Retain the complete experience
AgentForge and a qualified state/retrieval integration can retain input identity, configuration, failures, evidence and full execution cost.
S6 / SYSTEM CONNECTIONExplore all 64 product roles
Better proposals.
The same acceptance standard.
A useful feedback loop can improve which configuration, formulation or operation to try next. It does not need to let a learned guess establish feasibility, correctness or optimality.
Cross-product integration / customer qualificationExisting products · CLI · MCP · Headless
The products retain their available interfaces. These connections describe integration and evaluation opportunities—not a claim that every combined runtime has been delivered.
Recursive Engine is already the synthesis meta-system.
The eleven existing synthesis products include Gödel Optimizer for algorithm improvement, Neural Schedule for learned temperature schedules, SpinForge Warp for differentiable annealing, and Recursive Engine for cross-product improvement. This page describes how that existing family could connect to the wider checking and operating portfolio; it is not a proposal to invent a replacement meta-system.
Explore the eleven products ↗THE CUSTOMER CHALLENGEWhat separate
What separate
tools can miss.
A promising trial can fail on a different workload. Training-only improvements, hidden overhead and discarded failures can produce a confident but unhelpful default policy.
THE COMBINED APPROACH TO EVALUATEConnect the
Connect the
right responsibilities.
Retain authorized successes, failures and counterexamples. Propose a bounded change, independently check its claimed properties, evaluate the complete workload on held-out cases, and promote or reject it under a predeclared policy.
Accumulate useful computational experience while preserving explicit release, quality and evidence requirements.
HOW THE PRODUCTS RELATEThe useful
The useful
intersection.
Existing components have distinct jobs. The proposed integration connects their responsibilities without treating access interfaces as proof of a shared runtime.
Propose a different path
GarmoForge and documented UNISOLVE selection mechanisms can propose compatible plans. A larger model is not required for every policy change.
Change the supported execution
SpinKernel and relevant memory/runtime products can supply candidates within their published scope. Generated code still needs its own tests.
Accept only supported claims
SpinProof measures empirical quality and performance. Independent solver checkers establish their own mathematical properties. Neither type substitutes for the other.
INTERACTIVE STUDY / NOT A PRODUCT BENCHMARK
See the connection.
A local browser illustration of the idea. No customer model is uploaded and no SpinForge engine is invoked.
An illustrative review policy. The last state is a hypothetical passing candidate, not a reported SpinForge cross-product result. No runtime policy is changed by these controls.
WHAT YOU COMMISSIONAn evaluation.
An evaluation.
A defined delivery.
Not another mandatory product license. Scope the model, integration, acceptance and deployment needed for the actual customer workload.
A permissioned evidence ledger
Specify allowed collection and reuse; keep private customer data and held-out evaluations separate.
A fixed comparison protocol
Retain baseline, seeds, workload identities, failure outcomes and complete overhead.
A promotion decision record
Approve a candidate only for the measured scope after its correctness and held-out criteria pass.
A rollback and monitoring scope
Version the selected policy and define when a regression triggers review or rollback.
ILLUSTRATIVE CUSTOMER WORKFLOW
A new routing policy looks promising
Evaluate it beyond the examples used to fit it. If it misses the held-out target, preserve the evidence and keep the existing default. A good hypothesis is not a production mandate.
ACCEPTANCE MEASURES
- Held-out accepted outcomes
- Total cost including proposal overhead
- Preserved failures and regression behavior
- Promotion scope and rollback readiness
SCOPE & EVIDENCEClear at
Clear at
every boundary.
Documented components and proposed cross-product work are not the same evidence.
What the source supports
UNISOLVE’s inventory reports a partition selector that missed its improvement target and was not promoted. That is a demonstrated evidence boundary, not a successful cross-product self-improvement result.
The integration to establish
Permissioned telemetry, stable evaluation splits, cost accounting, explicit promotion rules and a versioned rollback path.
What must not be assumed
Predictions are proposals, not theorem authorities. No cross-customer learning is assumed without rights. Better tuning does not establish universal gains; promotion needs current workload, hardware and deployment evidence.
Products remain independently usable
Choose only the components that fit. Original and upgrade products are not additive speedups. No generalized benefit, hosted activation, certification or execution authority is implied.
SELECTED CAPABILITIESFind the
Find the
product role.
These links identify candidate contributors. They do not mean every product is required or that every connector already exists.
TAKE THE IDEA TO YOUR WORKLOADDefine what
Define what
must improve.
Start with an original model or workload, an accepted baseline and the protected requirements. Agree the comparison before promising a system-level gain.