--- title: Checked improvement feedback | SpinForge url: https://spin-forge.com/system-synergies/checked-improvement description: 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… --- S6 / SYSTEM CONNECTION [Explore all 64 product roles](https://spin-forge.com/system-synergies/product-map) # 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 qualification Existing 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. EXISTING PRODUCT CONNECTION ## 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 ](https://spin-forge.com/synthesis) THE CUSTOMER CHALLENGE ## 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 EVALUATE ## 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 RELATE ## The useful intersection. Existing components have distinct jobs. The proposed integration connects their responsibilities without treating access interfaces as proof of a shared runtime. 01 ### Retain the complete experience AgentForge and a qualified state/retrieval integration can retain input identity, configuration, failures, evidence and full execution cost. 02 ### Propose a different path GarmoForge and documented UNISOLVE selection mechanisms can propose compatible plans. A larger model is not required for every policy change. 03 ### Change the supported execution SpinKernel and relevant memory/runtime products can supply candidates within their published scope. Generated code still needs its own tests. 04 ### 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 COMMISSION ## An evaluation. A defined delivery. Not another mandatory product license. Scope the model, integration, acceptance and deployment needed for the actual customer workload. 01 ### A permissioned evidence ledger Specify allowed collection and reuse; keep private customer data and held-out evaluations separate. 02 ### A fixed comparison protocol Retain baseline, seeds, workload identities, failure outcomes and complete overhead. 03 ### A promotion decision record Approve a candidate only for the measured scope after its correctness and held-out criteria pass. 04 ### 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 & EVIDENCE ## 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. [Inspect the supplied UNISOLVE evidence](https://spin-forge.com/unisolve/validation) SELECTED CAPABILITIES ## Find the product role. These links identify candidate contributors. They do not mean every product is required or that every connector already exists. [SpinRoute](https://spin-forge.com/spinroute) [SpinKernel](https://spin-forge.com/spinkernel) [SpinProof](https://spin-forge.com/spinproof) [GarmoForge](https://spin-forge.com/garmoforge) [UNISOLVE](https://spin-forge.com/unisolve) [ANNForge](https://spin-forge.com/forge15/annforge) [SpectralForge](https://spin-forge.com/forge15/spectralforge) [TomoForge](https://spin-forge.com/forge15/tomoforge) [enkf_scale](https://spin-forge.com/forge15/enkf-scale) [StateForge](https://spin-forge.com/stateforge) [AgentForge](https://spin-forge.com/agentforge) [Unroute](https://spin-forge.com/solvers/unroute) [Krylov](https://spin-forge.com/solvers/krylov) [Batch tiny-instance](https://spin-forge.com/solvers/batch) [VRP](https://spin-forge.com/solvers/vrp) [View the full interaction map](https://spin-forge.com/system-synergies/product-map?mechanism=S6) [Related customer program](https://spin-forge.com/programs/compute-improvement) TAKE THE IDEA TO YOUR WORKLOAD ## 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. [Prepare an evaluation brief](https://spin-forge.com/system-synergies/evaluation?mechanism=S6)