How to Design a Pipeline Around Evidence, Not Activity

Who this framework is for

This framework is for teams replacing spreadsheet funnels, CRM habits, or automation sequences in which sent, opened, clicked, and assigned are mistaken for commercial progress. It helps product, sales, operations, and data owners agree on what a state actually means.

Contents

1 Start with the decisions the pipeline must support 2 Define every state as an evidence claim 3 Separate observation, evidence, interpretation, and action 4 Set evidence sufficiency and freshness 5 Keep Unknown, conflict, and hard gates visible 6 Build reversible transition decisions 7 Create accountable tasks from decision gaps 8 Measure evidence quality and controlled flow 9 Illustrative example 10 Common mistakes 11 What this does not mean 12 Common questions 13 Sources and methodology

1. Start with the decisions the pipeline must support

An activity-led pipeline begins with what a tool can count. An evidence-led pipeline begins with the decisions a team must make:

• Is this person or organization now a lead in the defined pursuit?

• Does the record satisfy the current qualification criteria?

• Did a new material fact change urgency, interpretation, or next action?

• Is meaningful human action required, and who owns it?

• Should outreach continue, pause, change contacts, or stop?

• Is the current state still accurate and fresh enough to use? For each decision, identify the consumer and consequence. A routing decision used only to prioritize a low-risk review can tolerate more uncertainty than a decision that triggers contact, suppresses a company, changes billing treatment, or tells a tenant that human action is required. The U.S. Government Accountability Office's Standards for Internal Control emphasizes defining objectives, using quality information, performing control activities, and monitoring. GAO, "Standards for Internal Control in the Federal Government" It governs U.S. federal internal control, not commercial GTM. The useful analogy is that a process needs explicit objectives, relevant information, controls, and monitoring rather than counts alone. Write one sentence for each supported decision: [Owner] decides [decision] for [scope] using [minimum evidence] under [rule version], because the result controls [action or exposure]. If no one can name the decision or consequence, the field may be telemetry rather than a pipeline state.

2. Define every state as an evidence claim

A state should be a claim that can be supported, challenged, and reversed. For SellersGPT:

• Lead claim: an identifiable person or organization has a recorded engagement or response event that satisfies the applicable lead-creation rule for one pursuit.

• Qualified-lead claim: the lead satisfies all applicable qualification criteria under one versioned rule, with no blocking exclusion.

• Signal claim: a material fact concerning a qualified lead changes interpretation or next action.

• Active-lead claim: meaningful human interaction or a verified required tenant/human action creates active status; owner assignment and acceptance remain separate routing controls. For each claim, write: Contract element Question Scope Which account, person, application, offer, geography, and pursuit? Predicate What must be true? Required evidence Which facts and sources prove it? Exclusions Which conditions block it? Rule version Which definition was effective? Authority Which process or person may confirm or reverse it? Freshness When must the evidence be reviewed? Consequence What work, visibility, or automation changes? Email sent fails this test as a lead state. It proves only that a send action was attempted or requested. Provider accepted message proves a delivery-system event, not human receipt. Opened may be technically ambiguous. Clicked can support an engagement event, but the lead and qualification claims still need their own rules.

action

Use four layers: 1 Observation: the raw event or artifact, such as a reply payload, form submission, click record, delivery code, or public document. 2 Evidence statement: the exact fact derived from the observation, with source, date, and limitations. 3 Interpretation: what the fact means under the current state or transition rule. 4 Action: what a named owner or controlled automation may do next. Google Analytics documents events and event parameters within its measurement product. Google Analytics, "About events" That is not a B2B qualification standard, but it demonstrates that an interaction can be stored with attributes before a business interpretation is applied. Example:

• Observation: reply body states, Please send the washdown specification and copy our maintenance supervisor.

• Evidence: sender requests a defined technical document and a second participant as of the reply time.

• Interpretation: material signal on one qualified lead; human response required; role of the maintenance supervisor remains Inferred until verified.

• Action: pause conflicting automation, assign the account owner a contextual response task, and update the evidence ledger. This separation prevents a classifier label from replacing source evidence. It also permits a later reviewer to disagree with the interpretation without denying that the reply occurred.

