Document Type: Framework
Status: Active
Version: v1.1
Authority: HeadOffice
Owning Brain: SIT Brain
Parent: SIT Brain Canon
Applies To: All MWMS Brains, AI Employees, AI Agents, Agentic Workflows, Model Routers, Review Systems, Automation Workers, Development Workflows, Research Systems And High-Risk Decision Processes
Last Reviewed: 2026-06-21
Source Of Truth: MCR
Source / Origin: MWMS Independent Model Review And Rescue Routing Framework v1.0 and AI Automations by Jack multi-model research, panel discussion, fact-checking, evidence comparison, specialist routing, and synthesis material. The existing v1.0 framework already establishes independent review, repeated-failure rescue, specialist routing, structured consensus, disagreement handling, and human authority controls.
Purpose
The MWMS Independent Model Review And Rescue Routing Framework defines how MWMS assigns independent AI review, detects unproductive reasoning loops, transfers failing work to a rescue model, compares separate interpretations, and governs disagreement between distinct reasoning systems.
The framework prevents an AI model from being treated as the sole authority for reviewing work it created.
It establishes the conditions under which MWMS must:
request independent review
route work to a different model or reasoning architecture
stop repeated failed attempts
invoke specialist review
separate evidence retrieval from interpretation
compare competing conclusions
identify agreement and disagreement
preserve minority evidence
route unresolved questions for further research
escalate high-risk work
preserve routing and review evidence
require human approval before execution
The framework exists to improve decision reliability, reduce self-confirming errors, prevent repeated failure loops, expose hidden assumptions, preserve credible disagreement, and ensure that important MWMS work receives appropriately independent scrutiny.
Core Principle
The model that creates an output must not automatically be treated as the only qualified reviewer of that output.
Self-review may improve presentation and identify obvious mistakes, but it does not provide true independence.
Where accuracy, safety, capital, compliance, system integrity, evidence quality, or operational continuity matters, MWMS must use a separate review path with sufficient cognitive and procedural independence from the originating model.
Independent review is not achieved merely by:
asking the same model to reconsider
opening a new conversation with the same unchanged instructions
asking the same agent to critique its own answer
repeating the same prompt
increasing response length
requesting more confidence
asking several models to vote without evidence review
allowing reviewers to see and copy the first conclusion before forming their own
allowing the same failed approach to continue indefinitely
Independent review requires a meaningfully separate:
model
role
context
method
prompt
source pathway
reasoning architecture
validation method
or specialist capability
Rule
Model diversity may improve perspective diversity.
It does not automatically create evidence diversity or factual independence.
Scope
This framework governs independent review and rescue routing for:
Brain decisions
Canon and governance changes
AI Employee outputs
agent-generated plans
research conclusions
fact-checking outputs
evidence synthesis
compliance assessments
financial recommendations
campaign decisions
software and automation outputs
database changes
tool permissions
deployment decisions
external communications
client-facing deliverables
high-impact operational actions
multi-model research workflows
work that repeatedly fails verification
work where models disagree materially
work where source evidence is weak
work where assumptions materially affect the answer
work where minority evidence may reveal risk
work requiring specialist interpretation
Framework Objectives
The framework has eight primary objectives.
Objective One: Separate Creation From Review
The originating model should not control both production and final approval of important work.
Objective Two: Stop Failed Reasoning Loops
Repeated attempts that do not produce verified progress must be stopped rather than disguised as persistence.
Objective Three: Route By Capability
Work must be routed to a reviewer with the required modality, domain, tool, context, or reasoning capability.
Objective Four: Expose Assumptions
Reviewers must identify the assumptions supporting their conclusions.
Objective Five: Separate Evidence From Interpretation
Models should not be treated as independent factual sources merely because they offer different interpretations.
Objective Six: Preserve Disagreement
Credible disagreement and minority evidence must remain visible until resolved or formally accepted as residual uncertainty.
Objective Seven: Maintain Authority
AI consensus must not override Canon, objective evidence, deterministic validation, or authorised human authority.
Objective Eight: Preserve Review Evidence
Review routes, findings, conflicts, decisions, and execution conditions must remain traceable.
Independent Review Definition
Independent review is a separate evaluation of an output by a reviewer that is sufficiently distinct from the originating route.
Independence may arise through:
a different model family
a different reasoning architecture
a separate specialist role
an independent source search
a blind review process
a deterministic validation tool
a separate prompt structure
a different modality-capable system
an authorised human reviewer
A review is stronger when several forms of independence are combined.
