On this pattern

System Pattern 09 / Documents, Knowledge & Context

No evidence. No conclusion.

Evidence-Bounded Inference

A generated conclusion is allowed to exist only when its material claims remain connected to current, permitted and sufficient evidence.

Maturity: Reference design / Specified
Critical uncertainty: The model may explain evidence. It may not manufacture missing evidence or convert uncertainty into a conclusion.
Human boundary: Interpret authorized ambiguity and make decisions within delegated scope
Applications: Support answers · Contract analysis · Incident hypotheses · Call compliance

Support, legal, incident, and compliance questions enter nine evidence-control layers and end as supported, contradicted, or insufficient evidence.

Problem shape

The problem beneath them

Bind every material claim to current, permitted, versioned evidence; search for contradiction; and abstain when the evidence cannot carry the conclusion.

Generation is cheap; warranted conclusion is not. Retrieval can return semantically similar but obsolete, unauthorized, incomplete, or contradictory material. A citation can exist without supporting the sentence beside it. Long context can contain the answer while a model attends to the wrong passage. Fluent wording then makes weak evidence appear settled.

The pattern treats a response as a set of claims, not one indivisible paragraph. Each material claim has an evidence contract: source types permitted, temporal scope, required specificity, minimum support, contradiction handling, and authority. Evidence retains stable identity, version, location, and access basis. The system validates whether cited spans entail the claim rather than merely share vocabulary.

Abstention is therefore an engineered outcome. “Insufficient evidence” should identify what is missing and who can obtain or judge it. Human review is not a ceremony that converts missing support into truth; a reviewer may approve a bounded interpretation, request evidence, record a policy judgment, or reject the claim.

The model may explain evidence. It may not manufacture missing evidence or convert uncertainty into a conclusion.

Evidence-bounded inference improves traceability and epistemic control. It does not guarantee that an authoritative source is factually correct, resolve conflicting authorities automatically, or confer legal, clinical, or operational authority on a generated explanation.

Different problems, same shape

The surface changes. The decision structure persists.

01 / Support

Which policy supports this answer?

A fluent response cites no current governing source.

02 / Legal

Which clause governs this conclusion?

Several versions or clauses may apply to the same question.

03 / Incident response

What evidence supports this hypothesis?

Symptoms suggest a cause but competing explanations remain.

04 / Compliance

Where in the call did this happen?

A consequential finding must resolve to timestamped source material.

All four questions converge into Evidence-Bounded Inference.

Exploded pattern

Open the mechanism at every decision boundary.

01
Define the claim contract

Question

What claims may be produced, and what evidence would make each permissible?

Responsibility

Specify materiality, source eligibility, temporal scope, required granularity, contradiction policy, and authority boundary.

Input

User task, domain policy, consequence, and requested output form.

Output

Machine-checkable claim and evidence requirements.

Stops when

The task is bounded or rejected as unsupported by an available evidence contract.
02
Retrieve current permitted sources

Question

Which sources may answer this question for this user and time?

Responsibility

Enforce authorization, jurisdiction, effective dates, tenancy, and source status during retrieval.

Input

Claim contract, query, identity, policy context, and source catalog.

Output

Eligible candidate sources plus retrieval trace.

Stops when

The bounded retrieval plan completes or source access is denied explicitly.
03
Preserve source identity and version

Question

Can every cited passage be recovered exactly as evaluated?

Responsibility

Retain stable source ID, version, checksum, effective period, location, and access basis.

Input

Retrieved documents, transcript segments, records, and metadata.

Output

Versioned evidence objects with resolvable locators.

Stops when

Later reconstruction does not depend on mutable “latest” content.
04
Locate supporting evidence

Question

Which exact spans support each proposed material claim?

Responsibility

Select minimal sufficient passages and map them to atomic claims.

Input

Eligible evidence objects and candidate claims.

Output

Claim-to-span support links with context.

Stops when

Each claim has candidate support or is marked unsupported.
05
Search for counterevidence

Question

What permitted evidence conflicts with, narrows, or supersedes the claim?

Responsibility

Run contradiction-oriented retrieval across relevant sources and versions.

Input

Candidate claim, supporting spans, source hierarchy, and known alternatives.

Output

Counterevidence links, unresolved conflicts, and supersession signals.

Stops when

Required opposing searches complete within the defined evidence scope.
06
Validate claim-to-source alignment

Question

Does the cited evidence actually entail the wording and scope of the claim?

Responsibility

Check entities, qualifiers, negation, dates, jurisdiction, numerical values, and citation coverage.

