MWMS Research Planning And Query Rewriting Standard

System: MWMS
Document Type: Standard
Status: Draft For MCR
Authority: HeadOffice
Applies To: Research Brain, Affiliate Brain, Ads Brain, Content Brain, HeadOffice Intelligence, Data Brain, Experimentation Brain, AI Employees, Deep Search Workflows, Future Client Facing AI Systems
Primary Location: MCR
Future Operational Destination: mwmsbrain.site, mwmsheadofficebrain.site, Research Brain, AI Employee Dashboards, Future Deep Search Systems
Parent Page: HeadOffice
Source Of Truth: MCR
Related Frameworks: MWMS Agent Loop Control Framework, MWMS Next Action Picker Standard, MWMS Agent Loop Context Schema, MWMS AI Work Session Persistence Standard, MWMS Deep Search Quality And Observability Framework, MWMS AI Employee Evaluation Scorecard Standard, MWMS AI Observability Metadata Standard
Course Source: Matt Pocock AIhero Build DeepSearch In TypeScript
Absorption Status: Approved For Integration

Purpose

The purpose of this standard is to define how MWMS plans research before searching.

Research quality does not begin when the AI Employee searches.

Research quality begins when the system understands the task, identifies what must be known, chooses what assumptions need checking, defines source requirements, and creates the correct search path.

This standard defines how MWMS AI Employees should transform a user request, business task, or intelligence need into a structured research plan and query set before external search begins.

The goal is to stop AI Employees from searching randomly and instead make them search with purpose.

Scope

This standard applies to any MWMS workflow where an AI Employee needs to gather, verify, compare, or update information.

This includes:

Deep Search research
Research Brain investigations
Affiliate Brain offer evaluation
Ads Brain compliance checks
Content Brain topic and market research
HeadOffice newsletter intelligence
Data Brain validation
Experimentation Brain test analysis
Finance Brain market or cost checks
Product or vendor research
Tool evaluation
AI platform monitoring
competitor research
source freshness checks
future client-facing AI research workflows

This standard does not define exact search tool implementation, scraper code, or model provider setup.

It defines the MWMS planning and query rewriting standard.

Core Rule

The core rule is:

Search queries are not just search queries. They are the execution path of the research plan.

An AI Employee should not begin important research by randomly searching a single broad phrase.

It should first create a research plan.

Then it should create focused queries that execute that plan.

Definition Of Research Planning

Research Planning is the process of deciding what must be discovered, checked, verified, or ruled out before an answer can be trusted.

A research plan should identify:

the real question
the decision being supported
known context
missing information
assumptions to verify
freshness requirements
jurisdiction or market requirements
source preferences
risk areas
query sequence
expected evidence
stopping criteria

Research planning happens before search execution.

Definition Of Query Rewriting

Query Rewriting is the process of transforming a user request or business task into better search queries.

A poor query often repeats the user’s wording.

A strong query is designed to find the evidence needed to answer the task.

Example:

User request:

Is this affiliate offer worth testing?

Weak query:

affiliate offer worth testing

Better research plan:

Find official vendor/product page.
Find affiliate network terms and payout information.
Check customer complaints or refund concerns.
Check competitor landscape.
Check compliance risk around claims.
Check current market demand.

Better queries:

official product name vendor affiliate terms
product name ClickBank affiliate payout refund rate
product name customer complaints reviews
product name competitors alternative products
product name ad compliance claims health income policy
product niche current demand USA 2026

Why MWMS Needs This Standard

Without research planning, AI Employees may:

search too broadly
search the wrong thing
rely on weak sources
miss official sources
miss current information
ignore jurisdiction
ignore compliance risk
answer from stale model memory
gather irrelevant evidence
waste cost on poor searches
produce confident but weak conclusions

With research planning, MWMS gains:

better source quality
better evidence coverage
better freshness awareness
better compliance safety
better cost control
better final answers
better eval results
stronger HeadOffice confidence

Research Planner Role

MWMS should treat research planning as a dedicated AI Employee action.

The Research Planner does not produce the final answer.