4. Set evidence sufficiency and freshness

Evidence sufficiency depends on the claim and consequence. Define:

• required fields and acceptable source types;

• whether corroboration is required;

• scope match across entity, site, application, geography, and time;

• maximum evidence age or review trigger;

• handling of indirect, derived, or self-reported evidence;

• reviewer authority for conflicts or exceptions. Use Known, Inferred, Unknown, and Contradicted as evidence states where appropriate. Do not convert confidence into truth. A 90% model confidence does not make a claim Known if the underlying source is not scope-matched. NIST's guidance on data integrity discusses protecting data from improper alteration or destruction and ensuring authenticity and integrity according to risk. NIST SP 1800-25, "Data Integrity: Identifying and Protecting Assets Against Ransomware and Other Destructive Events" It is cybersecurity guidance, not sales-data policy. Its scope supports the narrower need to protect material evidence from unauthorized or unexplained change. Freshness is not one global number. Employment evidence can become stale after an organizational change. A product claim changes after a revision. A buyer's timing statement may expire on the date they gave. Attach a review trigger to the fact rather than refreshing all records on one arbitrary cadence.

5. Keep Unknown, conflict, and hard gates visible

Evidence-led design avoids three common shortcuts. First, Unknown is not zero. If employee count is required but unavailable, a score of zero falsely states that the company has no employees or no fit. Preserve the missing condition and route it according to materiality. Second, conflict is not average. If a first-party facility page says one site and a current buyer reply names another, store both facts, scope them, and resolve the contradiction. Do not choose the source that produces the preferred state. Third, hard gates are not weights. Unsubscribe, company stop, prohibited geography, critical product mismatch, invalid identity, or an applicable exclusion cannot be offset by other positive evidence. The FTC's CAN-SPAM guide explains U.S. opt-out obligations for covered commercial email. FTC, "CAN-SPAM Act: A Compliance Guide for Business" The point for pipeline design is narrow: a communication stop needs precedence and auditability. It must not become minus 20 points inside a fit score.

6. Build reversible transition decisions

Every transition should create an immutable decision record containing:

• prior and proposed state;

• trigger event;

• evaluated criteria and evidence states;

• hard-gate result;

• decision and reason code;

• rule and data versions;

• deciding actor or process;

• timestamp and effective time;

• tasks created, canceled, or retained;

• next review or terminal disposition. NIST's log-management guide addresses establishing policies for log generation, storage, access, and disposal. NIST SP 800-92, "Guide to Computer Security Log Management" The document is not a sales-workflow specification. It supports preserving a reviewable record of material process events. Reversal is part of correct design. If an account was qualified using a distributor criterion but evidence later shows it is a direct end user outside the pursuit, append a reversal with the new fact. Do not erase the original decision. Measure why reversals occur: new information, stale evidence, rule change, identity merge, provider correction, or processing defect. Delivery-state corrections need similar care. RFC 3463 defines classes and subject detail for enhanced mail status codes. IETF, RFC 3463 Preserve the provider event and normalized class so a mapping correction can be audited.

7. Create accountable tasks from decision gaps

Tasks should close a decision gap, not reward motion. A complete task states:

• decision or evidence gap;

• subject and scope;

• current evidence and unresolved condition;

• required action and permitted boundary;

• owner and fallback;

• due time and escalation;

• completion evidence;

• state or next decision affected. Research account is weak. Verify whether Plant 2 uses 480 V, three-phase power before technical fit can be evaluated; owner: application engineer; due: August 28; acceptable evidence: current plant document or buyer confirmation is actionable. Completing the task does not guarantee a state change. It may confirm fit, disprove it, leave it Unknown, or identify a stop. Measure whether tasks resolve the stated gap, not only whether users click Complete.

8. Measure evidence quality and controlled flow

Add quality measures to activity measures: Dimension Example diagnostic Why it matters Completeness records with every required evidence field resolved or explicitly Unknown exposes hidden blanks Provenance material claims linked to accessible sources supports review Freshness claims past their review trigger prevents stale state Accuracy sampled decisions confirmed by independent review tests correctness Reversal transitions reversed by reason finds weak rules or data Control suppression propagated before next job tests hard-gate performance Ownership tasks accepted before service deadline tests operational handoff Age state and task age distributions identifies stranded work Do not combine these into one opaque pipeline health score. A dashboard can show them together, but a severe suppression failure should remain visible even if every other measure is green.