Input

Atomic claim, support, counterevidence, and claim contract.

Output

Supported, contradicted, or misaligned assessment per claim.

Stops when

Material wording is either grounded or removed from the proposed conclusion.
07
Measure evidence sufficiency

Question

Is the evidence adequate for the consequence, not merely relevant?

Responsibility

Evaluate authority, independence, completeness, recency, contradiction, and missing required evidence.

Input

Alignment results, source hierarchy, consequence, and sufficiency rules.

Output

Supported, contradicted, or insufficient-evidence route.

Stops when

The disposition is justified without using model confidence as a substitute for evidence.
08
Abstain or escalate

Question

What should happen when the contract cannot be satisfied?

Responsibility

Suppress unsupported conclusions and return precise evidence gaps or an authorized review packet.

Input

Failed sufficiency checks, consequence, deadline, and escalation policy.

Output

Bounded abstention, request for evidence, or human-review task.

Stops when

No unsupported claim leaks into an authoritative output.
09
Record the human disposition

Question

How did an authorized person resolve interpretation, conflict, or missing evidence?

Responsibility

Capture decision, authority, rationale, evidence reviewed, qualifications, and expiration.

Input

Review packet and reviewer judgment.

Output

Attributable disposition that does not rewrite source evidence.

Stops when

The decision and its scope are reconstructable and reviewable.

Decision forks

Every branch states why it exists and when it escalates.

Signals, decisions, reasons, and escalation conditions
Signal	Decision	Reason	Escalates when
Current authoritative source directly supports claim	Mark supported within stated scope	Evidence satisfies authority, time, and alignment	Consequence requires mandatory human approval
Authoritative source contradicts claim	Mark contradicted; do not soften conflict	Counterevidence governs the disposition	Authorities conflict or applicability is disputed
Citation is topically related but not entailing	Remove or rewrite claim	Relevance is not support	The requested conclusion cannot be narrowed safely
Required source is inaccessible	Return insufficient evidence	Permission failure must not be bypassed	An authorized person can lawfully obtain it
Sources apply to different versions or jurisdictions	Separate scopes or escalate	Combining them creates a false universal claim	Applicability needs domain authority
Support exists but a required qualifier is absent	Narrow wording	Claim must not exceed evidence	Qualification changes a consequential action
Model confidence is high but support is missing	Abstain	Confidence describes output behavior, not truth	Evidence can be gathered or specialist review is required

Operating paths

Clear, ambiguous, and failed work all reach explicit states.

Clear path

Define claim contract, retrieve eligible sources, map exact support, check contradiction and alignment, then publish bounded claims.

Final state: Supported or Contradicted.
Owner: Evidence service until disposition; consuming authority remains responsible for consequential action.
Evidence: Claim text, source IDs and versions, exact spans, retrieval trace, checks, and policy version.
Recovery: Re-evaluate when sources, effective dates, access, or claim wording change.

Ambiguous path

Expose conflicts or gaps, suppress a definitive conclusion, and send an evidence packet to an authorized reviewer.

Final state: Insufficient evidence.
Owner: Domain reviewer or source owner.
Evidence: Candidate support, counterevidence, missing requirements, applicability conflict, and model/system trace.
Recovery: Obtain evidence, narrow the claim, record a scoped human disposition, or retain abstention.

Failure path

Halt publication when provenance, permissions, indexing, or validation integrity fails.

Final state: Insufficient evidence.
Owner: Knowledge platform owner until trustworthy processing resumes.
Evidence: Failed component, affected corpus/version, queries, partial results, and containment event.
Recovery: Repair and re-index from authoritative sources, validate locators, and replay affected claims.

Authority map

Capability does not grant authority.

RULE

May decide: Enforce source eligibility, dates, required checks, and mandatory abstention
May not decide: Resolve substantive ambiguity not encoded in policy
Required evidence: Rule/version, context, evaluated conditions, and result

MODEL

May decide: Propose claims, find passages, compare wording, and explain conflict
May not decide: Invent evidence, override access, or authorize consequential action
Required evidence: Model/version, prompt/task, cited spans, and uncertainty

SYSTEM

May decide: Retrieve, preserve provenance, validate contracts, and suppress unsupported output
May not decide: Treat a retrieval score as factual sufficiency
Required evidence: Retrieval trace, source versions, mappings, checks, and state

HUMAN

May decide: Interpret authorized ambiguity and make decisions within delegated scope
May not decide: Backdate evidence, erase contradiction, or claim authority not held
Required evidence: Identity, remit, evidence reviewed, rationale, scope, and expiry

EXCEPTION

May decide: Hold missing, conflicting, stale, or inaccessible evidence cases
May not decide: Emit a definitive conclusion by default
Required evidence: Trigger, evidence gaps, owner, deadline, and recovery options

Failure modes and recovery

A failed path remains owned, evidenced, and recoverable.