Examples include:
different model plus independent source retrieval
specialist model plus deterministic tests
blind model review plus human adjudication
separate research agents plus evidence-focused synthesis
Independence Dimensions
MWMS should assess review independence across seven dimensions.
Model Independence
Was a different model or model family used?
Prompt Independence
Was the reviewer given a review-specific prompt rather than the original production prompt?
Context Independence
Could the reviewer form an initial view without being anchored by the original conclusion?
Source Independence
Did the reviewer inspect evidence independently rather than relying only on the creator’s source list?
Method Independence
Did the reviewer use a different analytical or validation method?
Capability Independence
Did the reviewer possess a specialist capability absent from the creator?
Authority Independence
Was the final decision made by an authorised role separate from the creator where required?
Rule
A different model receiving the same evidence, assumptions, errors, and framing may provide limited independence.
Review Trigger Classes
Independent review should be activated by one or more trigger classes.
Risk Trigger
The task may materially affect:
capital
legal exposure
compliance
privacy
security
client trust
public reputation
production systems
irreversible records
High-risk tasks may require independent review before any execution.
Failure Trigger
The same model or agent has failed to produce verified progress.
Failure includes:
repeating the same error
changing wording without correcting substance
ignoring supplied evidence
breaking an already working component
inventing unsupported facts
failing deterministic tests
failing to follow required structure
losing important context
claiming success without verification
Disagreement Trigger
Reviewers reach materially different conclusions.
Evidence Trigger
The original output relies on:
weak evidence
old evidence
dependent sources
missing primary evidence
unverified assumptions
conflicting sources
Evidence triggers may require new research rather than another opinion.
Confidence Trigger
The originating model or evaluator reports low confidence.
Novelty Trigger
The task involves an unfamiliar, emerging, complex, or poorly documented subject.
Authority Trigger
The work affects Canon, governance, permissions, production deployment, financial authority, or a client commitment.
Human Request Trigger
Martyn, HeadOffice, or another authorised operator explicitly requests independent review.
Review Modes
MWMS may use several review modes.
Mode One: Self-Check
The originating model checks:
format
internal consistency
obvious omissions
calculation transcription
instruction compliance
Self-check is useful but not independent.
Mode Two: Blind Independent Review
The reviewer receives:
the task
relevant evidence
governing rules
required output standard
The reviewer does not initially receive the originating model’s conclusion.
Blind review reduces anchoring and imitation.
Mode Three: Critique Review
The reviewer receives the original output and is instructed to identify:
errors
unsupported claims
missing evidence
assumptions
risks
contradictions
instruction failures
Mode Four: Alternative Solution Review
The reviewer independently creates another solution or recommendation.
The outputs are then compared.
Mode Five: Specialist Review
The work is reviewed by a model, tool, Brain, or human with the relevant specialist capability.
Mode Six: Deterministic Validation
The output is checked using:
calculations
schema validation
tests
database constraints
policy rules
source verification
code execution
known expected outputs
Mode Seven: Panel Review
Several defined reviewers independently examine different parts of the problem.
Mode Eight: Rescue Routing
A stalled or repeatedly failing task is transferred to a separate route with a controlled rescue package.
Mode Nine: Human Adjudication
An authorised human reviews unresolved evidence, risk, authority, or disagreement.
Review Selection Rule
The review mode must match the failure or risk.
More models are not automatically better.
The correct reviewer is the one with the capability and independence required to evaluate the actual weakness.
Review Role Separation
Multi-model review should assign defined roles rather than giving every model the same vague instruction.
Possible roles include:
Primary Producer
Creates the original output.
Independent Reviewer
Forms a separate assessment.
Challenge Reviewer
Searches for weaknesses, contradictions, and failure risks.
Evidence Reviewer
Checks whether sources support the claims made.
Source Independence Reviewer
Determines whether apparently separate sources rely on one underlying source.
Domain Specialist
Interprets evidence using subject-specific expertise.
Deterministic Validator
Runs calculations, tests, schema checks, or other objective validation.
Risk Reviewer
Examines safety, security, compliance, financial, privacy, and reputational consequences.
Synthesiser
Combines findings without hiding disagreement.
Adjudicator
Applies Canon, authority, evidence, and decision rules to determine the final outcome.
Rescue Reviewer
Takes over work that has failed under the original route.
