Lab 9's agent was asked for companies held by exactly three managers. Its Text2Cypher tool wrote m.name and c.name, which do not exist, and got no rows. The agent's own reasoning says so. The answer was three tidy rows anyway, one of them a company with nine owners in the database. The answers that came from a Cypher template were right. On the page the two kinds are indistinguishable.

Exporting the agent and running its templates from a REPL is what told them apart. The local version that followed keeps only the first kind: a small model chooses one of a person's templates and fills its parameters, and never writes a statement. One model of four passed the gate set for that.

"Uncommon companies" was answered by a generated template that requires at least two owners and sorts by name. The 805 single-owner companies never appear, and the first ten rows are the start of the alphabet. The agent stated the ceiling and not the floor. A vague word in a question is defined by whatever statement answers it.

Lab 8 says the instance behind the Gemini Enterprise agent holds the SEC filings and that creating a second "Bob" overwrites the first. By the time it was queried it held a CRM-shaped sample and 28 Bobs under four label schemes. Describe a database before believing the handout about it.

Lab 4 puts a node key on (:Company {cusip}). The Graph Builder's extraction, run into the same database for Lab 6, then fails the moment the model produces a Company from a 10-K: no CUSIP, no node. Neither tool mentions the other. Before a second source writes into a database, list its constraints.

The same shape bit twice. First, curl could not resolve the host, and the pipeline's exit status was jq's: a wait loop built on it broke out on its first, false, success. Then Aura's Query API reported a failed statement with HTTP 200 and an errors array beside empty data, and the same pipeline printed the empty data and exited 0.

The Form 13 loader returned early on an instance that already held the data, which is the only kind it had seen. On an empty local server it stopped at the first IN TRANSACTIONS statement, which needs an auto-commit session. Running the lab's load on an embedded Neo4j is what found it, in time to fix it and run it against Aura before the instance expired.

Two of the local graphs carry a vector store: 7,331 vectors of 768 dimensions from a small embedding model run locally. Over twelve questions, the vector index returned 78 % of the true ten nearest on one graph and 79 % on the other.

Measured 2026-10-11. At a few thousand vectors exact search is affordable and the index is the wrong default. The same run found a second thing: one book repeats a heading a dozen times with identical vectors, so which ten come back is arbitrary under any search. Duplicates in the source are ties in the answer.

Embedding through a shared gateway with a per-seat key was checked against the local model with seven gates written and committed first. Six passed: a question's vector agreed to six decimal places by either route. The seventh, "the same top five either way", failed for one question in twelve. A probe showed both routes agree exactly under exact search; the disagreement was the index above. Without the gate the miss would have been blamed on the gateway.

The order that worked: a file, then an embedded server in the same JVM, then anything hosted. Lab 4's database and Lab 9's agent had local stand-ins within three days, and the loader bug above was found only on the stand-in. The hosted environment is also the part that expired.

Three days on, the lab account is gone. What remains is the queries, the prompts as files, redacted screenshots, the agent's export and the numbered findings. A lab is repeatable to the extent those were written down while it ran. Committing in small pieces as each step works is the same rule applied to the tooling; the stretch where that lapsed is the stretch that would be hardest to rebuild.

The labs were driven over the DevTools protocol in a browser with an empty profile, so the lab's account never met a personal one. What that cost:

• One driver at a time. Two hands on one field garbled a sign-in, and a failed sign-in on a lab account counts against a small limit.
• A driver must not take focus. A script that raised the window sent the person's next keystrokes into the query editor, twice.
• Click where the page reports an element is, with a real input event. element.click() opens no tab, and coordinates read off a scaled screenshot land above the target.
• A screenshot of the instance-created page holds the password in clear.