model vs. most expensive. Evidence on every page.
4–6
Weeks to Baseline
From first session capture to calibrated cost model for your specific workloads
0.90
First-Pass Success Rate (WOC)
WOC = 1/(1+r) with r=0.115. 90% of sessions succeed first-pass.

Even when models use identical token volumes (~11K/session), per-token pricing diverges 25–75×.

The AI Maturity Ladder

The Accelerator engages at every level. Start where you are. Scale to where you need to be.

1

Individual Worker

Model selection, personal productivity, and tool consolidation. Equip every employee with the right AI at the right cost. Eliminate shadow AI spend and seat-license waste. The tooling profiles task types and recommends the cheapest model that gets the job done — measured (Rules 1–5), not guessed.

Model selectionPersonal ROITool consolidationSkills acceleration
2

Budgeting & Cost Controls

Token budgets, cascade routing, cost-per-outcome tracking across teams. Stop measuring cost per prompt — measure cost per completed task. Provider-specific caching, batch scheduling, and budget enforcement that cuts spend without cutting context. Chargeback models that connect cost to value.

Token budgetsCascade routingCost-per-outcomeChargeback
3

Governance

Policy enforcement, compliance guardrails, multi-team standards. Select models by Grit (integrity under degraded input). Prevent expensive reasoning models on trivial tasks. Audit trails that demonstrate compliance. Approval workflows for high-cost or high-risk operations. Track WOC ratio across teams and flag outliers before they become budget overruns.

Policy enforcementComplianceAudit trailsStandards
4

3-Year Projections & Planning

Cost modeling with N² codebase growth compounding, energy price inflation hedging, CapEx-to-OpEx transition. The tooling models the N² curve so your board sees it coming. Workforce readiness planning for augmented-to-autonomous transition.

N² growth modelingCapEx→OpExWorkforce planningScenario simulation
5

Augmented & Autonomous Workforce

Full AI-native transformation. Agentic CI/CD pipelines that auto-escalate through model tiers. Autonomous governance and evolution (AGE) — systems that self-monitor, self-correct, and self-optimize. The escalation cascade: 78.6% of tasks resolve at the cheapest tier. Only <5% ever reach a human. Continuous optimization that fine-tunes enterprise AI spend over time.

Agentic CI/CDAutonomous governanceEscalation cascadeContinuous optimization

Enterprise Projections

What Grit-routed model selection means in practice. Projections derived from our 227 historical worktrees (perturbation corpus) (Rules 1–5 measured) and WFM principles (Rules 6–10 modeled), applied to realistic enterprise team sizes.

Project your costs: Use the interactive calculator on the Operational Framework page to project costs for your team size and workload.

Autonomous Workloads

When AI works without human supervision — the rules, economics, and operational model

How Grit Controls Autonomous Workloads

The Grit framework directly governs unsupervised AI operations. In autonomous mode, you can't review every output. You need models that self-correct when given bad context — because bad context is inevitable in real systems (stale documentation, conflicting requirements, incomplete specifications).

Real-World Autonomous Patterns

Autonomous AI maps directly to established workforce management disciplines. The equations are the same. The resource under management is what's different.

Agentic Dynamics Concept
WFM Discipline
How It Maps
Budget Ceiling
Erlang-C staffing models
Throughput derived from budget, not headcount. Tmax = Budget / Cjob is the AI equivalent of Erlang's offered load → agents → service level. (derived from measured retry rate r=0.115, modeled on Erlang-C staffing)
Cascade Rule
Tiered support (L1/L2/L3)
Flash → Sonnet → Opus → Human is the same structure as L1 agent → L2 specialist → L3 engineer → supervisor. Grit gates determine which tier auto-resolves vs escalates. (modeled escalation architecture; cascade experiments not yet executed)
SLA Buffer
AHT × service level targets
Queue depth × average job time + retry buffer = completion window. Same formula as call center SLAs. (modeled from WFM queue theory; batch experiments not yet executed)
First-Pass Rule (WOC)
First-call resolution (FCR)
WOC = 1/(1+r) is the AI equivalent of FCR. Above 0.85 healthy. Below 0.70 critical. Track by task type, model, time of day — exactly how call centers track FCR.
Batch Discount
Back-office batch processing
Deferred work is cheaper. Real-time has a premium. The 50% batch discount mirrors the economics of night-shift back-office processing vs daytime real-time queues. (modeled; 50% discount from provider pricing tables)

Batch vs On-Demand Decision Table

Task TypeUrgencyComplexityModeRationale
Code review< 1 hourLow–MediumOn-demandBlocks developer workflow
Test generation< 4 hoursMediumBatchNon-blocking, can queue
PR description< 5 minLowOn-demandTrivial, no queuing benefit
Refactoring sweep< 72 hoursHighBatchLarge scope, schedule overnight
Documentation gen< 24 hoursLowBatchNon-urgent, perfect for idle windows
Incident triage< 2 minHighOn-demandTime-critical, needs fastest model
Dependency updates< 72 hoursLowBatchCan run weekly, huge cost savings
Security audit< 72 hoursMediumBatchQueue depth manageable, retry-safe

WOC Ratio — Rule 5 (First-Pass)

WOC = 1/(1+r) where r is retry rate. Higher WOC = better (more first-pass success). Measured: WOC = 0.90 from r=0.115.