Rule
A model must not silently perform producer, reviewer, evidence verifier, synthesiser, adjudicator, and final authority roles in one unobservable step for high-risk work.
Evidence And Interpretation Separation
Independent model review must distinguish between:
evidence retrieval
evidence verification
interpretation
recommendation
decision
Different models may produce different interpretations while relying on the same evidence.
This is perspective diversity.
It is not evidence independence.
Before accepting agreement between models, MWMS should determine:
whether each model used the same source set
whether sources share the same original evidence
whether each model independently inspected the sources
whether assumptions were shared
whether the task framing biased all reviewers
whether one model’s output anchored the others
Rule
Model consensus is not evidence consensus.
Independent Evidence Retrieval Standard
Where factual accuracy matters, at least one review path should independently search for or verify evidence.
The independent reviewer should be able to:
identify missing sources
find contradictory evidence
verify original sources
distinguish direct evidence from repeated reporting
check source freshness
check whether evidence supports the exact claim
A reviewer that only critiques the wording of the first answer has not independently verified the facts.
Blind Review Standard
Blind review should be used where anchoring risk is material.
The reviewer should initially receive:
the original task
governing framework
required evidence
decision constraints
known facts
The reviewer should not initially receive:
the first model’s conclusion
the first model’s confidence
the desired commercial result
the preferred answer
After the independent view is recorded, the reviewer may compare it with the original output.
Panel Review Standard
A model panel should not operate as an unstructured group conversation.
Each reviewer should first produce an independent record containing:
role
conclusion
evidence
assumptions
risks
confidence
limitations
missing information
recommended action
Only after independent records exist should synthesis begin.
The panel may then compare:
areas of agreement
areas of disagreement
different evidence used
different assumptions
different definitions
different risk tolerances
important omissions
minority findings
Rule
Reviewers must not be instructed to reach agreement.
Panel Discussion Risk
Allowing models to see each other’s conclusions too early may create:
anchoring
social imitation
premature consensus
suppression of minority evidence
repetition without verification
false confidence
For material decisions, initial review should remain independent.
Synthesis Standard
The synthesiser must not merely select the majority answer.
The synthesis should contain:
Original Question
Primary Conclusion
Supporting Evidence
Challenging Evidence
Areas Of Agreement
Areas Of Disagreement
Assumptions
Missing Information
Minority Finding
Residual Risk
Confidence
Recommended Decision
Required Conditions
Escalation Requirement
The synthesiser must distinguish:
consensus
partial agreement
unresolved disagreement
evidence dominance
authority decision
Rule
Synthesis is a structured comparison and decision aid.
It is not automatic truth creation.
Minority Evidence Preservation Rule
A minority reviewer may identify:
an overlooked risk
a conflicting source
an invalid assumption
a security weakness
a legal concern
a measurement error
an alternative explanation
Minority evidence must not be removed merely because most reviewers disagree.
The synthesiser should record:
the minority finding
the evidence supporting it
why it was accepted, rejected, or left unresolved
who authorised the final decision
Rule
A credible minority warning must remain traceable.
Assumption Exposure Standard
Each reviewer should identify material assumptions.
Assumptions may concern:
user intent
source reliability
data completeness
business rules
causation
future behaviour
technical capability
legal interpretation
risk tolerance
commercial impact
The review should classify assumptions as:
verified
reasonable but unverified
weak
contradicted
decision critical
Rule
Agreement created by shared assumptions must not be mistaken for independent confirmation.
Repeated Failure Detection
A task should enter failure review when the active model:
repeats the same incorrect action
fails the same verification test
ignores the same correction
continues using a disproven assumption
produces output without required evidence
changes presentation without improving the result
creates new damage while repairing old damage
claims completion without proof
Failure Counter Rule
The default MWMS failure counter is:
First Unverified Failure
Allow one controlled correction attempt.
Second Unverified Failure
Stop the originating route and initiate rescue review.
Immediate Rescue
Immediate rescue may occur after one attempt where the failure involves:
security risk
destructive action
privacy exposure
material financial risk
Canon violation
credential handling
production instability
fabricated evidence
false claim of completed execution
Rule
The same failed method must not be repeated indefinitely.
Verified Progress Standard
A correction attempt counts as progress only when evidence shows improvement.
Verified progress may include:
a failed test now passes
the unsupported claim is removed
the correct source is retrieved
the schema validates
the required format is restored
the contradiction is resolved
the user confirms the issue is fixed
a measurable failure count decreases
Longer output, different wording, or renewed confidence is not verified progress.
