Follow data ownership, pre-execution checks, recorded artifact outcomes, reusable recipes and the frontier/cost tradeoff in that order. Each record keeps its original scope.
01 / Five-point product explanation
The agent declares. Schemaxis marshals the data.
In the strict declared-work interface, the agent never runs code or directly handles source data. It declares where the data comes from or how it is generated, which operations it passes through, and how it is output. Schemaxis marshals immutable input captures, sealed execution and outputs retained by digest.
This describes who operates on the inputs: no agent-supplied arbitrary scripts, shell or direct source-filesystem manipulation. The agent can receive bounded observations and receipt content; this does not promise that data is hidden from the model. Sealed container execution requires an explicitly configured, qualified backend. Native applications execute trusted runtime code. The historical local MCP demonstration alone does not establish that isolation.
02 / Five-point product explanation
The plan is checked before anything executes.
Admission, the declared schema, task scope and budget reservation are checked before execution. A request outside its authority is refused before a native job can run. Planning and inspection can run without launching the applications; native result verification is a separate step that reports the checks actually performed.
AdmissionSchemaScopeBudget reservation
NumPy request. SymPy-only scope.
In the retained MCP demonstration, a NumPy request under a SymPy-only scope returned SCOPE_EXCEEDED. Calls stopped at planning: zero native execution, all four prior NumPy output digests unchanged, and no successful current invocation to claim.
Historical prepared-recipe local MCP demonstration, core 5cfe1b0. It is separate from both the comparative study and the hosted container-isolation proof.
03 / Five-point product explanation
Smaller models failed alone, then produced passing artifacts.
Three smaller models that failed this authoring task without Schemaxis passed every required check with it.
29 September 2026 · ParaView authoring · checks passed
Model version
Without Schemaxis
With Schemaxis
Haiku 4.5
22/110
110/110
Sonnet 4.6
22/110
110/110
GPT-5.6 Luna
68/110
110/110
29 September 2026. One authoring conversation per model and method; 110 checks across an initial task and a revision on small, purpose-built serial ParaView fixtures. These are checks, not independent tasks. Older versions; no universal or current-model success claim.
Without Schemaxis. Haiku 4.5’s initial output had 32/32 images rejected. The original data records its transcript calling the result “production-ready”.With Schemaxis. The recorded workflow passed the required authoring checks. One conversation for this model/method; no general guarantee.
Inspect the independent reference and source dataIndependent requested view used by the recorded grader. No pixels are edited in this prototype.
Some historical Schemaxis runs inspected implementation details or used helper code; one revision wrapper was broken despite passing artifacts. These are artifact outcomes, not qualification of the new strict no-shell authoring interface. The Opus native control was excluded after provider substitution.
The core supplies the verification state and receipt instead of relying on the agent’s success sentence. The recorded misleading claim and changed artifacts do not measure a general reduction in false-reporting propensity.
04 / Five-point product explanation
Reuse a supported recipe with varied parameters and a new receipt.
A supported declared recipe can be repeated with varied authorised inputs and parameters; each execution produces its own receipt. The reusable document and bounded parameter contract are the mechanism, not a measured advantage for junior staff. Anyone using the supported, configured environment and authorised inputs can repeat it. Compatible runtimes, scope and checks still apply; the record preserves honest failure and unknown states.
The downloadable two-case SymPy example checks local-core binding and host planning only: no native execution or new receipt was created by those preflight checks, and no hosted file-template endpoint is implied.
Keep the frontier tradeoff and the cheaper-model result together.
Frontier models passed either way in the retained paired studies, and Schemaxis was slower.
GPT-5.6 Luna with Schemaxis passed all 110 checks at about one tenth of Astra’s native historical scored API cost (US$0.0815 versus US$0.8828); n=1 conversation each, older versions and 29 September 2026 prices. This is not current pricing or a total compute/support-cost estimate.
October is a separate record. Current scripted/native acceptance and the later three-model unseen-task study do not rescore September. The October Luna failure remains below.
31 October 2026 · current engineering candidate
Native checks and client integration.
Current source: core 68e139a, private runtime f2b5f8e. Five new application images and 31 native phases passed in the retained local evidence. The native fragment catalogue contains 132 candidate operations.
The new local hosted matrix completed 119 cases and reports matrix_complete: true and passed: true. This is scripted local qualification of the candidate, not actual-host acceptance, customer activation or a released service.
Three scripted local native client lifecycles passed without model calls. Claude client evidence covers connection/status only. These are local engineering checks, not a released host or external customer result.
Source checkpoint: 31 October 2026. This static preview does not poll private reports. See the source map for precise report references and hashes.
Historical · core 3f514b4
The earlier five-app hosted matrix.
The retained retry10 matrix passed 119 cases covering 132 operations, with 360 settled jobs/events, zero holds and 119 physical task deletions. That result belongs to its original source and qualification conditions.
The historical ParaView EnSight case covers geometry only. It does not establish named-array interoperability or qualify the current candidate.
Historical · unseen-task model authoring
Two passes. One failure.
GPT-6.1 SolPassed both applications / all six groups
GPT-6 AstraPassed both applications / all six groups
GPT-6 LunaFailed without execution
Three settled conversations at the original 3f514b4 study scope. The old hosted-budget limitation remains part of that record. The later budget repair does not rescore the study.
No general model-success, speedup or customer-demand claim follows from those conversations. Earlier September results and October prepared-task comparisons retain their original labels and evidence in Schemaxis’s existing website work.
External customer activation
Still to be demonstrated.
A customer must bring and author its own authorised task, connect its own tested agent client, inspect a fresh check, retrieve its receipt and confirm deletion. Repeat use in week two, onboarding effort and the actual payment state are separate evidence.
A internally produced example, scripted lifecycle or internal model study does not count as independent customer activation.
Separate October record · prepared recipes
Six conversations under a controlled catalogue.
The retained paired comparison used six conversations and 24 turns across SymPy, NumPy, ExifTool and FFmpeg, with prepared recipes and fixed inputs. Its 96 application checks are a different experiment from September authoring and the later three-model unseen-task record. It does not rescore either.
Read all eight recorded study limitations
Six exploratory conversations; turns are not independent replicates.
Prepared recipes and fixed inputs, not application authoring or broad adaptor coverage.
Controlled two-function model catalogue; not a default-agent benchmark.
Elapsed time includes model and tool work; fresh sessions do not prove cold provider caches.
Included quota is not free computation or an API list-price estimate.
Local results do not establish hosted availability, cloud capacity or a speed advantage.
A preceding run was interrupted by provider transport failure after five completed turns and one incomplete turn; it was retained and excluded as a whole before this fresh study.
The interrupted turn has no final token total. Its C$0 additional charge is an explicit operator adjudication from the included-only path, not a provider-issued charge receipt.
schemaxis
