Simone Consumption Engine
Consumption Engine is Simone's main research project. We build and test ways to preserve project sources, carry relevant evidence into an agent's task and revisit work when its sources change.
Consumption Engine is in development. There is no paid product available today.
How it works
Local software preserves selected documents, finds source passages and prepares context that an agent can inspect while working. This workflow is still in development across separate implementations.
Keep originals and source identity.
Select relevant passages and their conditions.
Use the evidence in a concrete task.
Assess the output against the task.
Keep useful work, failures and corrections.
Explicit checks can detect changed or missing declared sources. Choosing indispensable evidence and deciding whether it supports a conclusion still require review.
Illustrative request: update a project's engineering guidance using the current source. This describes the workflow, not an additional experiment.
The hypothesis is that this preparation can make recurring work easier to inspect and correct. It must be tested against a capable agent using the same sources and tools, with preparation and review included. A reported local maintenance task used prepared evidence; it did not establish an efficiency advantage.
Reviewed October 7, 2026
Evidence
Dated reports preserve each experiment's methods, outcomes and limitations. The public fixture is a separate runnable package.
Latest reported work
Source-linked instructions helped agents follow synthetic project rules. A later real-task comparison did not establish a benefit from a compiled guide. The new report follows both results and the next research question.
Selected documents can be preserved, indexed and delivered as source-linked context. Local studies include successful manual procedure reuse and unsuccessful agent-selection and maintenance comparisons.
The public package checks a recorded correction, stale-snapshot refusal, preserved history and archive replay in a fresh directory. Other expected refusal states are manually specified in its oracle projection.
The one-hour fixture receipt records 60 cycles of four cases, with zero failures. Repeating those cases tests their repeatability; it does not establish performance under load or broader reliability.
The September 24 local demonstration also retained a same-machine restore. These checks do not qualify private-runtime recovery on a clean machine or operating-system-enforced offline operation.
One correction, made visible
In the synthetic example, a policy changes from three retries to two. The earlier work remains in history, its stale snapshot is refused, and a successor has its own check.
The interactive recording presents fixed values. The separate public Python fixture lets you run the bounded lifecycle.
The next milestone
Evaluate a concise briefing of current facts, decisions and next steps against ordinary source reading, including preparation and review effort.
The latest findings separate local application results from synthetic studies and unfinished evaluations. The private staged outputs are separate from the public fixture.
We want to measure supported task completion, correction handling and total effort against a capable ordinary workflow. Preparation and review count alongside execution. Read the current development status.
A possible future pilot
An exploratory evaluation could focus on one collection you have permission to use, one recurring workflow and one named reviewer.
For changing engineering guidance, compare whether answers cite the right sources, retain relevant conditions and stop relying on superseded guidance. Measure the review effort against your existing process.
Proposed deliverables include source records, evidence-linked outputs, an evaluation report and known limitations. Agree on data handling, manual steps, timing and commercial terms before work begins.
Availability
The synthetic replay is public. The product runtime and local study receipts are private. A pilot would define the workflow and deployment requirements before proprietary documents are shared.
Private hosting, access controls, retention and recovery must be agreed and tested for that deployment. They are not established by the public fixture.
Supporting research
Preservation and retrieval provide the source foundation. BigBoss explores human approval for agent actions, Council explores claim review, and SEED supplies evaluation methods. These are distinct projects; completed integration is not implied.
Proposed cloud work would run repeatable comparisons on authorized data and measure task quality, latency and total cost. It is a development objective, not a claim of qualified multi-tenant deployment. Discuss research support.
Format, size and a short SHA-256 identifier, measured when this page was built.