Rescue Routing Standard
When rescue is activated, the task must be transferred with a structured rescue package.
The rescue package should contain:
Task ID
Original Objective
Required Outcome
Current State
Work Already Completed
Originating Model Or AI Employee
Failure Count
Failed Approaches
Evidence Of Failure
Known Working Components
Known Broken Components
Governing Canon
Relevant Sources
Constraints
Prohibited Actions
Required Tests
Risk Classification
Deadline Or Urgency
Human Instructions
The rescue route must not be forced to rediscover the complete failure history.
Rescue Model Selection
The rescue model or route should be selected according to the failure.
Possible selection criteria include:
different model family
stronger reasoning capability
code-specialist capability
vision capability
large-context capability
source-retrieval capability
structured-data capability
domain expertise
deterministic tool access
lower susceptibility to the original failure mode
Rule
Availability does not create competence.
Rescue Reset Rule
The rescue route must receive enough context to continue safely but must not inherit every unverified assumption from the failed route.
The rescue reviewer should distinguish:
confirmed facts
operator instructions
working components
failed assumptions
unverified conclusions
prohibited actions
The rescue route may restart from the smallest validated checkpoint.
Rescue Outcome Classes
A rescue review may produce:
Recovered
The task was completed and verified.
Partially Recovered
A safe portion was completed, but unresolved work remains.
Reframed
The original task definition was found to be faulty and was corrected.
Escalated
The task requires specialist or human authority.
Blocked
Required access, evidence, capability, or permission is unavailable.
Abandoned
Further work is not justified by value, risk, or feasibility.
Specialist Routing Standard
Tasks involving specialised input should be routed to systems capable of handling that input.
Examples include:
images to a vision-capable reviewer
audio to transcription and audio-capable review
video to visual and temporal analysis
code to a code-capable reviewer plus deterministic tests
large document sets to an external retrieval engine
structured data to calculation or data-analysis tools
live web evidence to governed research tools
database changes to schema validation and data-integrity checks
legal questions to approved legal review
medical questions to approved medical evidence review
financial decisions to finance authority and validated calculations
A model must not pretend to have inspected a modality, source, system, file, or record it could not access.
Deterministic Verification Rule
Where objective validation is available, deterministic evidence should take priority over model opinion.
Examples include:
calculation result
test result
schema validation
database constraint
file comparison
checksum
policy rule
permission record
published source
deployment status
A majority of models must not overrule a failed deterministic test without proving that the test itself is invalid.
Consensus Mode
Consensus mode may be used where several independent perspectives are beneficial.
Each reviewer must initially provide:
conclusion
evidence
assumptions
risks
confidence
required conditions
limitations
Reviewers must not be instructed to agree.
The purpose of consensus mode is to reveal:
shared conclusions
meaningful disagreements
missing evidence
alternative interpretations
hidden assumptions
risk trade-offs
Consensus does not automatically create authority.
Final authority remains with the applicable MWMS governance role.
Consensus Classification
The synthesis should classify the result as:
Strong Evidence Agreement
Reviewers agree and rely on strong, sufficiently independent evidence.
Interpretive Agreement
Reviewers agree on interpretation but rely on the same evidence.
Conditional Agreement
Reviewers agree only if stated assumptions or conditions hold.
Partial Agreement
Reviewers agree on some components but disagree materially on others.
Unresolved Disagreement
Reviewers disagree and evidence does not resolve the issue.
False Consensus Risk
Apparent agreement may arise from shared prompts, sources, assumptions, or training patterns.
Rule
The type of agreement must be recorded.
Disagreement Resolution
When reviewers disagree, MWMS must determine whether the disagreement concerns:
facts
source reliability
source independence
interpretation
definitions
risk tolerance
authority
incomplete evidence
task definition
execution preference
time horizon
The adjudicator must resolve disagreements in this order:
Verify governing Canon and authority.
Verify the exact task or claim.
Verify source evidence.
Check source independence.
Run deterministic checks where possible.
Identify assumptions and definitions.
Request specialist review.
Reduce the decision to the smallest unresolved question.
Run additional research where justified.
Escalate to the authorised human or HeadOffice role when required.
A majority vote must not override:
clear Canon
objective evidence
deterministic validation
authority boundaries
credible severe-risk evidence
Unresolved Disagreement Rule
Where disagreement cannot be resolved, the final record must state:
what remains unresolved
why it remains unresolved
which evidence supports each position
what risk exists if either position is wrong
whether execution is restricted
who accepted the residual uncertainty
No unresolved disagreement may be disguised as consensus.
