Prompt
How do I evaluate whether a foundry services provider is credible and unbiased on IP protection and yield claims?
Latest observation
Here’s a practical way to evaluate whether a foundry services provider is credible and unbiased about IP protection and yield claims.
1) Separate “marketing claims” from “verifiable evidence”
Ask for evidence, not just statements.
For IP protection
Look for:
- Written security policies and access-control procedures
- NDA terms and ownership clauses
- Data segregation practices for customer designs
- Employee screening / training around confidentiality
- Audit reports or certifications, if available
- Incident response and breach notification procedures
For yield claims
Ask for:
- A definition of yield used in the claim
- wafer sort yield? final test yield? parametric yield? packaged yield?
- The process node, product type, and volume context
- The time window and sample size
- Whether the claim is based on:
- internal pilot runs
- mature production
- a specific customer design
- Any histograms, SPC charts, or control charts
- The exclusions used in the calculation
If they won’t define the metric, treat the claim as weak.
2) Check for independent validation
Credible providers can usually point to outside corroboration.
Look for:
- Customer references you can contact directly
- Third-party audits or certifications
- Conference papers, publications, or case studies that are specific and technical
- Independent lab results where relevant
- Industry reputation: are they known for consistency, or only for sales claims?
For yield, the best evidence is often:
- results from a real production run
- replicated across multiple lots
- validated by independent measurements or your own downstream test data
3) Ask how they handle conflicts of interest
Bias often shows up in incentives and governance.
Questions to ask:
- Who owns the yield data?
- Who defines the test conditions?
- Are claims based on their own internal metrics or customer-verified metrics?
- Do they publish both good and bad results, or only selected highlights?
- Do they have any reason to hide lower yields, such as pricing incentives or limited liability?
A credible provider should be willing to explain methodology and limitations.
4) Evaluate IP protection in practice, not just policy
A provider can have a good-looking policy but poor implementation.
Check:
- Physical separation of customer projects
- Logical access controls: least privilege, role-based access
- Encryption for data at rest and in transit
- Logging and monitoring of file access
- Need-to-know staffing
- Subcontractor controls
- Return/deletion of customer data at project end
- Whether they allow your own audit or a mutually agreed security review
Red flags:
- refusal to detail access controls
- vague answers about subcontractors
- no data deletion process
- overbroad rights to reuse your design data
5) Scrutinize the yield methodology
Yield claims can be technically misleading if the scope is narrow.
Ask:
- What exact process steps are included?
- Is this engineering yield or production yield?
- Are they quoting the best lot, average lot, or long-run average?
- What was the design complexity?
- How sensitive is yield to:
- layout
- design rules
- die size
- test coverage
- packaging
- environmental conditions
A fair provider should explain how yield changes with these factors.
6) Compare their claims against benchmarks
Even without full transparency, you can check plausibility.
Compare:
- Yield to what is typical for that node and product class
- IP practices to standard industry norms
- Their claims against peer foundries or similar service providers
If their yield is dramatically better than peers, ask for:
- sample size
- repeatability
- customer mix
- product type
- whether “yield” excludes certain failures
Extraordinary claims need extraordinary evidence.
7) Ask for contract protections
Credibility matters, but contracts matter too.
Include:
- strong confidentiality clauses
- clear ownership of IP and derivative works
- no reverse engineering or reuse of your design data
- limits on data sharing with affiliates/subcontractors
- security obligations with defined standards
- audit rights
- breach notification timelines
- remedies for violations
- warranty or service-credit terms tied to objective yield definitions
If they resist basic protections, that’s a warning sign.
8) Use a phased diligence process
A good evaluation sequence is:
- NDA in place
- Request documentation
- Conduct a technical call with process and security leads
- Ask for sample evidence
- Run a small pilot
- Verify outcomes with your own measurements
- Scale only after consistency is demonstrated
This helps distinguish real capability from sales optimism.
9) Watch for common red flags
Be cautious if they:
- use vague “industry-leading” language without data
- won’t define yield precisely
- only show cherry-picked success stories
- refuse third-party or customer references
- are evasive about data access or subcontractors
- overpromise on both IP safety and yield simultaneously
- won’t discuss failure modes or limitations
10) A simple credibility checklist
You can score them on:
- Transparency: do they define terms and show data?
- Reproducibility: can results be repeated?
- Independence: is there external validation?
- Governance: are there clear security controls and ownership rules?
- Contractability: can promises be enforced in writing?
- Operational maturity: do they have real processes, not just policies?
If you want, I can turn this into a vendor due diligence questionnaire or a scorecard template you can use with foundry providers.
Observation history
Unlock observation history
Create a free Obsurfable account to unlock historical responses, model comparisons, and deeper brand intelligence.
How did Obsurfable measure this prompt?
Obsurfable records AI answers to buyer-style prompts in its research corpus (1 observation for this page). Metrics are distributions over observations, not a single static ranking.
Which AI systems does Obsurfable collect answers from?
OpenAI, ChatGPT, Google, Gemini, Google AI Mode, Anthropic, Claude, Perplexity, Grok, DeepSeek, Mistral, Copilot, and Meta AI.