OpenAmer
An open-source self-verifying AI agent project that runs in a CPU-only environment. Its core feature is maintaining a publicly accessible structured outcome ledger that records the success or failure of each run along with the reasons, supporting a self-improvement loop.
Timeline (last 90 days)
OpenAmer is an open-source AI Agent, and every artifact it produces is actually executed and checked; gate run results are appended to a local ledger as one JSON object per line, containing timestamp, task, goal, Agent identifier, whether it passed, and a why field (which stores the verbatim raw output of the check, including exit code and error text).
The OpenAmer ledger's file is saved at memory/si/outcome_ledger.jsonl (the notebook's live file), and the repository's public snapshot is located at docs/EVIDENCE-LEDGER-SNAPSHOT.json.
The OpenAmer ledger's data spans from September 28, 2026 to October 13, covering 223 different goals, 833 tasks, and 79 Agents.
Goal gX has 187 recorded rows, a 100% pass rate, only one different task, and the why field of all records is 'exit=0 OK'; from October 4, 2026 to now, 187 runs have never failed, accounting for 10.1% of the entire ledger.
After excluding goal gX, the number of ledger rows drops from 1,852 to 1,665, the number of passes drops from 579 to 392, and the pass rate drops from 31.3% to 23.5%.
In the OpenAmer ledger, goals with 20+ different tasks (that is, goals that are truly working) have pass rates ranging from 0.0% to 29.8%; one of these goals ran 63 times with zero passes.
The OpenAmer ledger's 1,273 failures grouped by the first line of the why field have the following main failure signatures: exit=1 Traceback (uncaught Python exception) 761 rows accounting for 59.8%; JSON decode errors 199 rows accounting for 15.6%; exit=2 plus uv/python path 96 rows accounting for 7.5%; regex not found 51 rows accounting for 4.0%.
The OpenAmer repository's public snapshot docs/EVIDENCE-LEDGER-SNAPSHOT.json had 615 results as of September 30, 2026.
OpenAmer runs entirely CPU-only on a single Windows laptop, with no GPU, no cluster, and no cloud inference.
The rules derived from the audit: pass rates should be reported per goal rather than only aggregated; and '≥20 rows and 100% pass' should be treated as suspicious rather than successful.
12 more timeline events
All Facts (20)
OpenAmer is an open-source AI Agent, and every artifact it produces is actually executed and checked; gate run results are appended to a local ledger as one JSON object per line, containing timestamp, task, goal, Agent identifier, whether it passed, and a why field (which stores the verbatim raw output of the check, including exit code and error text).
50%UnverifiedThe OpenAmer ledger's file is saved at memory/si/outcome_ledger.jsonl (the notebook's live file), and the repository's public snapshot is located at docs/EVIDENCE-LEDGER-SNAPSHOT.json.
50%UnverifiedThe OpenAmer ledger's data spans from September 28, 2026 to October 13, covering 223 different goals, 833 tasks, and 79 Agents.
50%UnverifiedGoal gX has 187 recorded rows, a 100% pass rate, only one different task, and the why field of all records is 'exit=0 OK'; from October 4, 2026 to now, 187 runs have never failed, accounting for 10.1% of the entire ledger.
50%UnverifiedAfter excluding goal gX, the number of ledger rows drops from 1,852 to 1,665, the number of passes drops from 579 to 392, and the pass rate drops from 31.3% to 23.5%.
50%UnverifiedIn the OpenAmer ledger, goals with 20+ different tasks (that is, goals that are truly working) have pass rates ranging from 0.0% to 29.8%; one of these goals ran 63 times with zero passes.
50%UnverifiedThe OpenAmer ledger's 1,273 failures grouped by the first line of the why field have the following main failure signatures: exit=1 Traceback (uncaught Python exception) 761 rows accounting for 59.8%; JSON decode errors 199 rows accounting for 15.6%; exit=2 plus uv/python path 96 rows accounting for 7.5%; regex not found 51 rows accounting for 4.0%.
50%UnverifiedThe OpenAmer repository's public snapshot docs/EVIDENCE-LEDGER-SNAPSHOT.json had 615 results as of September 30, 2026.
50%UnverifiedOpenAmer runs entirely CPU-only on a single Windows laptop, with no GPU, no cluster, and no cloud inference.
50%UnverifiedThe rules derived from the audit: pass rates should be reported per goal rather than only aggregated; and '≥20 rows and 100% pass' should be treated as suspicious rather than successful.
50%UnverifiedThe project's stated reasons for making failures public include: self-measurement without failure records is merely optimism, failure distribution is a training signal, and anyone can verify via the ledger file
50%UnverifiedOpenAmer is a self-verifying AI agent running on a pure-CPU Windows laptop that maintains a public outcome ledger and has self-improvement capabilities
50%UnverifiedOpenAmer's self-improvement loop reads the outcome ledger, and improving the pass rate means overcoming the dominant failure records
50%UnverifiedOpenAmer's outcome ledger file path is memory/si/outcome_ledger.jsonl, where each run is recorded as one line containing a timestamp, task, objective, executing agent, and passed verdict
50%UnverifiedThe total number of outcomes recorded in the OpenAmer ledger is 1,701, of which 543 passed and 1,158 failed, giving a pass rate of 31.9%
50%UnverifiedFailure records in the OpenAmer ledger far outnumber success records, with nearly 70% being failures
50%UnverifiedFailure records in OpenAmer's ledger include a why field that categorizes failure reasons into human-readable information, such as task timeout, tool returning an unexpected data structure, path not reaching the objective, missing dependencies, etc.
50%UnverifiedOpenAmer's ledger records are written before side effects occur, and side effects are only retained when the ledger line is successfully persisted, thereby constituting a mechanism to prevent duplicate execution
50%UnverifiedOpenAmer's write-order design guarantees that a receipt will never correspond to an effect that never occurred, but the terminal state where an effect has landed while the receipt is missing may still arise
50%UnverifiedOpenAmer opts for at-least-once semantics, preferring a detectable inconsistent state over allowing ghost records
50%