The current product story
How can AI agents use what a project already knows?
Teams and agents produce information faster than they can verify, retain and reuse it. Consumption Engine is the main project in a group of local research implementations for preserving sources, finding relevant evidence and carrying it into an agent’s task.
The implementation and results described here were reviewed October 7, 2026. Jonathan Simone is the solo founder and developer, working with AI assistance. Consumption Engine is in development. There is no paid product available today.
The problem we are working on
A research finding becomes a summary. The summary becomes an instruction. An agent acts on it. When the source changes, the team needs to know which work depends on it, what still holds and what needs checking again.
Our goal is evidence-linked institutional knowledge for technical and research teams: preserve what was learned, make its support inspectable, and carry corrections into the next task. Source identity is one part of that process. Judging whether evidence supports a conclusion still requires substantive review.
What is built today
- Preserve and find. Local tools ingest selected documents, retain source identity, index text and return passages. An optional PDF page reader helps inspect prose against preserved originals; tables, equations and diagrams still need direct review.
- Prepare a task. Assemble selected source passages with their roles, qualifications and locators. Explicit source checks can refuse changed or missing material and incomplete coverage. Callers still choose indispensable evidence.
- Use and revisit work. Retain checked procedures and select them for later inputs. Bind recorded outputs to their sources and checks. A changed declared dependency can prevent reuse until the material is refreshed and the result checked again.
The September 30 ingestion and delivery report describes these local capabilities and their test conditions. This is a broader foundation than the original fixed catalog sample. It does not establish complete integration of every research collection, automatic contradiction discovery or autonomous learning.
Checks run at defined task steps or on explicit refresh. They do not continuously watch every source or automatically repair all downstream work.
What the results support
The project-rules study follows Claude-based synthetic experiments through a later real-task comparison. Supplying relevant instructions helped on synthetic tasks, but a whole-corpus baseline matched compiled guides on a small corpus. The subsequent real-task comparison did not establish a coding benefit. The aggregate summary records the methods, counts and corrections. These author-run studies remain separate from the practical work below.
The separate local application report describes a staged save/load feature and checks that a restored session could continue. A separate AI session checked the result. A corrected ordinary workflow also succeeded. This is useful software output in one task, not a demonstrated quality or cost advantage from context preparation.
The procedure studies recorded successful manual reuse on new inputs and refusal after selected source changes. Later assisted work used source-linked material to repair specific project documents.
The same research also retained unsuccessful agent-selection and maintenance comparisons. More prepared context did not consistently produce better work or lower cost. Preparation, retrieval, execution, correction and review all belong in the eventual efficiency measure.
These are author-reported local studies, with separate review within the research effort. Their private runtime and full receipts are not published. They are not external reproduction or measured customer benefit. Separate synthetic evaluations cannot be combined with the practical application into one integrated-system success rate.
What you can run
The public synthetic fixture is a small, separately distributed Python package. Its executed lifecycle carries a retry policy from three to two, refuses the earlier snapshot, preserves the original history and checks a saved archive after restoration in a fresh directory.
The package also contains manually specified expected refusal states. Matching that expected projection is an integrity check, not evidence that every state came from a new runtime query. The website’s interactive example presents recorded values; it does not run an agent or the engine in your browser.
The public package demonstrates this bounded contract. It does not reproduce the private studies, establish semantic truth or qualify recovery of the private runtime on a clean machine.
What comes next
Evaluate a concise briefing of current facts, decisions and next steps against ordinary source reading, including preparation and review effort. A briefing prototype is built; its benefit evaluation remains unrun. The latest dated note explains that research direction. The simulation application and public replay retain their separate evidence.
Broader transfer, automatic procedure selection, lower full effort and a demonstrated advantage on real-repository tasks remain unestablished. Private staged outputs are not a new public product release. The September 30 candidate conditions remain in that dated report.
Get involved
Researchers can examine the methods and results. Developers can run the public fixture and report discrepancies. Organizations interested in research support can help define a focused experiment and its publication requirements. Contact Jonathan to discuss the work.
A possible future pilot
Start with one document collection you have permission to use and one recurring workflow. For example, a team could track changing engineering guidance and check which saved answers need to be revisited.
This is an exploratory proposal. We would agree on scope, data handling, manual steps, deliverables and acceptance criteria before work begins. The proposed evaluation would compare source support, correction handling and reviewer effort with the team’s existing process. A report would retain failures and limitations alongside useful results.
Discuss a possible evaluation, explore the product or inspect the public fixture.
This brief is the website’s current account of the product. Dated evidence reports preserve the conditions and outcomes of individual studies. BigBoss, Council and SEED contribute supporting work; completed integration is not implied. The original September 23 implementation note remains available.