Your assistant. Your domain.
Bring domain guidance, evidence checks, and accumulated engineering knowledge into your existing conversation. Start with a plugin. Grow an operating environment your organization owns.
Open source · MIT License · Local plugin + skills + MCP
Your conversation, model, and tools.
Evidence · context · reviewed procedures
01 / Why it exists
The best general models bring broad knowledge and strong reasoning. Your engineering team brings something more specific: process history, local constraints, evidence standards, and experience with what actually worked.
Too much of that context has to be reconstructed in each conversation. FabQED provides a place to apply it, check work against it, and retain reviewed learning for the next investigation.
The aim: increasing gains from the knowledge your organization builds, even as the underlying models change.
The harness is a seed layer. Development continues toward reliable engineering workflows with frontier assistants and your organization’s operating context.
02 / Where it starts
Explain the evidence. Diagnose the problem and choose the next check. Plan a discriminating experiment. Each capability is available across all three domains.
Chemical mechanical planarization
Investigate removal-rate shifts, nonuniformity, defects, and interactions between consumables, equipment, and process conditions.
“Removal rate shifted after the pad change. What evidence would distinguish the likely causes?”
Lithography and pattern transfer
Organize evidence around critical dimension, overlay, defect signatures, and process windows to choose a useful next measurement.
“This defect appears after pattern transfer. Help separate the hypotheses and plan the next checks.”
Integration and interconnect reliability
Structure investigations into bonding, interconnect, warpage, thermal, and reliability issues across the integration sequence.
“Bond yield fell after a process change. Help design an experiment that can discriminate between causes.”
Example starting prompts, not model outputs or validated diagnoses. Actual work uses your evidence and engineering review.
03 / How it works
Describe the problem and supply the evidence you can share. The semiconductor skill brings relevant guidance into the conversation; your assistant retains its reasoning and existing tools.
Local tools organize attributed observations, check supported calculations and declared ranges, and preserve the investigation. Guidance helps choose the next check; it does not establish a cause.
Return to persistent investigations. An administrator can connect an organization’s private store and make approved knowledge and procedures available to subsequent work.
The operating runtime holds the persistent context. Skills guide use; MCP connects the runtime to a compatible assistant. Engineers keep the interface they already know.
04 / How it grows
The long-term value is an organization’s accumulated domain intelligence: which evidence matters, which checks discriminate, and which procedures work under particular conditions.
The runtime includes an operator-reviewed lifecycle for candidate knowledge, strategies, and procedures, with promotion, retirement, and rollback. An assistant’s answer alone cannot promote a lesson. This is the machinery for cumulative improvement; the performance gains still need to be demonstrated.
Keep operating state outside the installed code, in a store your organization controls.
Define evidence access, review criteria, and the scope in which a model or procedure is qualified.
Retain context as models change. Qualify each integration against the same engineering requirements.
05 / Get started
We are continuing to build and evaluate FabQED Semi Harness. Public package downloads are paused while this work progresses.
Use our contact form to discuss access, an evaluation, or an organization deployment. Mention FabQED Semi Harness, your domain, and the workflow you would like to support.
The plugin and harness are released under the MIT License. Use, modify, and distribute them, including commercially. Retain the copyright and permission notice. The software is provided as is, without warranty.
If this work helps yours, please credit FabQED Semi Harness, developed by semiAIfoundry and contributors. Public acknowledgment is a voluntary request; credit guidance and a software citation are included.
Build with us
Help grow this seed toward integrated capability native to semiconductor engineering. Bring a useful domain check, an engineering tool adapter, a reproducible case, or a rigorous frontier comparison.
The complete MIT source includes the harness, plugin, build tools, tests, and synthetic fixtures. Your organization’s proprietary evidence and learned context can stay private; MIT does not require publishing your modifications or data.
Share a contribution or evaluation proposal through the contact form. Include a link to your public patch, fork, or example if you have one.
Release status
Questions
FabQED Semi Harness is a domain operating layer. It adds skills and local tools to a compatible assistant and preserves structured engineering context. It does not change model weights or replace the host’s chat interface.
That is the design direction. The native package connects through local stdio MCP; the host selects and runs the model. The broader runtime also includes managed provider adapters. Each host and model combination needs its own integration and qualification. The plugin does not automatically switch your assistant to arbitrary models.
This release runs locally. It does not include a remotely hosted GPT Actions endpoint or an authenticated web connector. Those require additional hosting, access control, and platform-specific integration.
The package contains code and domain guidance, not a private semiconductor corpus. A personal starter begins empty; an administrator can connect an explicit organization store. The plugin does not call a model provider itself. Evidence passed into the assistant is subject to that host’s data handling and deployment policies, so local storage alone does not make inference sovereign.
No. Investigations can persist, but reusable learning moves through recorded outcomes, candidate creation, held-out trials, independent review, and operator approval. This keeps a plausible answer from silently becoming organizational knowledge.
Please use the contact form to discuss access, evaluation, or collaboration. Public plugin and source downloads are paused while development and validation continue. The existing MIT release retains its license terms.
Start with a bounded engineering question. For a proprietary environment, begin with your evidence policy, model choice, and criteria for a useful answer.
Discuss an organization deployment