Pre-flight spend quote
The gate checks every window first, so the customer gets a fixed, fingerprinted price before the heavy job runs, and can stop it if over budget.
Break-even 0.164 s of job per window; quote matched the run list
Features
Built and tested locally (350 feature tests passing on 2026-11-14). Not yet deployed to Databricks. The gate's 8-qubit circuit and Big J run on a classical simulator.
The gate checks every window first, so the customer gets a fixed, fingerprinted price before the heavy job runs, and can stop it if over budget.
Break-even 0.164 s of job per window; quote matched the run list
Runs the heavy job on a small circuit-chosen sample of skipped windows so "avoided cost" is measured, not estimated.
Sample estimate 41.63 s vs true 41.84 s; costs ~7.5% of the saving
Keep how close each call was; label windows solid or borderline; receipt counts only solid skips.
Margin is already measured inside the gate, currently thrown away
The cheap gate runs first and predicts exactly how much heavy work is coming: real wait times, right-sized workers.
Forecast matched the full run window for window
The gate marks every window run/skip right inside the customer's Databricks table; only "run" windows start the heavy job.
Gate ~2 ms/window; 223 windows gated in 0.34 s, 6 passed
An auditor checks any single decision on their own laptop without trusting our servers.
Matched the full run on all 8 windows
The circuit's own answer picks which skipped windows get spot-checked, so nobody can cherry-pick.
Same hash on 3 runs; a 0.001 change moved only that window's fingerprint
Data is closed off in blocks so Big J's answer for a past window never changes when new data arrives.
Big J's past reads changed in all 18 frames when data was added; sealed replay identical twice
A surprise daily input proves the engine running is the same engine, by how it behaves, not by file checksums.
Caught a code change; added internal fingerprint caught a tiny weight change the answer hash missed
After install, the circuit must reproduce a published fingerprint before it touches customer data, then sizes the saving.
Reproduced the stored fingerprint e82d0e9d
Only the circuit's answer leaves the customer, sealed per customer.
Found today two customers with the same data get the same hash (a leak this fixes)
Big J keeps a learned memory for each customer in their own locked Unity Catalog folder; replay still gives the same hash.
Memory replays identically; whether it improves calls over time not yet measured
An AI agent may act only on windows the gate passed, with a signed permit tied to the answer hash.
Skipped window got no AI call despite an "ignore the gate" injection; forged permit refused
Watches each stream's run/skip rate; a broken feed (zero-filled gaps, running totals, sorted rows) gets held before it costs money.
Broken feeds pushed gate pass rate from 2.7% to 53-98%
Hourly self-check on a pinned input plus feed-health signals; a skip-everything day from bad data pages someone.
Caught out-of-order, flat-sensor and duplicate feeds
Big J's QFT add-on on the gate's own 8 qubits catches repeating cycles the gate misses.
Gate ran 0-9 of 100 cycle windows; QFT caught 100/100; noise stayed skipped (standalone, not wired in yet)
Calls that flip under tiny wobbles go to the operator to decide; the choice goes on the chained record.
4-5 in 100 calls flip on noisy data
A new customer uploads one sample, sees the circuit replay it, skip scrambled and noise copies, and gets a first receipt in one sitting.
Replay identical; scrambled and noise both all skipped
Click any window: a plain sentence with the circuit's own numbers and Big J's 9 calls.
Same call on all 6 stored demo windows when recomputed
Any AI model (or Genie) explains; every sentence is checked against the circuit's answer, and the circuit wins.
4 of 4 planted errors caught
ideas_round1/IDEAS_20.md (round 1, 2026-11-14); each 'measured' line is copied word for word from a real local run