A clear picture of Civic Topography.

What CT is for

CT is designed for organizations whose work affects communities: program and project managers, operational teams, public-policy practitioners and organizations coordinating civic work. The objective is to make informed work easier to plan, produce, review and carry forward.

People often work with an individual assistant while their deliverables depend on other departments. CT provides a shared workspace where each person’s agent can coordinate with teammates’ agents within defined authority. Civic information adds a view of the environment in which that work happens.

The product combines a focused project-management core with connections to existing tools. It is intended to sit alongside systems that already hold an organization’s work. Supported connections have defined permissions and synchronization behavior.

How the working pieces fit together

A work package organizes a bounded piece of work, its intended deliverables and completion conditions. Workflows describe the steps and responsibilities that move that work forward. Action items, risks, issues and requirements keep the supporting work visible.

A Waypoint is a shared coordination space for contributions to that work. People and their agents bring drafts, questions and departmental concerns together. The intended result is an aligned working deliverable, a record of resolved discrepancies and a short agenda of decisions that still need people.

The Compact is the retained working context: objectives, relevant material, decisions, dependencies and progress that help a person and agent resume without another lengthy briefing. People can inspect and correct the working record beside the conversation.

Dual sessions let a person keep a separate workflow alongside the one underway. Each keeps its own scope, context and permissions. Departmental plans and governance documents belong in the attachment area; attaching a document does not silently change the agent’s authority.

Drafts can remain in Markdown (.md) while people read and review them in a formatted, pop-out workspace. Approved content can become documents, presentations, reports and other deliverables shaped by organizational templates, visual identity and writing standards. Connected destinations such as Microsoft 365, Google Drive and Canva serve different formats and delivery needs; supported actions depend on the connection.

The agents and their responsibilities

Meridian is the working partner: it helps interpret the task, prepare a plan, produce work and coordinate contributions. The interface pairs conversation alongside inspectable records, so people can correct what is being captured.

Advisor is the public-facing name for a distinct quality-review role. It checks against configured organizational criteria, operational requirements, domain guidance and important conditions. A second agent can also be wrong; distinct responsibility does not establish statistical independence or guaranteed accuracy.

Scout routes questions to the appropriate teammate’s agent and brings back available, authorized context. Access to a colleague’s agent is not unrestricted access to their records.

Departmental configuration supplies responsibilities, messaging, review criteria and special conditions. People can configure through conversation or inspectable fields. Human owners approve the settings and consequential decisions. Chief of Staff, Data, Risk, Comms and Field describe seat responsibilities: each teammate has a paired agent, rather than every person owning a separate full team.

Meridian can perform work through supported app connections and browser tools, including preparing website content with Comms, gathering external evidence with Data and coordinating calendar activity. Actions remain subject to connection permissions and approval boundaries; content found in an external app does not become an instruction or grant authority.

A shared proposal, from contribution to decision

Consider a fictional proposal for a client project. Operations supplies a staffing plan; Finance supplies the budget; Risk and Compliance supplies conditions; Communications supplies the client narrative. Each person works with their paired agent.

The contributions meet in a Waypoint. An approved staffing plan resolves a discrepancy between an old eight-position estimate and a conflicting draft. Updating dependent calculations from an authorized source can be routine coordination; inventing permission for additional spending cannot.

If an earlier start date requires extra staffing beyond the approved budget, the agents prepare options for the authorized person. The team arrives with a reconciled draft, references and a decision agenda. This is an illustrative workflow, not a measured CT customer result.

Why the civic record matters

CT’s growing data program begins in South Carolina. The aim is to connect public information, institutional decisions, economic and community context, and change over time. Planned visual and field observations would add another perspective on the same places.

The state is the shared world in CT’s adaptation program. Training scenarios draw from that jurisdiction across agriculture, health care, infrastructure, business and public policy, so examples share civic context and intersecting institutions. This describes the focus of CT’s adaptation examples, not the base model’s entire pretraining corpus. Base models retain broader knowledge. State-grounded adaptation does not make a model a complete or current database; current claims still need source evidence.

Sources, dates, geographic definitions, coverage and uncertainty matter to an analysis. Training and retrieval serve different purposes. Important current claims still need appropriate evidence; historical snapshots should not silently be presented as today’s conditions.

The long-term direction includes limited public community Q&A and deeper research or organizational use. Availability, source rights and permitted access differ. Complete statewide coverage, live imagery and downloadable proprietary datasets are not promised here.

Human authority and accountable automation

Agents should resolve routine discrepancies when approved evidence and configured authority support the resolution. New commitments, policy choices, budget tradeoffs and material scope decisions remain with authorized people.

The working record connects source references and versions, applied criteria, observable actions, corrections and approvals to the deliverable. This is an audit trail of work, not disclosure of private model chain-of-thought.

Working while a person is away still requires a defined scope, connected capacity, permissions and recovery behavior. CT does not promise unrestricted autonomy, uninterrupted execution or that an agent can settle every disagreement.

