11 confirmed
The result met the bar written before the check.
Consumption Engine: how the work is checked
How 34 research decisions were recorded
On September 27, 2026 the team building the Consumption Engine recorded 34 decisions. Each one wrote down the question, the change to try, the check and the pass bar before the check ran. The outcome was added afterwards. Fourteen results missed their bar and five were inconclusive. Both outcomes remain visible alongside the eleven confirmed and four open decisions.
Process evidence. Public export. Python standard library checker. No network after downloading the export
The day in numbers
30 decisions resolved. 11 cleared their bar.
The team is a small group of AI agents that Jonathan directs, using local compute. The ledger is its working record, not a report written afterwards.
14 refuted
The result missed the bar. The row records by how much and, where it was measured, why.
5 inconclusive, 4 open
The check could not run as written, or had not finished when the ledger was exported.
What the failures changed
A failed bar is where the next change comes from.
A gain that was a leak
Indexing the team's correction records appeared to lift answers to reworded questions from 7 to 12 of 15. A leak check found test files in the search corpus. With them excluded the gain was 0 of 15. The next two decisions made the index exclude test material by construction.
Decisions D-002, D-004, D-010.
A cheap checker that did not transfer
A small classifier trained on 2,000 synthetic errors learned them, but still accepted 27 of 31 real bad rows against a bar of 9. The real errors are claims that go beyond their source, which the synthetic edits did not model.
Decision D-008.
Onboarding that transmits current facts
A fresh agent given only a generated onboarding file answered 14 of 15 sealed questions about current project state, with 0 stale answers. A fresh agent given the existing hand-written docs answered 6 of 15, with 4 stale.
Decision D-019. The question bank and the onboarding file draw on the same records, so this shows the file carries what the records hold.
Claims from a real project hold up
A local model extracted claims from a real software project. 49 of 50 randomly sampled claims were supported by the source lines they cite, against a bar of 45. A separate source locator finds the file behind 83% to 85% of test passages in 4.4 ms at the 95th percentile.
Decisions D-029 and D-030. The 49/50 support judgment was graded by Claude-based agents on the same team, not externally reproduced. The locator passed on its fifth attempt; each earlier attempt is recorded.
Check it yourself
One command confirms the published record is intact.
The checker confirms the ids are consecutive, every row states a question, a check and a bar, every resolved row has an outcome recorded no earlier than the row itself, and the tally and file hashes match. Change one status and it fails. This verifies the public process record and ledger integrity; it does not reproduce the private outcome receipts.
git clone https://github.com/thisisntjon/consumption-demo
cd consumption-demo
git checkout 7df8429ba8de11ae3bf571cc9ee6bab03386b1ae
python ledger/verify_ledger.pyLedger README, DECISIONS.jsonl, verify_ledger.py, export rules and hashes
Qualification boundary
What this establishes
The ledger shows how the work is run: bars stated first, failures recorded, changes driven by measured misses. It does not show that the Consumption Engine outperforms other systems. That test, an agent given the engine's inventory against an agent given the raw project on sealed questions, had not passed when this was published.
Limits: timestamps were recorded by the agents themselves; the public commit is the first outside timestamp. Grading was done mostly by Claude-based agents on the same team, with the grader separate from the author where the row says so. That is a separation of roles, not independent reproduction. Receipts cited in the rows are private. Local file paths were redacted, and each row carries a hash of its private original.