The Research Planner decides how research should be performed.

Its job is to define:

what needs to be known
what sources should be preferred
what queries should be run
what order the queries should follow
what risks must be checked
what freshness level is required
what evidence will be enough to answer

The Research Planner prepares the path.

Other components execute the path.

Relationship To Next Action Picker

The Next Action Picker decides what the AI Employee should do next.

The Research Planner decides what the research should look for.

These are related but not identical.

Component	Main Job
Next Action Picker	Choose the next approved action
Research Planner	Define the research plan and query set
Search Pipeline	Execute the queries
Source Inspector	Open or crawl selected sources
Source Summariser	Compress evidence
Answer Generator	Produce final response
Evaluator	Judge output quality

The Next Action Picker may choose:

create_research_plan

Then the Research Planner creates the plan and query set.

Relationship To Agent Loop Control

This standard supports controlled agent loops.

In a controlled loop, research planning may happen:

before the first search
after a failed search
after evidence gaps are detected
when the user changes the goal
when sources conflict
when more current information is required
when the AI is about to answer but evidence is insufficient

The Agent Loop should not keep searching without revisiting the plan.

Relationship To Search Scrape Summarise

This standard sits before the Search Scrape Summarise evidence pipeline.

The normal flow is:

Understand task
Create research plan
Generate query set
Search
Inspect or scrape sources
Summarise evidence
Evaluate sufficiency
Answer or continue

The Research Planner defines what the pipeline should look for.

Research Planning Workflow

A standard MWMS research planning workflow should include:

Parse the user request or task
Identify the real decision being supported
Identify known context
Identify missing information
Identify assumptions to verify
Identify freshness and date needs
Identify jurisdiction, location, or market needs
Identify source type preferences
Identify risk areas
Create research questions
Convert research questions into search queries
Prioritise query order
Define expected evidence
Define stopping criteria
Pass query set to search workflow

Step 1: Parse The Request

The AI Employee should first identify what the user or system is really asking.

It should extract:

topic
task type
desired outcome
decision needed
implied context
constraints
risk level
required output

Example:

User request:

Is this course worth it for MWMS?

Parsed meaning:

Topic: course evaluation
Task type: business relevance research
Decision needed: buy, skip, park, or absorb
Required output: verdict and MWMS value analysis
Risk: low to medium cost/time risk
Required evidence: sales page, curriculum, overlap with existing MWMS knowledge

Step 2: Identify The Decision Being Supported

Research should support a decision.

Examples:

User Request	Decision Being Supported
Is this offer worth testing?	Proceed, reject, or park
Is this tool useful for MWMS?	Buy, trial, ignore, or monitor
Can we use this claim in ads?	Approve, modify, reject, or review
What does this newsletter mean for us?	Act, test, monitor, ignore
Is this source reliable?	Use, verify, avoid, or archive

If there is no decision, the research plan should still define the intended use of the answer.

Step 3: Identify Known Context

The planner should identify what MWMS already knows.

Examples:

existing MCR pages
prior course absorptions
known affiliate rules
current project context
user preferences
previous test results
existing Brain standards
known product rules
known compliance restrictions
current MWMS tools

This prevents duplicate research.

Step 4: Identify Missing Information

The planner should identify the information needed before answering.

Examples:

official product details
current price
affiliate payout
refund rate
policy rule
release date
source publication date
vendor credibility
customer complaints
competitor evidence
current market demand
jurisdiction-specific rule

The search queries should be designed to fill these gaps.

Step 5: Identify Assumptions To Verify

The AI Employee should list assumptions that may be unsafe.

Examples:

the product is still available
the policy has not changed
the source is current
the vendor is reputable
the affiliate offer is active
the payout is accurate
the course teaches something new
the tool integrates with MWMS
the claim is compliant
the market demand still exists

Unsafe assumptions should become research questions.

Step 6: Identify Freshness Requirements

The planner must decide whether current information matters.