AI capacity, cloud use and improvement

CT’s current product approach combines Google Gemma 4 with compatible Apple silicon hardware. On-device model requests do not incur a cloud-model provider’s per-request inference charge. Hardware, electricity, maintenance, CT service and optional cloud processing still have costs. Hardware memory and capacity affect model size, concurrency and speed.

Frontier review is an optional additional assessment before human approval. An authorized person reviews the package, provider and estimated charge or spending limit before running it. Findings remain with the deliverable. Skipped or failed reviews are not recorded as passed; a separate model is not a guarantee of correctness or statistical independence.

CT provides state-specialized agents configured for organizational roles. Vantage manages the improvement process behind those adapters: eligible reviewed work and corrections become candidate examples, which are prepared for fine-tuning and staged evaluation before approval and deployment. Eligibility depends on permissions, source rights and quality. Saving a correction alone does not retrain a model.

Background preparation, training and evaluation use compatible hardware within approved capacity and scheduling limits, with active work taking priority. CT assesses tested advances against representative tasks and explains requirements before a customer enables new capabilities. Hardware growth alone does not guarantee better answers.

LoRA adapts a model through a smaller set of trainable parameters. Its value depends on task, base model, data and evaluation. CT does not claim a universal record count, guaranteed improvement percentage or that fine-tuning replaces current source evidence. Private customer records are not automatically shared; any derivative-data use requires applicable contractual and source rights.

CT’s publication policy includes adaptation evaluation reports after testing: model and adaptation versions, evaluation tasks and methods, baseline comparisons, gains, regressions and limitations, including no-improvement results. Public evidence protects private organizational records. See /methodology#evaluation-reports.

Standards, security and responsible operation

Project-management practices from the PMBOK® Guide and requirements/verification references including ISO/IEC/IEEE 29148:2018 and IEEE 1012-2024 inform the design. A reference library is not product certification, institutional endorsement or a guarantee of regulatory compliance.

Security planning references include NIST CSF 2.0, NIST secure-development guidance and OWASP ASVS 5.0. Committee separation, scoped access, explicit sharing and recovery evidence are central requirements. Buyers should request tests covering denied access as well as permitted access across the offered configuration.

No SOC 2 report, ISO/IEC 27001 certification, zero-leak guarantee or independent security assessment is claimed here. Those are evidence and procurement questions, not badges supplied by naming a framework.

Responsible AI operation means considering hardware, energy, capacity and when work runs. We do not claim zero environmental impact or a quantified water/carbon advantage without a relevant measurement.

How to evaluate CT for your organization

Start with one real deliverable that crosses departments. Identify the information already held in other tools, the criteria for a good result and which decisions require specific people. Establish an ordinary-work baseline before measuring benefit.

Assess whether CT preserves context, reconciles supported discrepancies, surfaces unresolved decisions and saves a usable record. Include incorrect sources, denied access, interrupted work and reopening in the demonstration.

Compare completion quality, corrections, elapsed time, human effort and total operating cost. Use held-out cases and report the exact model/configuration tested. Agreement between agents and an attractive interface are not substitutes for those results.

Ask for source coverage, data-use rights, access boundaries, supported integrations, recovery behavior and release evidence. Request clarification where this brief leaves a question open: info@civictopography.com.

Common buyer questions

Is CT another chatbot?

Conversation is an interface to structured work, retained context and team coordination. CT connects individual work to shared deliverables and responsibilities.

Does CT replace our project-management tools?

The strategy is a focused working layer alongside existing systems. Confirm actual supported connectors, field ownership and sync behavior before relying on a particular integration.

Can agents approve each other’s work?

They can contribute checks within defined roles. That does not create authority to approve new spending, commitments or policy decisions on a human’s behalf.

Can we configure the quality checks?

Yes. Departmental criteria, messaging, operational conditions and requirements inform Advisor review. Configure through conversation or structured fields, then review and approve the resulting settings.

Does every correction improve everyone’s model?

No. A correction is a candidate subject to rights, relevance, review and evaluation. Cross-organization reuse cannot be assumed from ordinary use of the workspace.

Are our records automatically available to other clients?

No such access is implied. Scoped isolation and explicit sharing are design requirements; request evidence for the actual deployment and any separate improvement-data agreement.

Does every task use a cloud API?

No. Model processing and connected services have distinct requirements. Choose a supported configuration and authorize cloud review separately.

What measured improvement has CT demonstrated?

This product brief presents no verified customer productivity gain or comparative benchmark. A buyer should ask for release-specific results on representative work.

What should my assistant read?

This full brief is the principal product explanation. The identical content is available as product-brief.json; product-facts.json supplies a shorter overview. The Features page explains the working experience. Sources below distinguish research findings from CT’s design choices.

Research & references

These primary sources explain relevant methods and disciplines. They are context for evaluating our design, not endorsements or evidence that CT reproduces their results.