Stuart Winter-Tear → simulation components

Track A — SWT Framework Alignment

How each of Stuart Winter-Tear’s principles maps to a concrete simulation component — and what we measure to know we’re following the framework.

01

The flow

1Calibrate
Value anchoring. Choose the business lever first, not the AI use case.
2Deploy
Bounded workflow. No reorg. AI enters existing structures.
3Observe
An LLM scans each sprint for operational patterns — bottlenecks, trust breakdowns, exception spikes, informal workarounds, coordination gaps — plus patterns it wasn’t pre-programmed to look for. Falls back to threshold checks if the LLM is unavailable.
4Learn
Discovery Engine + THEARI capture evidence. Update the playbook, governance, frameworks. LLM heuristics with a deterministic fallback.
5Redesign
LLM-proposed redesigns when available, else a hardcoded conservative seam-fix. One change, evidence-backed.
6Compound
The next deployment starts from accumulated knowledge.
02

Current run — strongest signal

Agent-surfaced
gaps
The clearest divergence: Track A’s rate declines to 0 as evidence-built redesigns complete the seams durably. Track B completes them too — a-priori, in waves — but its fixes revert under change fatigue. In the codifiable regime B ties A; in the tacit regime A still wins. Employee AI trust is no longer hand-set to diverge — as of the Round 5 de-rigging it emerges for both tracks from the same signal, and comes out close and earned (A ~6.8, B ~6.7). Honest caveat: only 2/11 agent profiles diverge behaviorally, attrition never triggers, and run-to-run variation now comes only from LLM sampling — the model’s randomness is fully seeded, so same-seed runs reproduce exactly.
Trace
system
Every run emits a self-contained trace.jsonl audit trail with 13 event types across all layers. The schema is documented in .claude/trace-schema.md for AI-assisted analysis.
03

THEARI value-thesis validation

Every discovery advances through five phases. Track A is evidence-gated — it must earn each advance with sprint-log evidence. Track B is timeline-gated — it auto-advances on the steering-committee schedule regardless of proof. Same five phases, different key that unlocks them. Track B now runs full discovery too — its missing-information discoveries come from agent-surfaced gaps, not a hardcoded detector.

1Theoretical
Hypothesis proposed by the Discovery Engine.
2Empirical
Evidence from the sprint log — Track A must earn it.
3Applicable
Works inside the bounded workflow.
4Replicable
Holds across deployment cycles.
5Impact
The business lever moved.
04

Capability state progression

What each track’s Discovery Engine is authorized to do — unlocking on validated patterns, successful redesigns, and the absence of incidents.

1Observe
Read the sprint log, detect patterns.
2Recommend
Propose actions and hypotheses.
3Bounded
Act within authorized scope.
4Autonomous
Redesign from validated evidence.
5Emergent
Shape the operating model itself.
05

SWT principles ↓ components11 mappings

Operating model
is discovered
TrackAController deploys AI into unchanged org structure. No target operating model, no future-state design upfront. Structure follows evidence — the operating model is extracted by the Discovery Engine from observed patterns, not drawn on a whiteboard. Source: “The AI Operating Model Is Discovered, Not Designed” (Jul 2026)
Bounded, complete
workflow
BoundedWorkflow: “single-vehicle auto claims under $10K, no injury, clear liability,” complete from FNOL through subrogation closure. ~30% of claim volume. Accountable owner with operating authority. Source: “The AI Operating Model Is Discovered, Not Designed” (Jul 2026)
Value anchoring
business_lever = "loss_ratio". The first deployment is chosen because it connects to the #1 P&C metric, not because “AI can do estimation.” Track B starts with “where can we use AI?”; Track A starts with “where does the business need better intelligence?” Source: “From AI Activity to Operating Leverage” (Jul 2026)
No reorg before
evidence
DeployWithoutReorganizing: no new roles, no reporting-line changes, no RACI redesign. AI gets access to the same systems the human role has; output enters at the same handoff point; existing roles interact with it using existing authority. Source: “The AI Operating Model Is Discovered, Not Designed,” “You Are Not Deploying Agents”
Capability + org
knowledge
LearningLoop captures both outputs. The Discovery Engine proposes LLM-driven actions (recommended_action, predicted_impact, counterfactual) that feed governance and decision frameworks. The Playbook accumulates patterns; the next deployment starts from evidence, not opinion. Source: “The AI Operating Model Is Discovered, Not Designed” (Jul 2026)
Translation
debt
Seam-risk scores on every workflow handoff; a translation-incident counter on every decision. Agent-surfaced gaps surface information that should exist but doesn’t — the agent notices the missing context, not a detector. Diana’s informal review is the canonical example, explicitly modeled. Source: “You Are Not Deploying Agents. You Are Redesigning Work” (Mar 2026)
The bottleneck
always moves
Bottleneck detection is LLM-driven in observe(): the Discovery Engine identifies which step is the constraint each sprint instead of a hardcoded threshold check. Track A redesigns the seam, not the AI; Track B treats bottlenecks as “teething issues.” Tracked via the bottleneck_migration discovery type. Source: “Speeding One Cog Breaks the Machine” (Mar 2026)
Permission and
coordination
Agent psychological profiles (OCEAN, biases, motivations, relationship graphs) drive behavior more than AI capability. The same AI under different organizational conditions produces different outcomes; Diana’s trust cascade propagates through influence networks. Source: “The Agentic Operating Model” (Mar 2026)
Stop and adapt
Stop conditions: no measurable business impact within 3 sprints → stop and redirect. Explicit criteria, not framed as failure — framed as evidence that this insertion point wasn’t the right one. Source: “The AI Operating Model Is Discovered, Not Designed” (Jul 2026)
Measured by
outcomes
Outcome metrics: cycle time, loss ratio, cost per claim, retention. Activity metrics (licenses, training completion) are tagged with SWT warnings and enabled only for Track B. Track A never measures adoption. Source: “From AI Activity to Operating Leverage” (Jul 2026)
Accountable owner
deploys learning
Kathryn as Forward Deployed Engineer: embedded in the work, not managing from a dashboard. Primary metric: learning velocity — patterns discovered, validated, and operationalized per sprint. Whether she’s in the work or above it is the difference between a discovery driver and a project manager with a different title. Source: Klatt, “Human-in-the-Lead Forward Deployment” (Jun 2026); Marampelly, “Rise of the FDE”
06