Freshness levels:

Level	Meaning
Historical	Older sources acceptable
Evergreen	Recency not critical
Recent	Prefer current-year or recent sources
Current	Must verify latest available information
Live/Highly Current	Needs current web/search data and date awareness

Current information is required for:

pricing
affiliate offers
policies
laws
compliance
product availability
platform rules
AI tool capabilities
market trends
current events
ad platform behaviour
software documentation

If freshness matters, the query set should include date or current-context terms where useful.

Step 7: Identify Location And Market Context

Some research depends on location, country, currency, language, or jurisdiction.

The planner should identify:

target country
user country
campaign market
legal jurisdiction
currency
timezone
language
platform region
shipping region
compliance region

Examples:

Google Ads policy may vary by country or product class.
Affiliate offers may be USA-only.
Finance or insurance offers may depend on state or country.
Product availability may differ by region.
Compliance may differ across Australia, USA, EU, UK, Canada, or Asia-Pacific.

Rule

If location, market, or jurisdiction changes the answer, the research plan must include it.

Step 8: Identify Source Preferences

The planner should define what source types are preferred.

Source preference examples:

Task	Preferred Sources
Platform policy	official platform documentation
Product details	official vendor page
Affiliate offer	affiliate network page, vendor affiliate page
Compliance	official policy, legal/regulatory source
Market trend	credible industry reports, news, search data
Tool feature	official docs, changelog, vendor page
Complaints	reviews, forums, Reddit as signal only
Competitors	official competitor pages, comparison pages

Rule

Official sources should be preferred when the task depends on factual or policy accuracy.

Commercial and user-generated sources should be treated as signals, not proof.

Step 9: Identify Risk Areas

The planner should identify research risk areas.

Examples:

compliance risk
financial risk
platform policy risk
client-facing risk
outdated source risk
biased source risk
weak evidence risk
hallucination risk
high-cost research risk
privacy risk
overclaiming risk

Risk areas should influence query design.

For example, if compliance risk exists, one query should target official policy or regulatory sources.

Step 10: Create Research Questions

Before writing search queries, the planner should create research questions.

Example for affiliate offer evaluation:

What is the official product and vendor?
What are the current affiliate terms?
What claims does the sales page make?
Are there refund, complaint, or scam concerns?
What competitors exist?
Is the niche suitable for Google/YouTube traffic?
Are there compliance risks?
Is the offer still active and current?

Research questions are easier to evaluate than raw queries.

Step 11: Convert Research Questions Into Search Queries

Each research question should become one or more search queries.

Good queries should be:

specific
evidence-seeking
aligned with source preference
not overly broad
not overly narrow
freshness-aware when needed
jurisdiction-aware when needed
designed to find verifiable sources

Example:

Research question:

What are the current Google Ads rules for this type of claim?

Query:

official Google Ads policy health claims supplements misleading claims 2026

Step 12: Prioritise Query Order

Not all queries are equal.

Recommended order:

Official source query
Current/freshness query
Evidence or claim verification query
Risk or complaint query
Competitor or market context query
Supplemental expert/source query

The system should not start with low-trust sources when official sources are likely available.

Step 13: Define Expected Evidence

Each query should have an expected evidence target.

Examples:

Query Type	Expected Evidence
official vendor query	product page, vendor statement
policy query	official policy page
complaint query	customer complaints or review signals
competitor query	alternative products and market positioning
pricing query	current price or pricing page
affiliate query	payout, network, terms, restrictions
trend query	recent market or demand evidence

This helps the evaluator judge whether the query succeeded.

Step 14: Define Stopping Criteria

The research plan should define when enough research has been done.

Stopping criteria may include:

official source found
enough independent sources found
current policy confirmed
source conflict detected and escalated
required evidence missing after search attempts
max query count reached
max source count reached
cost or time limit reached
confidence threshold reached
human review required

Without stopping criteria, research can drift or become too expensive.

Standard Research Plan Output

The Research Planner should return structured output.

