Mellivora One

ImportForge is a self-improving warehouse import agent built during this hackathon. A model generates varied synthetic warehouse CSV files with deliberate quantity, unit, packaging, date and context errors. The importer executes tools, evaluates the results and can propose up to two bounded harness revisions per task. The product is the import agent; the history viewer exposes its evidence. DETAILS → AGENT STEP → MEASURED RESULT 1. Recursive Harnessing — change the way the agent works. Requirement: Automatically adapt rules, context policies and tool access to a task. Agent step: After improving file A, the procedure fails on a new file C. The agent uses execution feedback to revise its procedure, adding a channels JOIN between documents and suppliers before validation. Measured result: In recorded run f240172c, C improves from 11/60 to 60/60 correct decisions. The whole batch improves from 151/240 to 240/240. Independent controls remain 90/90, with zero regressions. 2. Recursive Harnessing — task-specific adaptation that transfers. Requirement: Adapt the agent environment to a specific task or use case. Agent step: Persist accepted procedure v6, then execute it on new rows and a blind holdout, rather than reuse cached row answers. Measured result: File B and blind file D each score 60/60 without another importer-model call; each has zero unsafe accepts. 3. Long Horizon Engineering — limited durable-memory evidence, not a full-scale claim. Requirement: Preserve coherent memory and optimize toward a goal using hard metrics. Agent step: In a separate recorded run, restore the same task from an Atlas checkpoint after a process restart, retaining progress, feedback and remaining request/tool limits. Measured result: Run 89ca64f0 progresses from 33/180 to checkpoint 129/180, then 180/180. The Run ID and checkpoint/history hashes match across different process boot IDs. Limit: This demonstrates safe-checkpoint recovery, not hours or days of autonomy or billion-token memory. Our primary statement is Recursive Harnessing. 4. Atlas Sandbox — use the event-provided environment. Requirement: Finalists must build using the provided MongoDB Atlas Sandbox. Agent step: Store runs, feedback, versioned procedures, checkpoints and validated receipts in Sandbox Cluster0, database importforge. Measured result: Atlas readback matches 125/125 receipts for run f240172c; saved harness v6 has the same Run ID. The 240 correct decisions include rejected rows and are not 240 imported records. 5. New hackathon work, public source and a one-minute demo. Requirement: Show work built during the event, provide a public repository and an accessible demo video. Agent step: Demonstrate the generator, feedback-driven harness revision, regression gate and Atlas persistence built for ImportForge. Measured result: Public GitHub repository, a 60-second demo video and an accessible read-only viewer. Recorded agent executions are labeled; fresh Atlas readback is available. Public AI calls and writes are closed. Scope: The agent edits bounded harness configuration, not application source code or model weights. Synthetic data only. Demo: https://importforge-demo-894fcf.azurewebsites.net/ Evidence: https://importforge-demo-894fcf.azurewebsites.net/history.html?run=f240172c-ec44-46c4-b2d5-f29aa90321bf GitHub: https://github.com/BorisKovStudio/importforge Video: https://drive.google.com/file/d/1ziLl94ePZAPxbG12CfpesoYW5R_yarS7/view?usp=drivesdk