SWT’s specific concerns — how we measure them12

Translation
debt
Translation-debt index: handoff failures and reconciliation incidents. Handoff-failure rate: claims with at least one translation incident. Declining over time = seams are being redesigned.
Bottleneck
migration
Bottleneck detection each sprint: which step is the constraint? Tracked across sprints. At least 2 migrations expected in 16; stagnant = not enough deployment cycles.
Exception
inflation
exception_rate: % of decisions requiring human intervention beyond the designed process. An initial spike after deployment, then decline as seams are redesigned. A flatline at zero = agents aren’t producing realistic friction.
Interruption
readiness
interruption_readiness (0–1): can named individuals stop, narrow, or override AI within one business day? Designed after first deployment; autonomy widens only after interruption is designed.
Evidence
completeness
evidence_completeness: % of AI decisions with a full audit trail (what, why, authority, inputs, changes, escalation). “The receipt is the product.”
Agent authority
width
agent_authority_width: dollar threshold for autonomous AI decisions. Should widen with evidence — starts at $5K, widens to $15K+ as interruption is proven. Widening without evidence = deferred risk.
Learning
velocity
Learning-loop closure rate: discoveries become redesigns within 2 sprints. Pattern-reuse rate: later deployments informed by earlier evidence. Flat = learning isn’t compounding.
Absence
detection
AbsenceDetector: surfaces decisions made from incomplete information — stale data, missing pauses, unrecognized gaps. Type 2 emergence: the system didn’t know what it didn’t know.
Coordination tax
coordination_tax: count of handoffs + review layers + duplicate entry points + approval steps per workflow — steps that exist only because humans served as the interface between teams. Track A eliminates them after observation; Track B preserves them in the process diagram.
Decision latency
decision_latency: mean sprints from pattern detection to action taken. Track A → same-sprint (0–1); Track B → 1–4 sprints (steering-committee cadence). The operational metric behind the THEARI timeline divergence.
Elimination-first
check
Post-observation filter: “Does this step exist only because humans were the interface between teams?” If yes → eliminate, don’t redesign. Zero-based questioning grounded in deployment evidence, not upfront design. Track B doesn’t ask this.
Outcome
radiators
DiscoveryFeed: a continuous cross-team pattern broadcast. Empirical+ discoveries visible to all adjacent workflow owners. Mitigates Track A’s local-optimization risk without central governance; Track B broadcasts via monthly, filtered steering-committee presentations.
07

Six gaps we filled

SWT’s framework implies these capabilities but doesn’t specify them operationally. We made them explicit so the simulation can implement them. These need validation.

Gap 1 —
Interruption design
Implementation: after the first deployment, explicitly design both the veto dimension (who can pause / override / stop) and the responsibility dimension (who owns purpose, judgment, and learning). Sanjay can pause; Diana can pause individual decisions; Kathryn can stop the whole deployment. “Human-in-the-lead does not mean humans approve every AI decision. It means humans remain responsible for purpose, judgment, and learning.”
Gap 2 —
Translation capability
Implementation: after each redesign, score the workflow on SWT’s four translation capabilities — structure extraction, meaning portability, routing accuracy, throttling. Scored from observation data, not hardcoded.
Gap 3 —
Evidence & dockets
Implementation: every AI decision produces a structured audit trail as part of the workflow itself, tracked via the evidence_completeness metric. Not an after-the-fact compliance exercise.
Gap 4 —
Learning failure modes
Implementation: learning extraction can fail. Busyness (stress above threshold → 50% chance of missed capture). Politics (Diana suppresses discoveries that threaten her informal authority). Recognition (low AI literacy → patterns not recognized).
Gap 5 —
Stop & adapt logic
Implementation: explicit stop criteria evaluated every 2 sprints. No measurable business impact within 3 sprints → stop and redirect. Not failure — data. Move to the next opportunity.
Gap 6 —
Value anchoring
Implementation: Track A’s first decision is “which business lever needs improvement?” → then whether AI helps. Track B’s first decision is “where can we use AI?” → then connect to a business lever. This is the fork in sprint 1.
08

What Track A does NOT do

No steering
committee
Decisions made by an accountable owner with bounded authority, not by a monthly committee reviewing dashboards.
No future-state
design
No target operating model designed upfront. No new roles created before evidence justifies them.
No adoption
metrics
License activation, training completion, AI-assisted claim % — activity, not outcome. Track B measures them; Track A doesn’t.
No vendor
platform
No RFP process, no enterprise-platform selection. AI capability deploys into the existing tech stack within the bounded scope.