Rescue Escalation Ladder
Level One: Originating Model Self-Check
Allowed only for low-risk correction and formatting issues.
Level Two: Independent Model Review
A separate model or reviewer evaluates the work.
Level Three: Alternative Method Or Model
The work is recreated or verified through a different reasoning or tool path.
Level Four: Specialist Review
A specialist model, Brain, tool, or qualified human reviews the issue.
Level Five: HeadOffice Or Human Authority
HeadOffice, Martyn, or another authorised decision-maker determines the final action.
Higher-risk tasks may begin at a higher escalation level.
Risk Path Rules
The following work must not rely on single-model approval:
Canon promotion
financial commitments above approved limits
legal or compliance interpretations
security permission changes
destructive database operations
credential handling changes
production deployment with material external impact
high-volume customer communications
irreversible record deletion
system-wide architecture changes
suspension or replacement of a Brain or AI Employee
decisions where a reviewer identifies credible severe harm
public factual claims with material legal or reputational exposure
high-impact client recommendations based on conflicting evidence
Model Review And Evidence Review Separation
MWMS must distinguish two different review questions.
Model Review Question
Did the model reason and respond appropriately from the information available?
Evidence Review Question
Was the information available accurate, sufficient, current, independent, and relevant?
A strong model may reason correctly from false or incomplete evidence.
A weak model may reach a correct conclusion accidentally.
Rule
Both reasoning quality and evidence quality must be evaluated where factual reliability matters.
Fallback Rules
When the preferred reviewer is unavailable, MWMS must:
identify the missing capability
select the closest approved alternative
disclose the reduced review strength
increase deterministic validation
restrict execution where review remains insufficient
escalate when the risk exceeds the fallback reviewer’s authority
Availability does not create competence.
Review Evidence Record
Every formal independent review must preserve:
Review ID
Task ID
Originating Model Or AI Employee
Reviewer
Reviewer Role
Model Or Reasoning Route Used
Prompt Or Instruction Version
Date And Time
Risk Classification
Trigger Class
Review Mode
Evidence Reviewed
Sources Independently Retrieved
Assumptions Identified
Conflicts Found
Minority Findings
Review Outcome
Confidence
Required Conditions
Escalation Decision
Final Authority
Execution Status
Rescue records must additionally preserve:
Failure Count
Failed Approaches
Known Working Components
Known Broken Components
Rescue Package Version
Rescue Route
Rescue Outcome
Unresolved Work
Review Outcome Classes
Approved
The output meets the required standard.
Approved With Conditions
The output may proceed only if stated conditions are satisfied.
Revision Required
Correctable issues exist.
Additional Evidence Required
The conclusion cannot be accepted without more evidence.
Specialist Review Required
The current reviewers lack necessary capability or authority.
Escalation Required
Human or HeadOffice authority is required.
Rejected
The output is materially unreliable, unsafe, or non-compliant.
Blocked
Required access, evidence, permission, or capability is unavailable.
Execution Gate
Independent review does not automatically authorise execution.
Before execution, the system must confirm:
review outcome permits execution
required conditions are satisfied
authority is correct
risk is within approved limits
deterministic checks have passed
required human approval exists
review evidence is stored
No Review Bypass Rule
Review may be bypassed only where the bypass is:
explicitly authorised
time-bounded
risk-assessed
recorded
reviewable by SIT Brain
approved by the applicable authority
An emergency does not remove accountability.
It changes the required record and escalation speed.
Human Authority Rule
AI models may:
produce
review
challenge
compare
synthesise
recommend
They may not create final authority where authority has not been granted.
Human or HeadOffice authority remains required for:
Canon changes
major capital decisions
material legal or compliance judgments
irreversible destructive actions
high-risk public commitments
suspension or replacement of governed systems
acceptance of significant unresolved risk
Relationship To Other MWMS Frameworks
This framework should operate with:
MWMS AI Agent Failure Handling And Escalation Protocol
MWMS AI Output Validation Standard
MWMS AI Employee Evaluation Scorecard Standard
MWMS Deep Search Quality And Observability Framework
MWMS Source Visibility And Evidence Display Standard
MWMS Research Synthesis Documentation And Distribution Framework
MWMS External Knowledge Engine And Reasoning Agent Separation Framework
MWMS AI Multi Agent Role Design Framework
MWMS AI Agent Orchestration Framework
MWMS AI Tool Permission And Access Framework
MWMS AI Observability Metadata Standard
MWMS AI Guardrail And Preflight Check Standard
MWMS Plus Drift Control And Human Challenge Protocol
MWMS Brain Authority Matrix
MWMS Decision Record System
This framework does not replace those pages.
