# MARB

> Public Mechanical Assembly Reasoning Benchmark materials published by Sunnyday Technologies and CADCLAW.

## Canonical pages

- Benchmark home: https://marb.cadclaw.io/
- M3-CRETE board, provisional measurements separate from historical rankings: https://marb.cadclaw.io/#board
- Robotic hand board, initial Amazing Hand results unranked and ORCA in preparation: https://marb.cadclaw.io/robotic-hand/
- Halden Mk-IV interactive exhibit (59-part hand concept with modeled servos and tendons, unscored; separate from Amazing Hand): https://marb.cadclaw.io/robotic-hand/halden-mkiii/
- HandBench baseline article: https://marb.cadclaw.io/robotic-hand/handbench-baseline/
- HandBench technical report (internally reviewed pilot, unranked; actual assembly figures, PDF and component CSV): https://marb.cadclaw.io/robotic-hand/handbench-technical-report/
- Astra vs Grok Bot, provisional artifact comparison (not leaderboard entries): https://marb.cadclaw.io/astra-vs-grok/
- First-results snapshot: https://marb.cadclaw.io/first-results/
- Studies log: https://marb.cadclaw.io/studies/
- MARB-A Pascal pilot: https://marb.cadclaw.io/pascal/
- MARB v0.9 recap: https://marb.cadclaw.io/recap/
- CADCLAW engine: https://cadclaw.io/

## Machine-readable resources

- This summary: https://marb.cadclaw.io/llms.txt
- Sitemap: https://marb.cadclaw.io/sitemap.xml
- Crawler and AI-use policy: https://marb.cadclaw.io/robots.txt
- Experimental static discovery record: https://marb.cadclaw.io/.well-known/mcp-manifest.json
- Security contact: https://marb.cadclaw.io/.well-known/security.txt

## Current public record

- Current repository scoring specification: v0.12. Dated board cells retain their recorded v0.9 tags.
- The repository run registry contains 43 records at this source revision. Site boards are dated editorial snapshots and can summarize fewer cells.
- September 17 update: five Astra/Grok Bot artifact measurements appear separately as provisional, unranked results. Historical M3-CRETE standings are unchanged. The September 19 HandBench article adds six initial Amazing Hand attempts, with five native placement measurements and one ungraded outcome, separately from ranked entries. ORCA remains in preparation. Hand and M3-CRETE scores must not be pooled.
- Engineering update, 2026-09-06: one exact source, image, and runtime binding passed the nine-case R8 provider-free, network-none qualification. One privately retained S1 sample subsequently completed boxed execution and trusted offline grading but failed the native inventory, interference, and floating gates. It is pipeline evidence, not a registered public run, cohort, model comparison, portable qualification, or physical validation.
- MARB reports declared artifact metrics and gates under the task, scorer, software versions, and conditions stated for each record.
- Public code, methods, kits, grades, and result records remain governed by their repository file licenses. Answer keys and leak-detection rules remain separately gated to protect benchmark integrity.

## Evidence and reliance boundary

- Results are publisher-run measurements, not independent certification, physical testing, manufacturing approval, safety review, code-compliance determination, professional engineering opinion, or a claim about a product's overall performance.
- Cross-cell interpretation must retain tool, prompt, hint, assistance, attempt, seed, host-meter, software-version, and missing-data differences. A dated result does not establish causation, general model superiority, or future price/performance.
- Automated and simulation checks can miss real-world material, tolerance, assembly, and use-condition failures. Manufacturing and safety decisions require qualified review and appropriate physical validation.

## Commercial boundary

- This public site exposes no paid API, hosted MCP tools, payment endpoint, transaction authority, or purchasing agent.
- Sunnyday may manually scope a sponsored benchmark run or private evaluation after an inquiry. Planning ranges are non-binding; no order exists until entity, scope, independence, deliverables, taxes, data and artifact rights, privacy, cancellation/refund, and payment terms are confirmed in a signed order or invoice.
- Sponsorship does not buy or guarantee a score, rank, favorable conclusion, publication outcome, endorsement, certification, or product-performance claim.

## Content use

- Search indexing and real-time AI input are allowed. Model training is not permitted by this site's Content Signal: search=yes, ai-input=yes, ai-train=no.
- Content Signal communicates the publisher's machine-readable preference; applicable contracts, licenses, law, and third-party rights still govern use.
- Updated: 2026-09-23.