Threshold-Governed Automation Doctrine

0. Purpose, Scope, and What This Is Not

Purpose

Every automated system that touches money, customers, or irreversible state needs an answer to one question: under what declared conditions does the machine act alone, and under what conditions does a human enter? Systems that answer it implicitly drift toward one of two failure poles - full-autonomy creep (the machine quietly accumulates authority nobody granted) or approval bottleneck (every action queues behind a human and the automation's economics collapse).

Threshold-Governed Automation (TGA) is the named pattern that answers the question explicitly: automate beneath a declared threshold, govern the exceptions through a defined authority ladder, verify through an independent observer, and publish the policy. This doctrine specifies the pattern's contract surface, anchors it in its canonical production instance, positions it within the executor-governance stack, and defines the proof-asset extraction rules that make it sellable.

Scope

• The pattern definition and its four mandatory contract elements (§1, §3).
• The canonical production instance: a national moving company's claims routing (§2), stated from that company's canonical claims-department context, not from operator recall.
• The pattern's position relative to human-in-the-loop handoff contracts, runtime dispatch governance, and the Darwinian Six-Rule Observation Gate (§4).
• The generalization registry (§5).
• The direct-response layer: TGA as the trust mechanism that converts AI anxiety into a governance story (§6), and the proof-asset rules (§7).

What This Is Not

• Not a system prompt or implementation spec. The conversational craft of the automated slice (the nine-stage conversational framework, the negotiation ladder) lives in the instance's own canon.
• Not a substitute for human-in-the-loop handoff contracts. TGA decides when a human enters; the handoff contract governs what the human node is once entered. They are complementary layers (§4).
• Not a license for fabricated proof claims. §7 defines exactly what may be asserted about the pattern's instances and from which canonical source.

---

1. The Pattern in One Page

A TGA instance is a routing policy at an automation boundary, composed of four declared elements:

1. The tractability gate. A deterministic test deciding whether the case is algorithmically resolvable at all. Cases failing the gate route to mandatory human handling - the automation never sees them. (Path A in the worked instance below.)
2. The authority bounds. The automated actor's decision space, stated in the domain's own units (dollars, percent-of-order, discount depth, schedule deviation) and derived from an external anchor where one exists (regulation, contract, policy). The machine may act freely within bounds and may not act outside them - structurally, not aspirationally. (In the worked instance: offers computed from state-regulated liability limits; conversational negotiation only within authorized limits.)
3. The acceptance gate. The counterparty's explicit acceptance closes the case automatically; rejection routes to the next rung - a richer automated mode or a human - never to silent retry. (Paths B and C below.)
4. The exception path and audit artifact. Every above-bounds, out-of-policy, or emotionally-escalated case has a named escalation destination, and every automated action emits an audit artifact reviewable by a human after the fact, verified through an observer independent of the writer (per the canonical-observer principle).

The pattern's economics: the automated slice handles the high-volume, low-variance middle of the distribution at near-zero marginal cost; human attention concentrates on the tails, where it is irreplaceable. The pattern's psychology: a published threshold is itself a trust instrument - customers, clients, and regulators are not asked to trust the AI; they are shown the policy that bounds it.

---

2. The Canonical Production Instance - Moving-Company Claims Routing

Stated from the canonical claims-department context of a national moving and storage company, which is authoritative over any conversational recall of these facts:

• Volume context: ~250,000 moves/year across 90+ branches; ~3% generate claims (~7,500/year).
• Tractability gate + authority bounds: the mobile claims journey collects structured inputs (items, weights, photos, desired resolution); an algorithm determines whether an automated settlement offer can be generated and computes it from state-regulated legal limits of liability (e.g., $0.60/lb for Released Value coverage in Texas), with elevated offers permitted based on what the customer seeks and move cost. Automated offers range from 0.1% to 11% of total move cost.
• Three-path routing: Path A - no offer generated (algorithmically intractable) → mandatory human oversight; the AI never handles these. Path B - offer accepted in the mobile journey → resolved automatically. Path C - offer rejected → the AI claims agent conducts the resolution conversation, within authorized limits, under the nine-stage framework, with escalation to a senior specialist as a defined rung.
• The governing insight: "Empathy is not a cost. It is a financial lever." The automated slice exists because the algorithm already did the evaluation; the conversational layer's job is the human-grade resolution experience, not re-adjudication.

This instance demonstrates every contract element of §1 in production.

---

3. The Contract Surface

A conforming TGA instance declares, in its instance canon (a note with role: tga_policy or equivalent frontmatter on the owning artifact):

Coverage-flavor declarations (instance-specific, on top of the base six). A coverage instance (the reply and lead flavor of TGA) declares these additional parameters within the contract surface above; they extend the base six elements, they do not replace them:

• Response-time SLA: the target window the published policy promises. For the AI SDR positive-reply touch it is 3 to 5 minutes; the form, chatbot, and lead-magnet touch is a separate, tighter SLA of under 60 seconds.
• Force-AI max-time-thresholds: the hard threshold above which the AI acts regardless of roster state (for example, greater than 2 minutes, or greater than 20 minutes by offer type). This is the coverage invariant made structural: no gap goes uncovered past the threshold.
• Handler roster plus waterfall and priority assignment: who is eligible to take a touch, and in what order the work falls across handlers.
• Per-channel and per-actor on and off toggles: turn the AI off entirely, or the humans entirely, per channel and per actor.
• The driving structure: the graph plus role configuration, which is where the user authors the coverage policy (the same OCP role and org structure that governs the rest of the platform).

