π Devlog #3 β Building the Material World
Structuring planning judgement, one consideration at a time.
π‘ Overview
One of the foundational challenges in automating planning support is modelling how human officers actually reason about development impacts. Not just what the policies say, but how competing issues are weighed, judged, and sometimes traded off.
In this update, Iβve tackled that problem head-on by implementing a full system for representing material considerations β the legally recognised planning factors that shape decisions. These arenβt just fields in a form; theyβre the active ingredients of discretion.
To support this, Iβve created a complete library of Material Consideration Templates β structured, modular JSON files that encode the domain logic, assessment questions, typical impacts, and mitigation strategies for each key planning issue.
This gives The Plannerβs Assistant a reusable vocabulary of planning judgement. It lets the system:
- Detect relevant topics in application documents
- Trigger the right reasoning logic or specialist agent
- Generate officer-style commentary
- Feed into the overall planning balance matrix
- Remain fully overrideable and traceable
Itβs a structured simulation of how officers think β not a replacement, but a mirror.
π Structured Schema per Consideration
Each module defines:
- π A plain-English explanation of the issue
- π Trigger keywords for LLM document scan
- β Sub-questions that simulate how officers assess relevance, severity, and mitigation
- π A schema hint: how to extract structured data from application documents
- πΊ Typical positive/negative impacts and common mitigation measures
- π€ Agent routing hint (e.g.
ClimateActionAgent,HeritageAndArchaeologyImpactAgent)
This enables:
- Targeted AI prompting
- Structured commentary generation
- Planning balance synthesis
- Fully explainable and overrideable output per impact category
Each consideration is treated as a modular reasoning unit. Think of it like a judgement card β activated when relevant, scored if triggered, weighted in the final recommendation.
π Full Coverage (May 2025)
π± Environment & Climate
- Air Quality
- Climate Change & Resilience
- Flood Risk & Drainage
- Energy & Carbon Reduction
- Water Resources & Quality
- Ecology & Biodiversity
- Ground Conditions & Land Stability
- Land Contamination & Remediation
- Noise & Vibration (Environmental)
π Design & Built Form
- Urban Design & Townscape
- Heritage Impact (Built & Archaeological)
- Trees, Hedgerows & Landscaping
- Public Realm & Community Infrastructure
- Neighbour Amenity
π Policy & Delivery
- Planning Policy Compliance & Balance
- Housing Delivery
- Development Viability
- Cumulative Impacts Assessment
- Utilities & Infrastructure Capacity
- Waste Management (Construction & Operational)
- Sustainable Construction & Materials
π Transport & Access
- Transport & Accessibility
- Servicing & Delivery
- Parking Provision
- Active Travel & PTAL
π’ Economic & Social
- Economy & Employment
- Retail Impact & Town Centre Vitality
- Public Safety & Crime Prevention
π What This Unlocks
AI Reasoning by Topic Each consideration invokes its own logic and sub-agent. Structured, traceable outputs.
Planning Balance Matrix Integration All considerations are scored, weighed, and surfaced in an explainable matrix.
Interactive Frontend Rendering Each material issue becomes a collapsible card in the UI, with toggle + override.
Override + Audit Layer Planners can amend or replace reasoning, which is logged and versioned per consideration.
Extensibility New considerations can be added with no system refactor. Just drop in a new JSON.
It also serves as the foundation for real explainability. When AI says βthis application is harmful to heritageβ, it can show which heritage asset, what type of harm, what level of harm, and why that matters β in language and structure planners recognise
π Next Steps
- Integrate material consideration modules into the Master Reasoning Model (MRM) pipeline
- Implement agent invocation and fallback logic per consideration category
- Build frontend UI for structured consideration cards with progressive disclosure
- Add per-consideration weighting + visibility toggles in the planning balance matrix
- Enable traceable overrides and planner comments per consideration
- Begin logging and evaluation of LLM-generated reasoning against officer-style outputs
- Draft a public-facing explainer or walkthrough showcasing material consideration handling in action
π΅ Closing Note
This work might not be glamorous, but it's foundational. If planning decisions are going to be AI-assisted, we need systems that reflect the actual structure of judgement β not just generate policy citations or hallucinate harm.
Material considerations are how planners make sense of complexity. Now the system can too.
"Not everything that counts can be codified. But most of it can be modularised."