Recommended format:

{
  "research_goal": "",
  "decision_supported": "",
  "known_context": [],
  "missing_information": [],
  "assumptions_to_verify": [],
  "freshness_requirement": "",
  "location_or_market_context": "",
  "source_preferences": [],
  "risk_areas": [],
  "research_questions": [],
  "query_plan": [
    {
      "query": "",
      "purpose": "",
      "preferred_source_type": "",
      "freshness_need": "",
      "expected_evidence": "",
      "priority": 1
    }
  ],
  "stopping_criteria": [],
  "human_review_triggers": []
}

This is a conceptual schema only.

Implementation may vary.

Minimum Research Plan

For early MWMS implementation, the minimum research plan should include:

research goal
decision supported
missing information
freshness requirement
source preferences
3 to 5 queries
expected evidence
stopping criteria

This is enough to avoid random searching.

Query Count Rule

Most research plans should begin with 3 to 5 focused queries.

One query is often too shallow.

Too many queries can waste cost and create noise.

The correct number depends on:

task complexity
risk level
freshness requirement
source availability
decision importance
budget and time limits

High-risk or complex research may need more queries, but only with clear reason.

Query Quality Rules

Good MWMS queries should:

target evidence, not vibes
prefer official sources where needed
include product, vendor, platform, or policy names where known
include jurisdiction or market where relevant
include current-year terms only when freshness matters
separate different research purposes into different queries
avoid overly broad generic phrases
avoid asking the search engine to “decide”
avoid emotional or biased wording

Poor query:

best product scam or legit

Better query set:

product name official vendor page
product name affiliate terms payout
product name customer complaints refund reviews
product name competitors alternatives
product niche Google Ads policy claims

Query Rewriting Modes

MWMS should support different query rewriting modes depending on task.

Mode 1: Clarifying Query Rewrite

Used when the user’s request is vague.

Goal:

make the query more specific
identify likely intent
avoid wrong assumptions

Mode 2: Current Evidence Rewrite

Used when information may have changed.

Goal:

find current sources
include recency terms where useful
prioritise official/current pages

Mode 3: Source Type Rewrite

Used when a specific kind of source is needed.

Goal:

find official docs, policy pages, vendor pages, reviews, competitors, or legal sources

Mode 4: Risk Check Rewrite

Used when the task has compliance, financial, or reputational risk.

Goal:

find warnings, policy limits, complaints, restrictions, or conflicting evidence

Mode 5: Competitive Research Rewrite

Used for market and offer analysis.

Goal:

find competitors, alternatives, market signals, and positioning

Mode 6: Internal Plus External Rewrite

Used when both MWMS records and web research matter.

Goal:

define what to search internally and externally

Agentic Dial Rule

This block adds an important principle:

Not everything should be agentic.

MWMS should decide whether a research task should be handled as:

Mode	Meaning
Deterministic Workflow	Fixed path, little AI choice
Assisted Workflow	AI helps inside controlled steps
Controlled Agent	AI chooses next approved action
Agentic Workflow	AI has broader control under limits
Autonomous Agent	Only for mature, evaluated, low-risk or approved systems

Research planning helps decide the right mode.

High-risk tasks should use more controlled workflow.

Low-risk exploratory tasks may allow more agentic behaviour.

Agent Vs Workflow Rule

The AI should not make every decision if a workflow rule can make it better.

Examples:

Situation	Better Approach
Search always needs source inspection	combine search and inspect workflow
Official source is required	deterministic official-source-first rule
Compliance risk exists	mandatory policy check
Sources conflict	mandatory review/escalation
Offer evaluation always needs vendor page	fixed vendor check step
Newsletter always needs routing	fixed extraction plus routing proposal

The rule:

Use AI judgement where judgement is useful. Use workflow rules where predictability is better.

Action Consolidation Rule

If two actions almost always belong together, MWMS should consider consolidating them into one workflow step.

Example:

Instead of:

Search
Ask AI whether to inspect source
Inspect source

Use:

Search and inspect top eligible sources

This reduces:

unnecessary AI decisions
cost
latency
failure points
agent drift

Possible consolidated actions:

search plus inspect
inspect plus summarise
source summary plus trust rating
newsletter extract plus routing proposal
offer page inspect plus claim extraction
ad review plus compliance risk check

Search Scrape Summarise Principle

The Search Scrape Summarise pattern is:

Search broadly enough to find candidate sources.
Inspect or scrape selected sources.
Summarise each source into compact evidence.
Feed compact evidence into final reasoning.
Store source summaries for reuse.