It governs when and how independent reasoning, model challenge, rescue routing, synthesis, and disagreement escalation occur.
Observability Requirements
Formal review workflows should record:
models used
model versions where available
review roles
prompt versions
source sets
independent source searches
tools used
assumptions
review findings
agreement classification
disagreement classification
minority findings
confidence
cost
latency
retry count
rescue activation
human intervention
final decision
execution status
Rule
A multi-model review that cannot show who reviewed what and why is not governed review.
Quality Measures
SIT Brain and HeadOffice may monitor:
percentage of high-risk outputs independently reviewed
percentage of reviews using blind initial assessment
percentage of factual reviews using independent evidence retrieval
repeat-failure rate
rescue activation rate
rescue success rate
average attempts before rescue
review disagreement rate
minority finding acceptance rate
false consensus incidents
review-to-execution time
cost per independent review
percentage of reviews requiring human adjudication
percentage of model agreements supported by independent evidence
The goal is not to maximise model usage.
The goal is to apply the smallest reliable review architecture that protects decision quality.
Failure Conditions
The independent review process should be treated as failed where:
the reviewer copies the original conclusion without assessment
reviewers are instructed to agree
all reviewers rely on the same unverified evidence
model voting replaces evidence review
minority evidence is deleted
assumptions remain hidden
the wrong specialist capability is used
failed deterministic tests are ignored
the rescue model receives incomplete failure history
the failed model continues after the rescue threshold
the review record is missing
human authority is bypassed
execution occurs before conditions are satisfied
Drift Protection
This framework protects MWMS from:
self-review being treated as independence
repeated reasoning loops
model-confidence theatre
model-consensus theatre
majority voting without evidence
shared-source agreement being treated as independent confirmation
specialist tasks being reviewed by general systems
minority evidence suppression
premature consensus
uncontrolled rescue attempts
review without authority
execution without validation
Drift Signals
Watch for:
“Ask the same model one more time.”
“Open a new chat and try the same prompt.”
“Three models agreed, so it must be right.”
“The majority answer wins.”
“They all used the same source, but that is fine.”
“The reviewer does not need to search independently.”
“Do not show the reviewer the conflicting evidence.”
“The minority view will confuse the user.”
“The models should debate until they agree.”
“The rescue model can work out what failed.”
“The code looks correct, so tests are unnecessary.”
“The reviewer is available, so it must be qualified.”
“We can approve it without storing the review.”
“The AI panel approved it.”
Rule
When these drift signals appear, return to independence, evidence separation, specialist capability, deterministic validation, minority evidence preservation, and authority.
Architectural Intent
MWMS is designed to operate as a governed intelligence system rather than a collection of unmonitored AI outputs.
Reliable intelligence requires:
separation between creation and review
controlled escalation after failure
model diversity where it adds value
evidence diversity where factual verification matters
independent initial assessment where anchoring risk exists
deterministic verification where available
specialist routing where required
preservation of minority evidence
structured synthesis
clear disagreement records
preservation of review evidence
clear authority boundaries
human control over high-impact decisions
This framework ensures that MWMS does not confuse:
repetition with persistence
confidence with accuracy
agreement with evidence
model diversity with source independence
synthesis with truth
AI consensus with authority
Final Rule
When an AI model produces important work, MWMS must ask whether that work requires independent review.
When the same model fails twice without verified progress, MWMS must stop the loop and route the task for rescue.
When several models are used, each must have a defined role and should form an independent assessment before synthesis where material anchoring risk exists.
When models agree, MWMS must determine whether the agreement is supported by independent evidence or merely shared sources and assumptions.
When reviewers disagree, MWMS must preserve the disagreement, inspect the evidence, expose assumptions, and route the smallest unresolved issue for specialist review, additional research, or human authority.
When risk remains unresolved, execution must not proceed.
