SAGE Suite public site

Public, nonproprietary product and evidence hub for SAGE Suite. The site is written for manufacturers, engineers, design partners, researchers, future customers, and technical reviewers encountering the platform for the first time.

npm install
npm run generate:simulation
npm run generate:replay
npm run dev

The homepage begins with the promise, “Make every difficult cut a decision you can defend.” It then shows the customer outcome: a structured job model, ranked process-plan alternatives, and an evidence-linked review packet. Simulation, open-data research, and the five public-safe assurance checks support that product story rather than replacing it.

The homepage publishes a concrete offline pilot format: scope one representative job, compare alternatives in shadow, review an evidence packet with a qualified engineer, and decide whether to stop, repeat, add measurements, or design a separate physical trial. It also states the customer-data boundary and requires written permission for cross-customer training or a public case study. The public email link is for scoping only; controlled production files wait for agreed data-use terms and a separate secure intake route.

The pilot section does not imply live controller or ROS access, customer proof, measured process improvement, or authority over production equipment.

The product page leads with a deterministic public aggregate from the full disclosed NVIDIA Isaac Sim shadow campaign:

• 2,542 program comparisons with matching provenance records
• 659 verified shard archives
• nine shape/material coverage cells
• 46.7% median reduction in the campaign's composite synthetic objective
• 2,542 of 2,542 programs below their own same-simulator baselines

The reviewed aggregate is published at public/data/sage-public-nvidia-simulation-v1.json. It contains counts, coverage, distribution statistics, runtime scope, integrity status, and the fail-closed model decision. It excludes raw command schedules, private geometry, customer telemetry, objective composition, internal thresholds, and model internals.

A separate retrospective aggregate at public/data/sage-public-nvidia-surface-integrity-v1.json covers all 2,542 NVIDIA robot episodes. Its median modeled finish-pass interval is 1.78–3.43 µm, with a 2.64 µm midpoint. Historical corner radius was not recorded, so the result is explicitly assumption-bounded and modeled_not_measured; it is not measured Ra or a surface-quality guarantee.

Robot articulation ran in NVIDIA Isaac Sim. Cutting loads were mechanistic and high-frequency vibration was model-synthesized. The campaign therefore demonstrates within-simulator shadow optimization and evidence capture; it is not held-out policy generalization, physical cutting, measured cycle time, part quality, or safety proof.

The homepage also embeds public/media/sage-simulation-replay-v5.mp4, a silent 120-second SIMULATED public Isaac Sim capture with tool-stock cutting contact. It alternates three CNC surrogate-training scenes with three NVIDIA UR10e robot shadow-optimization scenes over the disclosed circle, rounded-rectangle, and slot classes.

public/media/sage-isaac-capture-manifest-v5.json records the six jobs and 2,880 showcase frames. The capture is newly generated and public-safe; it is not a private campaign USD, customer program, physical cutting, measured machine footage, or a hardware command stream.

The current robot scenes use a fixed-base UR10e. A separate public engineering schematic introduces a proposed force-aware relative-pose study comparing the same synthetic job across a fixed-base cell, a robot on a seventh-axis linear rail, and a part on a two-axis tilt-rotary positioner. That comparison is not in the current replay and is not a live ROS path, physical-machining result, or measured efficiency claim.

/simulation/ begins with the complete NVIDIA campaign and surface-integrity aggregates, then publishes a separate deterministic 120-row generic CNC-like teaching trace. Every teaching row is labeled SIMULATED, the observed count is zero, surface roughness Ra is always null/unmeasured, and the advisory threshold is fully disclosed as illustrative public-demo logic rather than SAGE production policy.

/research/ summarizes the public-safe SAGE research program:

• traceable machining and surface measurement;
• digital-twin verification, validation, and uncertainty quantification;
• transfer across machine, tool, material, geometry, pose, and time;
• force-aware relative pose across fixed-base, rail, and articulated-workpiece cell arrangements;
• evidence exchange and non-actuating shadow pilots;
• NIST CRADA, SBIR, and MEP collaboration routes; and
• a counsel-gated TechRxiv, Zenodo, and peer-review release sequence.

The site never publishes the founder-private NIST draft, patent material, weights, objective coefficients, calibration internals, private source, raw programs, or customer data.

npm run generate:simulation
npm run check:simulation
npm run generate:replay
npm run check:replay

The simulation generator writes CSV, JSON, and SHA256SUMS.txt under public/data/. The replay command verifies the reviewed Isaac Sim MP4, poster, captions, capture manifest, and web-binding manifest under public/media/; it does not regenerate or substitute the captured renderer output.

npm run verify

The verifier checks the NVIDIA aggregate schema and reviewed values, public simulation artifacts, evidence labels, unknown-value handling, source hashes, static downloads, crawler policy, security headers, canonical URLs, and public-copy disclosure boundaries.

Every exported file under public/data/ is checked byte-for-byte against its reviewed source, including aggregate JSON downloads outside the teaching simulation's historical checksum manifest. npm run check:export-integrity requires an existing build and tests corruption of each exported data copy; it restores every tested copy and leaves reviewed source artifacts unchanged. The regression check also runs as part of npm run verify.

Vercel is the sole production host:

https://sage-public-evidence.vercel.app

The retired GitHub Pages deployment workflow is intentionally absent. Preview deployments should be verified before promotion to production. Custom-domain preparation lives in docs/CUSTOM_DOMAIN_CUTOVER.md.

• All cooperative crawlers, including AI and archive agents, are permitted by app/robots.js.
• Global metadata requests indexing and following without cache/archive limits.
• app/llms.txt/route.js provides an AI-readable public site index and keeps the SIMULATED, modeled-not-measured, and non-actuating boundaries visible.
• app/sitemap.js publishes the search sitemap.
• Visible-content-matched JSON-LD describes the product and public datasets.
• Google Analytics loads only after visitor consent when configured.
• Vercel response headers enforce CSP, framing, permissions, referrer, and content-type boundaries.

• Physical CNC and robot command authority is disabled.
• Simulation evidence remains labeled and missing values remain unknown.
• Campaign gains are not presented as measured machine outcomes.
• The saved surrogate was not promoted after independent held-out and generated-boundary checks.
• No customer data, raw campaign geometry, command schedules, private thresholds, proprietary algorithms, or model internals are published.
• Kernel descriptions stop at the user-facing question and fail-closed outcome.