This helps MWMS use more sources without overloading the model context window.

Evidence Compression Rule

Source summarisation is evidence compression.

The summariser should compress long source content into useful research notes without inventing missing facts.

A source summary should preserve:

source title
source URL
query that found it
retrieved date
source date if visible
key facts
key claims
relevant figures
limitations
what the source does not answer
trust rating
freshness rating
relevance to research goal
whether it supports final conclusion

Rule

Summaries must compress evidence without inventing evidence.

Source Summary Record Concept

This standard points toward a future source summary record.

Recommended fields:

Field	Description
source_id	Source record ID
research_session_id	Parent research session
query_used	Query that found source
source_url	URL
source_title	Title
retrieved_at	Retrieval time
publication_date	Source date if visible
source_type	Official, expert, commercial, user-generated, unknown
trust_rating	Low, medium, high
freshness_rating	Outdated, acceptable, current, unknown
evidence_summary	Relevant summary
key_claims	Claims from source
limitations	What source does not prove
used_in_final_output	Yes or no

This may later become part of a dedicated MWMS Search Scrape Summarise Evidence Pipeline Standard.

Query Rewriter Evaluation

Research plans and query rewrites should be evaluated.

Evaluation questions:

Did the plan identify the real decision?
Did the plan identify missing information?
Did it include freshness requirements?
Did it include market or jurisdiction where needed?
Did it prefer the right source types?
Did it include risk checks?
Were the queries specific?
Were the queries diverse enough?
Did the query set avoid duplication?
Did the queries lead to useful sources?
Did the plan support a better final answer?

Deterministic Checks

Some research plan outputs can be checked deterministically.

Examples:

research_goal exists
decision_supported exists
freshness_requirement exists
query_plan has at least one query
each query has a purpose
each query has expected evidence
source preferences are included
risk areas are included where needed
stopping criteria exist

LLM Judge Checks

Some research planning quality requires judgement.

Examples:

Did the plan understand the real task?
Were the chosen queries good?
Did the plan miss obvious evidence needs?
Did it over-search?
Did it under-search?
Did it ignore compliance risk?
Did it ignore current information needs?
Did it use the right source priorities?

This should connect to the MWMS AI Employee Evaluation Scorecard Standard.

Failure Conditions

A research plan should be marked failed or weak if:

it repeats the user wording as the only query
it does not identify the decision being supported
it ignores freshness when freshness matters
it ignores jurisdiction when jurisdiction matters
it does not prefer official sources when needed
it misses obvious risk areas
it creates duplicate or vague queries
it creates biased queries
it lacks stopping criteria
it cannot explain what evidence is expected
it sends the search workflow in the wrong direction

Human Review Triggers

Human review should be triggered when:

the research task has compliance risk
the research affects budget decisions
the research affects campaign launch
the research affects client-facing output
source evidence conflicts
official sources cannot be found
freshness cannot be confirmed
the plan requires assumptions that may be unsafe
the AI Employee is operating outside its approved scope

Relationship To Research Brain

Research Brain should eventually use this standard as an operational layer.

Research Brain should not merely “search.”

It should:

plan research
generate queries
inspect sources
summarise evidence
evaluate source quality
produce decision-ready findings

This standard strengthens Research Brain’s planning function.

Relationship To Affiliate Brain

Affiliate Brain should use research planning before evaluating offers.

Offer research plans should usually include:

official product/vendor source
affiliate network terms
payout or commission data
refund or complaint signals
competitor landscape
traffic platform fit
compliance and claim risk
current market demand

This prevents weak offer decisions.

Relationship To Ads Brain

Ads Brain should use research planning before compliance or campaign recommendations.

Ad-related research plans should usually include:

platform policy source
claim type being reviewed
target jurisdiction
product category
compliance risk
allowed and prohibited wording
examples or enforcement signals where available

This prevents risky ad decisions.