These are the coverage flavor's instance-specific declarations; the base six elements still apply unchanged.

Enforcement is layered: structural bounds via dispatch gates and use-case invariants for agent executors; human-in-the-loop handoff contracts for the human rungs; the audit artifact per the independent-observer discipline.

---

4. Position in the Executor-Governance Stack

Four documents govern four different questions, and conflating them is the most common confusion at this boundary:

• This doctrine (TGA) - when does execution stay automated vs. route to a richer mode or a human? The routing policy at the boundary.
• Human-in-the-loop handoff contracts - what is the human node once routed: its contract surface, exit criteria, operator experience. TGA hands off; the handoff contract receives.
• Agent runtime governance - how the agent executor's bounds are mechanically enforced: dispatch gate eligibility, use-case domain invariants. TGA's authority bounds are implemented here for agent rungs.
• The Darwinian Loop's Six-Rule Observation Gate - the campaign-mutation instance of TGA: agents may not sunset, publish, or open cohorts except through threshold-and-criteria gates. The Loop's discipline is TGA applied to the system's own evolution.

One sentence: TGA is the policy, runtime governance is the enforcement, the handoff contract is the human rung, and the Darwinian gate is the pattern eating its own dog food.

---

5. Generalization Registry

Instances and candidates, scored honestly:

1. Moving-company claims routing - LIVE. The canonical instance (§2). Instrumentation status: routing structure and authority bounds canonical; outcome-rate reporting pending formal instrumentation (see §7).
2. After-hours call handling for a mobile-services company - CANDIDATE (strongest second instance). The routing matrix, daily transcript review ticket, and escalation roster already exist in that org's canon; what remains is declaring the contract surface (§3) explicitly against those artifacts and verifying bounds enforcement.
3. LeadsBlaster delivery make-goods - CANDIDATE. The KPI system's delivered-vs-committed intervention ("potential make-good") is a threshold with a human rung; a TGA declaration would bound any future automated make-good issuance.
4. Commission-threshold auto-tendering (3PL logistics) - HYPOTHETICAL. The direct transfer of the claims-routing pattern to freight tendering; it remains paper until real gates fire. No engineering hours.
5. Scheduling exception policy - HYPOTHETICAL. Reschedule/cancel/no-show handling bounded by declared policy before automation deepens.
6. Reply and Lead Coverage with Waterfall Assignment: the productized flagship instance (the SalesBlaster software value proposition). The TGA pattern applied to inbound reply and lead handling: automate beneath the declared time-threshold (the response-time SLA), govern exceptions through a configurable handler roster, route to a human rung, and audit every automated touch. The role-config and coverage model: a clock-in roster (human reps log in, the AirCall and RingCentral analog); waterfall and priority assignment of work across handlers; interleaved human and AI fallback; max-time-thresholds above which the AI fires no matter what (for example, greater than 2 minutes, or greater than 20 minutes depending on offer type); per-channel and per-actor on and off toggles (turn the AI off entirely, or the humans entirely); all of it driven by the graph structure and role configurations the user sets, which ties this instance to the OCP role and org structure. The coverage invariant is no uncovered gap, formalized as the human/AI coverage state machine: sign on, become responsible within the SLA window, hit the off-button, the AI covers the gap. Relation to the live claims instance (§2): claims routing is governed by a dollar-authority threshold; coverage is governed by a time-SLA threshold plus a handler roster; same pattern, different threshold dimension. Status, scored honestly per §7 and the mutation rule below: the mechanism is canonical and designed, and this is the core product the platform sells, but it is not yet a single instrumented production deployment; it does not count as an instrumented instance until its six §3 contract elements are declared per instance and the audit artifact is flowing.

Registry mutation rule: free-edit with dated evidence lines; an instance counts only when all six contract elements are declared and the audit artifact is flowing.

---

6. The Direct Response Layer

Kennedy/Shah governance, stated plainly: never automate what you have not first made accountable. A threshold without its intervention pairing is decorative (the KPI pre-commitment rule - no metric ships without its intervention - is the measurement-side sibling of this doctrine); an automation without its declared bounds is a liability compounding quietly.

The trust mechanism for Stage 1–2 markets: in verticals where AI offers still convert on direct claims, the buyer's dominant objection is not capability but control - what happens when it's wrong? TGA is the answer productized: lead the sales conversation with the routing policy, not the model. "The system resolves what the policy authorizes, routes everything else to your people, and every automated action is reviewable" is a governance story, and governance stories close operators whom capability stories frighten.

