Factor

Factor helps US residential and light-commercial electrical contractors turn a DWG/DXF package into an evidence-linked takeoff, a reviewable cost calculation, and an approved PDF proposal delivered by email.

The pilot uses the fictional Factor Demo Electric workspace. Its USD purchase prices and labor rates are demo assumptions, not verified supplier invoices or measured field costs. Unknown geometry, scale and electrical connectivity remain visible. A priced allowance does not establish technical completeness.

The selected public demonstration uses ETSU's Tiny Electrical Plan.dwg, an educational electrical drawing; the recording plan includes its source attribution. Electric Symbols.dwg is uploaded as reference CAD; its examples are excluded from quantities. Both unchanged, openly licensed files and attribution are in fixtures/public/etsu. Private customer drawings remain local audit inputs.

Explore the recorded ETSU demo without an account: drawing evidence, assumptions, recorded prices and a display PDF. This public mode does not run analyses or send email. The live workspace requires an authorized Cognito account; judges receive separate private credentials. See trial access. For the separate local synthetic walkthrough, run the local fixture console with pnpm --dir apps/web dev. Set the public client settings in apps/web/.env.example to use a deployed API instead.

The graph v3 flow is company setup → CAD/reference upload → autonomous research and policy resolution → decision/cost summary → exact PDF → one final approval. An approved company policy can apply labeled symbol scenarios, header-unit assumptions and budget alternatives. Only unresolved material decisions require input. Optional scope or commercial changes produce a new calculation and PDF. One main CAD is selected explicitly. CAD files may be up to 64 MiB, packages up to 128 MiB; the supporting PDF limit remains 20 MB. Large packages use direct authorized S3 upload rather than email attachments.

Only Amazon Nova 2 Lite (us.amazon.nova-2-lite-v1:0) is callable by model verification and production paths. Strands 1.52.0 runs the native graph; deterministic nodes own CAD, quantities, company rules, pricing and packaging. Nova investigates immutable evidence with bounded tools and returns advisory claims, never quantities, coefficients, prices or sending authority.

The 2026-09-09 synthetic verification record records actual AgentCore tool use, reuse of CAD/takeoff during a 5% commercial revision, native approval/resume and one SES-delivered proposal. Two earlier IAM refusals are retained. The decoded email attachment equals the approved PDF; two replays of the completed delivery record left its attempt and provider message unchanged. This is one controlled demo, not a field accuracy or reliability study.

The 2026-09-10 autonomous ETSU verification records real reference-tool use, three commercial revisions reusing five analysis checkpoints, a superseded approval rejection, native resume and a single SES delivery of the exact approved PDF. Nova did not demonstrate additional priced scope beyond the policy baseline in that comparison. The drawing's horizontal route scenario and missing conductor inputs remain explicit; this is a conditional demo estimate, not independently verified electrical quantities.

The repository is intentionally split by trust boundary:

• apps/web — browser review and approval experience.
• services/control-plane — authenticated API and workflow control plane.
• services/agent-runtime — Strands/AgentCore orchestration and model tools.
• tools/cad-inspector — isolated .NET CAD inspection worker.
• packages/contracts — canonical schemas and language adapters.
• packages/pricing-engine — deterministic money and pricing rules.
• infrastructure/cdk — AWS infrastructure as code.
• fixtures and tests — synthetic/openly licensed verification material.

Verify

Prerequisites: Windows with PowerShell 7, Node 22, pnpm 11.19.0, Python 3.12 and .NET 8. Install the locked JavaScript packages and Python development tools:

pnpm install --frozen-lockfile
python -m pip install -r requirements-dev.txt
python -m pip install -e packages/pricing-engine -e services/control-plane -e "services/agent-runtime[dev]"
pnpm --dir apps/web exec playwright install chromium

Run the complete no-AWS/no-model verification suite:

pwsh -File scripts/verify.ps1 offline

The offline target validates the contract spine in Python, TypeScript, and C#; compiles and verifies fixtures and CAD gates; exercises deterministic pricing, agent graphs, catalog provenance, control-plane/package delivery, both browser modes, CDK synthesis, and controlled-local end-to-end scenarios. It never runs AWS preflight, Nova/AgentCore, SES, or deployment operations.

Live model checks are separate and require an explicit target plus -Live (e.g. agentcore-spike or catalog). Pull-request and push CI runs only the offline target; the live-model-verification workflow is manual and should be protected by the factor-live-verification GitHub environment.

Safety boundary

No proposal is sent from model output alone. Money is represented as decimal strings, calculations are deterministic, evidence is addressable by hash, and an approved proposal is bound to the exact calculation and artifact hashes.

S3/DynamoDB artifacts are the domain source of truth. Strands session state tracks execution and native approval interrupts. Leases and conditional writes protect concurrent runs and reviews. Commercial changes create a new run and reuse validated CAD, raw takeoff, evidence research, optional topology review and policy resolution when their analysis dependencies remain unchanged. Quantity, scale, reference, interpretation, rule or model/prompt changes invalidate affected analysis. Explicit reanalysis preserves prior results. Earlier sent proposals remain immutable. See the graph v3 architecture.

The durable delivery worker checks the authenticated approval, native resume, current review version, PDF and MIME hashes before SES. Provider acceptance, delivery and bounce are distinct states. A sending timeout becomes SEND_UNKNOWN; the worker does not blindly resend.

Release evidence and operation

• Pre-video audit, 2026-09-10: deployed fixes, regression results, current product limits and submission checklist.
• Current closure ledger: implemented behavior, live verification and remaining publication work.
• Architecture and deployment: AWS boundaries, reproducible deployment and controlled delivery verification.
• Five-minute demo outline: what to show and which claims require evidence before recording.

Private customer drawings and local live-test evidence are excluded from source and runtime images. The Turkey project audits are local robustness checks, not US market validation or an expert-verified accuracy benchmark. No 98% accuracy, 50% time saving, 30-project evaluation or human usability success rate is claimed.