Relationship To Content Brain

Content Brain should use research planning before generating strategy or content.

Content research plans may include:

search intent
target audience
competing content
source credibility
freshness needs
SEO angle
monetisation angle
compliance risk
content gap

This improves content quality and authority.

Relationship To HeadOffice Intelligence

HeadOffice should use this standard when external signals require investigation.

Newsletter or market signals should not automatically become actions.

They may need research planning first.

HeadOffice research plans may include:

signal verification
business relevance
Brain affected
risk level
source freshness
market impact
action potential
monitoring need

Relationship To Geolocation And Jurisdiction

Research planning must include location when location changes the answer.

Examples:

Australia versus USA compliance
Victoria versus other Australian states
EU privacy rules
UK advertising rules
US state-specific finance or insurance offers
local product availability
time-zone dependent events
currency-sensitive pricing

This supports MWMS global compliance coverage.

Relationship To Work Session Persistence

Research plans should be stored inside the AI work session where useful.

A persistent work session may include:

original request
research plan
query set
sources found
source summaries
final answer
review notes
Kaizen note

This makes research resumable and reusable.

Relationship To Agent Loop Context

The research plan should be part of the Agent Loop Context.

Useful fields:

research_goal
missing_information
assumptions_to_verify
query_plan
source_preferences
risk_areas
stopping_criteria
evidence_sufficiency

The Next Action Picker can then use the plan to decide whether to search, inspect, summarise, answer, or escalate.

Minimum Starting Implementation

MWMS does not need a complex query rewriter immediately.

Minimum implementation:

{
  "research_goal": "",
  "decision_supported": "",
  "missing_information": [],
  "freshness_requirement": "",
  "source_preferences": [],
  "risk_areas": [],
  "queries": [
    {
      "query": "",
      "purpose": "",
      "expected_evidence": ""
    }
  ],
  "stopping_criteria": []
}

This is enough to improve research quality quickly.

Future Enhancements

Future enhancements may include:

MWMS Search Scrape Summarise Evidence Pipeline Standard
MWMS Deep Search Source Record Standard
MWMS Source Summary Record Schema
MWMS Research Plan Evaluation Dataset
MWMS Research Planner AI Employee Role Card
MWMS Research Query Quality Scorecard
MWMS Geolocation And Jurisdiction Context Standard
MWMS Agentic Dial And Workflow Control Standard

These should only be created when course material or implementation need justifies them.

Drift Protection

This standard prevents the following drift:

random searching
one-query research
vague query generation
ignoring official sources
ignoring current information needs
ignoring jurisdiction
ignoring compliance risk
using search snippets as proof
overloading the model with raw source content
treating query rewriting as wording cleanup only
letting AI answer without a research plan
making every workflow too agentic
wasting cost on unnecessary AI decisions

Research quality starts before search.

If the plan is weak, the evidence will be weak.

If the evidence is weak, the final answer cannot be trusted.

Architectural Intent

The architectural intent of this standard is to make MWMS research intentional.

MWMS is not building an AI system that simply searches the web and answers.

MWMS is building a governed intelligence system that plans research, gathers evidence, compresses sources, evaluates quality, and produces decision-ready outputs.

Research planning is the first control point in that chain.

A strong research plan creates better queries.

Better queries create better sources.

Better sources create better evidence.

Better evidence creates better decisions.

Better decisions make MWMS stronger.

Change Log

v1.0 Initial Draft

Created the MWMS Research Planning And Query Rewriting Standard based on absorbed insights from Matt Pocock AIhero Build DeepSearch In TypeScript.

Integrated principles from course blocks covering:

geolocation and contextual awareness
agents versus workflows
reducing unnecessary AI decision points
search plus crawl consolidation
search, scrape, and summarise evidence flow
query rewriting as research planning
source preference planning
freshness and jurisdiction requirements
research questions before search queries
evidence compression
action consolidation
agentic dial control
evaluation of query quality

Established this standard as the MWMS governance page for turning research tasks into structured research plans and focused query sets before search execution.