Failures, detection, containment, recovery, and owners
Failure	Detection	Containment	Recovery	Owner
Stale policy remains searchable	Effective-date or checksum audit disagrees	Exclude stale version from current-answer route	Re-index authoritative version and replay affected claims	Content owner
Citation does not support adjacent claim	Alignment check or review finds scope mismatch	Suppress claim and preserve trace	Rewrite narrowly or locate sufficient evidence	Evidence service
Chunking removes decisive qualifier	Source-span comparison reveals omitted context	Expand context; reject fragment-only support	Reprocess source with versioned segmentation	Knowledge platform
Cross-tenant retrieval exposes evidence	Authorization audit or tenant mismatch	Block response, isolate index, start incident process	Repair partitioning and assess exposure	Security owner
Counterevidence search is skipped	Missing required stage in trace	Prevent final disposition	Run required search and revalidate all claims	Evidence service
Human approval lacks rationale	Required disposition fields absent	Keep item non-final	Return to authorized reviewer for complete record	Domain owner

Invariants and guarantees

Properties the structure is designed to preserve.

Every material published claim maps to recoverable, permitted, versioned evidence.
Source locators resolve to the content actually evaluated, not an untracked latest version.
Contradictory evidence remains visible and affects disposition.
Missing access, missing evidence, and low support never become affirmative conclusions.
Model confidence never substitutes for source sufficiency.
Human dispositions retain scope, authority, rationale, and expiration where relevant.
Consequential action remains outside the model’s authority unless separately delegated to an accountable role.

What changes between implementations

The constraints determine the final mechanism.

The claim contract changes with domain and consequence. Support may require one governing policy, several independent observations, a complete contract hierarchy, or exact transcript timestamps. Retrieval architecture follows corpus size, update frequency, source permissions, language, and latency. Alignment may combine deterministic checks, specialist models, and human review.

Sufficiency thresholds depend on false-conclusion cost, review capacity, and reversibility. Some outputs can safely quote evidence without concluding; others must withhold the entire response. Retention and trace detail follow privacy and legal requirements. The invariant is that implementation convenience cannot widen the claim beyond current permitted evidence.

Evidence chain

Follow the pattern into systems and software.

Architecture and controls specified; no production results published.

version-aware-support-rag relates to current-policy answers; incident-triage to evidence-backed hypotheses; contract-compliance to clause-level conclusions; and on-prem-sales-call-compliance to timestamped findings. These are reference-design relationships and do not establish that this exact pattern was implemented, tested, or measured in those cases.

Entity resolution can preserve which real-world subject the evidence concerns. Risk-aware cascades can allocate model depth after evidence and authority floors are enforced. Private inference constrains where the entire evidence path may execute.

This is a REFERENCE DESIGN with specified controls; no production results are published. No open-source implementation is linked. No evaluation fixture, test, or benchmark is linked. Case-study references identify conceptual composition points only and carry no standalone implementation claim.

Reference designEvidence-Bounded InferenceThis specified pattern and its controls.
Reference designSupport drafts from current policyRelated case-study system.
Reference designIncident triage with linked evidenceRelated case-study system.
Reference designContract review across linked clausesRelated case-study system.
Reference designCall review inside a private networkRelated case-study system.

Known boundaries

Limitations and non-fit

Grounding cannot make a false source true or settle conflicts between authorities. Retrieval may miss decisive evidence, especially in incomplete, scanned, multilingual, or rapidly changing corpora. Claim decomposition and alignment checks reduce risk but are themselves fallible and need domain-appropriate evaluation before automation.

The pattern is not a substitute for professional judgment, discovery obligations, incident investigation, or policy ownership. It is excessive for unconsequential creative generation where factual grounding is not requested. It is non-fit when source rights prohibit the required processing, provenance cannot be retained, or the consuming process demands certainty that available evidence cannot provide.

Related patterns

Continue through the adjacent decision structures.

Evidence-bounded inference changes the unit of trust from a fluent response to a reconstructable claim. The useful output may be support, contradiction, or a precise refusal to conclude. All three are operational results when the system preserves what is known, what is missing, and who has authority to decide next.

Adapt the pattern

Bring the problem, the boundary, and the consequence of being wrong.

Let's build something real