NEPQ application: the diagnostic battery for any prospective automation engagement includes eliciting the prospect's implicit thresholds ("at what dollar amount would you want a human to see it first?" / "which caller types must never be handled automatically?"). The prospect authors their own §3 contract during discovery; the close presents it back as the deployment policy. The prospect's policy in the prospect's words - Carnegie's root layer, operationalized.

Deliverable-stacking: in proof-saturated Stage 4–5 conversations, the named sub-deliverable is the currency. "Threshold-governed automation policy: tractability gate, authority bounds, acceptance gate, exception ladder, independent-observer audit" is five named deliverables in one sentence.

The value-prop trinity (Power Dialer plus Coverage plus A/B). The platform's entire software value proposition is three named mechanisms in one offer, each carrying an engineering form and a direct-response form:

• Power Dialer. Engineering: channel-agnostic outbound execution (voice, email, text), every step an agent walking the graph. Direct response: Kennedy's follow-up discipline industrialized (more follow-up than feels comfortable, automated to saturation) plus Hormozi's variant assembly line.
• Coverage. Engineering: this doctrine's own TGA trust mechanism applied to the reply surface, the human/AI coverage state machine with no uncovered gap. Direct response: the trust and governance story productized; no lead is ever dropped; the published policy is that a human handles it within the SLA or the AI does, and nothing falls through. Coverage is the promise; the SLA is its teeth.
• A/B. Engineering: ceteris-paribus variant selection. Direct response: the Darwinian Loop's selection mechanism turned on the offer itself.

Deliverable-stacking, applied to the trinity. "Power Dialer, Coverage engine, A/B selection" is three named mechanisms in one value proposition, the named-sub-deliverable currency for Stage 4 and 5 markets. The trinity is what the platform sells; this doctrine is the governance spine of its middle term.

---

7. The Proof Asset - Extraction Rules

The pattern is a marketing asset (grant applications, case studies, sales decks) only when extracted under these rules:

1. Mechanism claims come from canon; results claims come from instrumentation. The following are canonical and citable today: the three-path routing architecture; algorithmic offers computed from state-regulated liability limits; the 0.1%–11%-of-move-cost offer range; mandatory human handling for algorithmically intractable claims; conversational AI restricted to the rejected-offer slice within authorized limits. The following may NOT be claimed until formally instrumented and recorded in canon: resolution rates, acceptance rates, settlement deltas, CSAT movements, cost-per-claim savings.
2. The mechanism one-liner (approved form): "Claims are resolved by a published policy: the algorithm offers within legally-anchored bounds, customers who accept are done in minutes, customers who decline get a human-grade AI conversation within authorized limits, and anything the policy can't bound goes straight to a person."
3. No figure enters a proof asset that does not appear verbatim in a canonical artifact. (This rule exists because a figure from conversational recall once briefly entered the canon pipeline and was corrected against the canonical context before publication. The pipeline worked; the rule makes it permanent.)
4. Results slots in collateral are rendered as instrumentation-pending placeholders, never as estimates.

---

8. First Principles

TGA-1. The threshold is declared, anchored, and published - never emergent. Bounds derived from an external anchor (regulation, contract, policy) survive scrutiny; bounds that emerged from prompt iteration do not.

TGA-2. Bounds are structural, not behavioral. The agent is unable to exceed authority (dispatch gate + use-case invariants), not instructed not to. Instructions are tone; structure is governance.

TGA-3. Rejection routes upward, never around. Every "no" from the counterparty moves the case up a declared rung. Loops that re-offer without escalation are the automation arguing with a human.

TGA-4. The writer never grades its own work. Audit flows through an observer independent of the acting system (canonical-observer principle). Transcript review, settlement sampling, and exception postmortems are the pattern's immune system.

TGA-5. Human attention is spent on the tails by design. The pattern's ROI is the redistribution of judgment from the tractable middle to the consequential edges - measure it there.

TGA-6. The published policy is part of the product. What the counterparty can read, they can trust; what they can trust, they accept faster (the claims instance's empathy-as-financial-lever insight, generalized to policy transparency).

---

9. Anti-Patterns

TGA-AP1 - Threshold theater. A declared threshold with no enforcement layer and no audit artifact. Looks like governance; is a prompt suggestion.

TGA-AP2 - Silent threshold drift. Bounds widened in config or prompt without a dated canon entry. All bound changes are logged with rationale, per the system's append-only discipline.

TGA-AP3 - Automating the unaccountable. Deploying the automated slice before the measurement and intervention pairing exists (violates the measurement-first sibling rule).

TGA-AP4 - Full-autonomy creep. Each "the AI handled that edge case fine" expanding de-facto authority without a declaration. Authority grows by amendment, not by anecdote.

TGA-AP5 - Bottleneck inversion. Routing the tractable middle to humans "to be safe," collapsing the economics and burning the human attention the tails needed. The gate exists to be trusted within bounds.

TGA-AP6 - Recall-sourced proof. Marketing numbers from memory instead of canon (§7 rule 3).

---

End of Threshold-Governed Automation Doctrine v1.0.0.

Automate the middle. Govern the tails. Publish the policy. Let the audit watch the machine, and let the machine earn its authority one declared amendment at a time.