Source Absorption Basis
The original framework was created from the AI Automations by Jack course block covering:
multi-model routing
independent review
repeated-failure rescue
specialist model assignment
risk-path escalation
structured consensus
The v1.1 update absorbs additional course material covering:
multi-model research panels
fact-checking workflows
independent research routes
evidence comparison
model perspective diversity
assumption exposure
agreement and disagreement synthesis
minority evidence preservation
source-backed adjudication
The source material reinforces that multiple models should be used to reveal different perspectives, assumptions, omissions, and risks.
It does not justify treating model voting as factual truth.
MWMS System Change Log
Version: v1.1
Date: 2026-06-21
Author: HeadOffice
Change
Updated the MWMS Independent Model Review And Rescue Routing Framework from v1.0 to v1.1 using the AI Automations by Jack multi-model research, fact-checking, browser copilot, panel discussion, evidence synthesis, and independent-review material.
Expanded the framework beyond repeated-failure rescue and general model disagreement to define a governed multi-model review architecture.
Added standards covering:
• independence dimensions
• blind independent review
• defined reviewer roles
• evidence and interpretation separation
• independent evidence retrieval
• structured panel review
• panel anchoring risk
• synthesis requirements
• minority evidence preservation
• assumption exposure
• agreement classification
• unresolved disagreement records
• model review versus evidence review
• source-supported adjudication
• review observability
Added explicit doctrine that:
• model diversity is not automatically evidence diversity
• model consensus is not evidence consensus
• majority voting must not replace evidence review
• reviewers should not be instructed to agree
• minority evidence must remain traceable
• synthesis must preserve disagreement and residual risk
• AI consensus does not create authority
Expanded the review evidence record to capture:
• reviewer role
• prompt or instruction version
• review mode
• independently retrieved sources
• minority findings
• confidence
• rescue package version
• unresolved work
Added quality measures for:
• blind review use
• independent evidence retrieval
• false consensus incidents
• minority finding acceptance
• evidence-supported model agreement
Change Impact Declaration
This update strengthens SIT Brain governance without changing final authority.
SIT Brain remains the owning Brain for independent model review and rescue routing.
HeadOffice and authorised humans retain final authority for high-impact decisions.
The update does not require multiple models for every task.
It requires the smallest review architecture capable of addressing the actual risk, failure, evidence weakness, or capability gap.
The update does not allow model panels to override:
• Canon
• deterministic evidence
• source evidence
• Brain authority
• authorised human decisions
Pages Created
• None
Pages Updated
• MWMS Independent Model Review And Rescue Routing Framework updated from v1.0 to v1.1
Pages Deprecated
• None
Standalone Pages Not Created
The following standalone pages were not created because their durable intelligence is governed within this updated framework:
• MWMS Multi Model Panel Discussion Framework
• MWMS Model Consensus Framework
• MWMS Multi Model Fact Checking Framework
• MWMS Independent Research Panel Framework
• MWMS AI Debate Framework
• MWMS Minority Model Evidence Framework
• MWMS Blind Model Review Framework
• MWMS Model Assumption Comparison Framework
Registries Requiring Update
• SIT Brain Page Registry
• MWMS Canon Index
• MCR Page Registry
• MCR Copy Map where the framework version is recorded
• MWMS Course Absorption Decision Registry
Canon Version Update Required
No immediate SIT Brain Canon version change is required unless the current Canon directly records framework versions or contains review rules that conflict with v1.1.
The model-consensus, evidence-independence, minority-evidence, and blind-review controls should be included during the next scheduled SIT Brain Canon alignment review.
Change Log Entry Required
Yes.
The v1.1 update must be recorded in:
• MWMS System Change Log
• SIT Brain Page Registry change history where applicable
• MCR Page Registry change history where applicable
• MWMS Course Absorption Decision Registry
Strategic Absorption Result
The AI Automations by Jack material concerning multi-model research, panel discussions, fact-checking, alternative interpretations, and synthesis has been absorbed into the existing MWMS Independent Model Review And Rescue Routing Framework.
The absorption preserves the useful multi-model architecture while rejecting:
• model voting as truth
• consensus theatre
• reviewers being instructed to agree
• multiple models relying on one evidence chain
• early reviewer anchoring
• suppression of minority findings
• unstructured model debate
• synthesis that hides disagreement
The resulting v1.1 framework establishes that MWMS independent review must be:
• role-defined
• sufficiently independent
• evidence-aware
• source-aware
• assumption-aware
• anchoring-aware
• disagreement-preserving
• specialist-routed
• deterministic where possible
• observable
• authority-controlled
END OF FULL FILE OUTPUT