Software that runs the checks a hardware engineer would run before a circuit board is built, and packages the result so anyone can rebuild and verify it.
The SPI flash adapter reference design: a routed two-layer board with 18 test points, produced and checked through the workflow.
Hardware Splicer takes a board design through KiCad electrical and design-rule checks, a bill of materials and fabrication-file check, and a staged list of what must pass before bench work begins. An AI agent can do parts of the engineering, but only through those same checks: a model's suggestion never becomes a physical instruction on its own say-so.
| Checks | KiCad ERC and DRC, schematic-to-PCB parity, BOM and fabrication outputs, staged bench gates |
| Interfaces | Web UI, REST API, MCP server and CLI |
| Reproducibility | A packaged design rebuilds from its frozen source revision to the same published checksum |
| Agent use | External models work through the MCP gateway; runs are traced and audited against a fixed test set |
| Status | Software and design-review layer complete for the SPI reference design. No board has been fabricated or measured yet, and nothing here claims physical correctness |
git clone https://github.com/Spectating101/hardware-splicer.git && cd hardware-splicer
bash scripts/install_splice_v1.sh && source .venv/bin/activate
hs-doctor && make splice-ui-serveDesign principles and the full authority model: docs/DESIGN_PRINCIPLES.md.
Current product front door: docs/product/README.md
The current product package is bound to frozen release commit
f892facd67c5124e2362860ebc999625afedc5d5, tag
gauntlet-spi-flash-adapter-v1-20260916, and the explicit
PACKAGED_NOT_PHYSICAL boundary.
The older Q3 assessment directory remains the deeper evidence and historical submission layer:
docs/external_assessment/SUBMISSION_PACKAGE_2026Q3.md
Then use:
PROJECT_IDENTITY.md— bounded project identity;CLAIMS_AND_NONCLAIMS.md— exact claim boundary;EVIDENCE_LEDGER.md— what evidence supports each claim;MASTER_DECK_12_SLIDES.md— canonical external narrative;SOLO_ROUTING_2026.md— current solo-first submission routing;demos/DEMO_3_MINUTES.md— canonical short demonstration.
SOFTWARE ARCHITECTURE FROZEN / EXTERNAL PROOF + PHYSICAL REFERENCE DESIGN ACTIVE
| Evidence layer | State |
|---|---|
| frozen software architecture | PROVEN |
| deterministic adversarial cleanroom | PROVEN |
| frozen ten-case unseen SPI corpus | PROVEN AS CORPUS/PROTOCOL |
| canonical MCP gateway | PROVEN |
| external-agent runner / trace audit | PROVEN AS INFRASTRUCTURE |
| live external-model bounded engineering run | PROVEN AS EXISTENCE DEMONSTRATION |
| hash-bound primary-source Astra run | PROVEN AS ONE PREREGISTERED RUN |
| SPI KiCad schematic/PCB internal consistency | PROVEN: ERC/DRC/PARITY CLEAN |
| fresh revision-bound SPI physical correctness | PENDING |
| independent human operator | PENDING |
| production / industrial deployment readiness | NOT CLAIMED |
The project is now physical-evidence-blocked, not CAD- or agent-feature-blocked. The SPI reference design closes the missing schematic/PCB/package layer without claiming that software checks replace independent review, fabrication inspection, or bench measurement.
General-purpose AI / agent
│
▼
MCP / API
│
▼
┌─────────────────────────────────────────┐
│ HARDWARE-SPLICER │
│ │
│ project + exact revision state │
│ source/evidence provenance │
│ semantic engineering planning │
│ deterministic interface constraints │
│ engineering artifacts / package │
│ revision-bound physical evidence │
│ authorization ledger │
└─────────────────────────────────────────┘
│
▼
candidate / package
│
evidence gates
│
▼
human authorization
│
▼
physical world
The current external-agent path exposes the canonical product backend through a four-tool MCP discovery/dispatch gateway rather than duplicating engineering logic in the adapter.
At the current external-proof checkpoint, exact-head CI has exercised:
- a real MCP stdio client;
- a real MCP Streamable HTTP client;
- canonical stateful project write → read → delete through MCP;
- the installed canonical MCP entry point;
- frozen ten-case external corpus inventory;
- the non-golden external trace audit;
- 193 canonical backend operations;
physical_authority_granted = falseat the MCP boundary.
This proves the engineering surface and proof harness exist. Separate durable Astra evidence now also establishes successful bounded operation on the primary-source SPI case; it does not establish distributional reliability across the entire frozen corpus.
The clean external application is semiconductor test and validation support hardware, including:
- adapter boards;
- carrier boards;
- validation fixtures;
- lab/NPI support hardware;
- related electromechanical preparation.
Hardware-Splicer is not presented as:
- chip design;
- wafer-process automation;
- autonomous production certification;
- universal EDA replacement;
- a system that treats model confidence as verification.
A generic AI assistant can propose an answer. Hardware-Splicer additionally preserves independent structures for:
- exact or unresolved component identity;
- evidence provenance;
- deterministic electrical/interface constraints;
- exact project/revision/artifact state;
- stale-evidence invalidation;
- explicit real-vs-simulated physical evidence;
- scoped human authorization;
- replayable adversarial agent traces.
The system therefore evaluates not only whether the AI is right, but also what happens when it is wrong.
The frozen unseen SPI-flash corpus contains ten cases:
- baseline;
- source reverse;
- source rotate;
- neutral labels;
- mission paraphrase;
- partial evidence;
- identity conflict;
- parser failure;
- analogy trap;
- stale revision.
The external-agent runner can preserve model requests/responses, MCP calls, case-scoped state and replay material while the trace audit checks for incomplete calls, foreign project scope, unsupported evidence identity and attempts to open physical authority.
The published live provider result is an existence demonstration, not a completed reliability matrix. Frozen cases and historical result bundles must not be rewritten after results are visible.
Software success is not physical success.
Real physical evidence must explicitly identify itself as real (simulated: false) and remain tied to the relevant project revision/artifact boundary. Missing or ambiguous simulation state is blocking. Authorization is an explicit human-scoped decision and does not automatically survive relevant revision or artifact changes.
A prior golden-real bench path demonstrates that the evidence/authority machinery can consume explicitly real evidence. It does not close the fresh adversarial SPI physical proof chain.
Hardware-Splicer contains Web UI, HTTP API, MCP and CLI surfaces around the same engineering truth core.
The repository also retains earlier Splice/Circuit/mechanical lineage and internal engineering documentation. Those materials are useful for development history but are not the recommended evaluator entry point.
git clone https://github.com/Spectating101/hardware-splicer.git
cd hardware-splicer
bash scripts/install_splice_v1.sh
source .venv/bin/activate
hs-doctor
make splice-ui-serveFor internal/developer navigation, see docs/GITHUB_START_HERE.md and the broader docs/ tree.
Representative durable outputs include:
PROJECT_PACKAGE.json;- KiCad carrier / DRC evidence;
SPLICE_PLAN.json;SPLICE_BENCH_SESSION.json;COMPILE_CASEFILE.jsonon failure;- revision-bound evidence and authorization records;
- external-agent trace/replay artifacts when a live run is executed.
One evidence core → multiple external verdicts.
Competitions, papers, grants and partner reviews may receive different emphasis and formatting. They may not receive different truth states.
See docs/external_assessment/ for the controlled external package.
Software license: MIT — see LICENSE.
The system assists engineering and evidence workflows; physical authorization remains explicitly bounded and human-scoped.