Healthy

WOC > 0.85

Good first-pass rate. Scaling is safe.

Investigate

WOC 0.70–0.85

Retry rate rising. Tune thresholds.

Critical

WOC < 0.70

Too many retries. Pause autonomy until root cause found.

Calculated: WOC = 0.90 from retry rate r=0.115 across 249 historical experiment sessions.

Escalation Cascade

Visual flow from cheapest to most expensive resolution path.

T1Flash / DeepSeek v4 Pro×1.0 base cost
T2Sonnet / GPT-4o×8–12 cost
T3Opus / GPT-5.6×28.2 cost (price ratio)
HumanHuman Review$25–150/hr context-switch cost

Key insight: 88.5% of tasks resolve at T1. Only 11.5% escalate. Of those that escalate, most resolve at T2. The human sees < 5% of all tasks.

Workstream Capacity

How many jobs can you process per day at a given budget?

Throughput = Budget / (costper_job × (1 + retry_rate))
ScenarioBudget/DayCost/JobRetry RateJobs/Day
Light coding tasks$5.00$0.01511.5%299
Code review + fix$10.00$0.0811.5%112
Architecture generation$20.00$0.4211.5%43
Full-stack PRD → deploy$50.00$1.0111.5%44

The Measurement Tools

This Agentic Dynamics instrument isn't theory. It's calibrated through 249 historical sessions, 10 perturbation operators, and 7 recovery signals to measure what AI costs and why. Run it against your own sessions and fine-tune the machine to your workloads.

10 perturbation operators

4 manifold (alien vocab, framing shifts, causality reversal, forced abandonment) and 6 semantic (false premises, inverted constraints, removed constraints, phantom success, competing goals, contradiction injection).

7 recovery signals

Surface markers, behavioral convergence, qualifiers, tech term convergence, step length convergence, structural pattern convergence, trajectory distance. Every trajectory step classified: EXPLORATION, RECOVERY, STABLE, or UNKNOWN.

Session capture & fine-tuning

Run the tools against your organization's own AI tool sessions. Calibrate cost models to your specific workloads, models, and task types. Every session captured becomes training data for a continuously improving cost intelligence engine.

249
Historical Experiment Sessions
Every session recorded and inspectable
3
Architectures
GPT dense · DeepSeek MoE · Claude dense
8
Model Variants
Cross-architecture AST-level comparison
34
Configs
Open source · Apache 2.0 license

Why the Accelerator over DIY

Without the Accelerator

  • Cost measured per prompt, not per outcome
  • Model routing decisions made on price, not pass-rate data
  • Token budgets cut without understanding reasoning impact
  • 3-year projections built in spreadsheets, not simulations
  • Governance is reactive — budget overruns found after the quarter closes
  • Vendor benchmarks used as cost estimates for your specific workloads

With the Accelerator

  • Cost-per-outcome tracked by task type, model, and team
  • Calibrated pass-rate data drives every routing decision
  • Budget enforcement that preserves reasoning depth
  • N² growth modeling with scenario simulation built in
  • Calibrated pass-rate data drives every routing decision
  • Real-time dashboards flag WOC outliers before they become overruns
  • Your own sessions calibrate your own cost models

Open source. No vendor lock-in. 34 configs, 249 historical sessions. The instrument runs locally — no API keys required to verify. If you find a gap, extend it. The data is yours. The instrument is Apache 2.0 licensed. Run it against any model, any provider, any task.

Cost compounds. Are you measuring it?

Traditional engineering optimizes for CPU, memory, storage, and network. AI introduces a new resource that deserves the same discipline: reasoning. The instrument measures what that reasoning costs — by model, by task, by session — so you route work where it delivers results. No more guessing which model is right for the job.

Explore Operational Principles →

Open Questions and Honest Nulls

What the instrument has not yet measured is published as a named state, never a zero or a stale fallback.

untriggeredEscalation arm. All six escalate cells passed on the first attempt — a premium-cost escalation estimate is not measured.
unmeasuredLSP diagnostics. No current canonical LSP series — published as "not measured", never "0 errors".
underpoweredCalibration. The 0/3 → 2/3 arc is descriptive under n=3 — a result, not statistical clearance.
no canonical outputLab panels. Several analysis surfaces have no canonical current output — the absence is rendered explicitly.

Each null names the signal it needs. When the instrument measures it, the hypothesis moves from proposal to evidence — see The Question and The Evidence.