9. Illustrative example

Illustrative example - explicitly fictional. IronVale Sensors reviews a fictional 100-record lead cohort. Its old dashboard reports 70 engaged records because it treats any click, reply, or form as progress. The evidence-led review finds: Evidence result Records Identifiable leads with valid event provenance 52 Anonymous or ambiguous events not safely matched 11 Duplicate event-derived records merged 7 No engagement event; activity label caused by send/open telemetry Total 100 Of 52 leads, 18 satisfy the versioned qualification criteria, 22 remain leads with material Unknowns, 8 fail an application criterion, and 4 have hard stops. The equation is 18 + 22 + 8 + 4 = 52. Six of the 18 qualified leads have material signals, and three require human action and become active leads. The review does not claim that the old team lost 67 opportunities. It shows that one activity label concealed several conditions. The team now reports 52 leads, 18 qualified leads, six qualified leads with signals, and three active leads, while retaining the anonymous, duplicate, failed, and stopped records for audit and learning. These counts are fictional, not benchmarks.

10. Common mistakes

• Naming a stage after a completed action.

• Treating clicks, opens, assignments, or tasks as commercial proof.

• Allowing a score to hide Unknown or override a hard gate.

• Storing only classifier output and discarding source evidence.

• Updating state without rule version or effective time.

• Deleting a prior decision after correction.

• Measuring task completion rather than evidence resolution.

• Creating one pipeline-health score that hides severe control failures.

• Copying another platform's labels without defining consequences.

11. What this does not mean

Evidence-led design does not eliminate judgment, uncertainty, or buyer complexity. It does not mean every low-risk event requires manual review or that more fields always improve quality. Collect and retain only what the decision needs. This framework does not prescribe market, ICP, buying-role, billing, or AI-governance policy. It does not promise that better evidence creates more revenue. It makes the pipeline's current claims, limitations, and responsibilities more trustworthy.

12. Common questions

Can an activity ever create a state transition? Yes. An activity can create an event that supplies the evidence required by a rule. The transition occurs because the evidence claim is satisfied, not because the activity category has inherent commercial meaning. Should every pipeline field have a source link? Every material evidence claim needs provenance. Derived operational fields should link to their input events and rule version. Pure display or system fields may need lineage rather than a public URL. Can a score be used in an evidence-led pipeline? Use transparent scores for prioritization when inputs, weights, Unknown handling, and limits are visible. Do not let a score establish a hard factual state or override an exclusion, suppression, or mandatory approval. How often should evidence quality be audited? Set frequency from consequence, volume, error history, change rate, and control risk. Combine ongoing exception monitoring with scheduled independent samples; do not assume one universal cadence.

13. Sources and methodology

The state claims, four-layer model, evidence contract, transition record, and quality diagnostics are SellersGPT editorial tools. Sources support limited principles concerning objectives, event data, integrity, log records, mail status, and communication stops. They do not validate commercial outcomes. 1 PP05-GAO-GREENBOOK - U.S. Government Accountability Office, "Standards for Internal Control in the Federal Government." https://www.gao.gov/products/gao-14-704g Accessed August 25, 2026. 2 PP05-GA4-EVENTS - Google Analytics Help, "About events." https://support.google.com/analytics/answer/9322688 Accessed August 25, 2026. 3 PP05-NIST-LOGS - National Institute of Standards and Technology, SP 800-92, "Guide to Computer Security Log Management." https://csrc.nist.gov/pubs/sp/800/92/final Accessed August 25, 2026. 4 PP05-NIST-DATA - National Institute of Standards and Technology, SP 1800-25, "Data Integrity: Identifying and Protecting Assets Against Ransomware and Other Destructive Events." https://csrc.nist.gov/pubs/sp/1800/25/final Accessed August 25, 2026. 5 PP05-RFC3463 - Internet Engineering Task Force, RFC 3463, "Enhanced Mail System Status Codes." https://www.rfc-editor.org/rfc/rfc3463 Accessed August 25, 2026. 6 PP05-FTC-CANSPAM - U.S. Federal Trade Commission, "CAN-SPAM Act: A Compliance Guide for Business." August 25, 2026. https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business Accessed