Presentation Notes: Governed Workflow Systems

Architects of Intent

From one limited change to a composable factory of verified workflows.

Notation, not proof

A Model of Bounded Change

The notation names a proposed transition. The workflow around that transition determines what was allowed, what evidence counts, and whether any effect becomes real.

Observed state: Zn

The pinned base, current Terrain, and adopted intent used for this run.

Bounded process: P

One finite workflow proposes, checks, refines, or stops under fixed authority.

Terminal result

A supported proposal, no_change, blocked, or failed. The Run Record is sealed.

Separate authority

Admission may apply a candidate. Adoption may change future intent or policy.

Feedback adapts the next proposal. It does not widen the active authority, lower the evidence requirement, or admit or adopt its own result.

Mission Object & The Loop

A bounded refinement loop: generation produces candidates, validation produces findings, and a supported proposal remains separate from Admission.

This panel shows the Bounded Refinement Loop: probabilistic generation proposes, validation emits findings, and routing either permits another attempt or reaches a terminal result. A supported proposal remains separate from Admission.

Deterministic Sandwich

The Deterministic Sandwich: Prep/Mission Object → Model/GenAI → Validation/Gate
Model / GenAI

One probabilistic step works within named limits. Its output remains a candidate; Validation supplies evidence about the Mission's properties.

Operational Notes

The model explores candidates from a bounded Context Packet; no side effects happen at this layer.

  • Draft outputs are proposals, not accepted state.
  • Passing checks can support proposal closure; separate authority still controls Admission or Adoption.
In this minimal Sandwich, only the middle is stochastic. Larger workflows may use probabilistic evaluation inside enforcement and authority boundaries the run cannot change.

Substrate - Work becomes enforceable when intent, observed reality, and retained evidence are explicit.

Map

Intent surface

A structured, versioned representation of intent or observed structure, authoritative only for declared properties and scope.

Terrain

Operational reality

Reality-as-code: source, configs, schemas, and the behavior your systems actually execute.

Ledger

Reconstructable evidence

Append-only intent, input fingerprints, effects, structured results, transitions, and provenance. Under protected-sink and trust-root assumptions, missing or contradictory required evidence fails closed.

Semantic Density, Correctness, and Relevance

Three distinct properties of mission-ready context. Optimize all three; none can compensate for a missing one.

Semantic Density

Distinct, decision-bearing meaning per unit of context.

Correctness

Evidence-bound claims, current enough for the decision, with contradictions exposed.

Relevance

Material tied to the active Mission, property, and decision.

Dense falsehood misleads. Correct irrelevance distracts. Relevant repetition wastes attention.

authority | provenance | freshness | budget

Rule of Seven

Branching-Factor Heuristic
Inspect layers outside roughly 7 (±2) siblings. The band is a mnemonic, not a target or gate; observed failures decide whether to regroup.
Observed siblings
7
1 5 7 9 12
Low fan-out: inspect High fan-out: inspect
Illustrative Topology
Heuristic band

Example project tree

Inference Cost in Recurring Workflows

Per-call inference cost may fall while total workflow demand rises. Once inference becomes a recurring dependency, its cost, latency, and availability belong in the operating model.

Recurring inference can become infrastructure. Infrastructure exposure should be measured and bounded.

When inference becomes structural
  • Pricing volatility → operational risk
  • Provider lock-in → systemic dependency
  • Latency → architectural friction
The relevant unit is cost per admitted outcome, not price per model call.
Use the illustrative Hedge Builder to explore how different operating assumptions change cost and dependency risk.
Hedge Builder
Compute exposure under selected assumptions
Autonomy 60%
Expected cost / outcome
Index (lower is better)
P95 cost / outcome
Tail premium: + ( vs spot)
Cost certainty
%
Inverse of pricing volatility
Risk profile
Pricing volatility /100
Provider lock-in /100
Latency friction /100
Frontier exposure /100
Recommendation

The Limit Case: Acceleration Without Bounds

When candidate volume rises without enforceable boundaries, small errors can compound into systemic risk. Governed loops bound effects, require evidence, and stop when declared acceptance cannot be established.

Candidate generation rate
75/100
Unguarded AI
Entropy / risk index
Governed loops
Entropy / risk index

Illustrative model, not a measured risk curve. The dashed marker indicates the high-pressure region.

System Growth Raises the Cost of Broad Loops

The project tree grows across many changes. A dependable workflow does not revisit or rewrite the whole tree on every run; it operates on a declared slice with explicit effects.

Iteration cycle
0/12
Folders
Files
Touches / cycle
This cumulative view shows system growth. Each actual workflow should touch only the surfaces its Mission authorizes.
Project filesystem (cycle 0)
Seed

Side view: the system grows across authorized changes
Seed
A few files. One executable path. Minimal structure.

The Judgment Bottleneck

When candidate production gets cheaper, value can move toward specifying intent, defining evidence, governing admission, and maintaining the workflow system.

The shift is empirical, not automatic: teams must measure whether the added controls reduce total cost and risk for comparable work.

Anatomy of Delegated Autonomy

Routine work can proceed with less interruption when the activated Mission keeps authority, effects, checks, and stopping conditions unchanged.

1. Adopted Intent
Names authority, effects, checks, and budgets
2. Finite Workflows
Converge toward acceptance or stop with evidence
3. Separate Authority
Admits candidates and adopts any rule changes separately
Govern change

Shaping the Optimization Terrain

This is the geometry view of slicing. You are not dumping the whole repository into the window. You reshape the terrain so valid paths are easier and drift is harder.

Intent defines a target region. Slices, contracts, and Validators bias candidate production toward relevant states and expose selected forms of drift. They do not guarantee the shortest path, a global optimum, or a complete specification.

Intent preset: Central Optima
Illustrative 3D terrain: peaks represent outcomes that satisfy the selected properties; unmodeled properties may still matter.

Authority and Enforcement for One Property

Authority is declared for the property being decided. The control plane protects the comparison and transition; it does not become a universal source of truth.

1 / Declare
Map
Desired condition, authoritative source, owner, and transition policy.
2 / Observe
Terrain Sensor
Pinned observed state with provenance. No authority is inferred from the observation.
3 / Compare
Validator
Structured finding: pass, fail, or inconclusive, with evidence references and protocol identity.
4 / Decide
Route and Seal
Reach a terminal result under the Mission's routing; any supported proposal still needs separate authority.
Ordinary Mission
May change only the named Terrain surface. An attempted change to a protected control is denied and recorded before the Run Record is sealed.
Governance Mission
May emit a supported proposal. Separate authority adopts it; Adoption either atomically applies and selects it or selects an exact copy that was staged non-authoritatively.
Ledger: under protected-sink and trust-root assumptions, reconstructs authority, observations, effects, findings, and decisions; missing or contradictory required evidence fails closed.