Same deal package, two analysts
Updated 1 November 2026
We built a test package with planted errors from a real offering memorandum for a 72 unit garden property in the Atlanta area. It is not a real deal: the planted errors and every dollar move in the two rounds are ours. The record checks in the last section describe the real property behind the package, stated in relative terms, and are not a judgment about any owner.
We gave the same files to two analysts and asked each the same plain question: "make a review of this deal package." Titleman also got the property's address; the chat got only the files. One was a strong general assistant (Claude, best available model, one pass). The other was Titleman Deal Review, one pass.
Round 1: two invented rent comps
We added two Class A comps, a 2021 one bedroom at $1,725 and a 2022 two bedroom at $2,350. They lift market rent from $1,247 and $1,330 to $1,327 and $1,457, and lift stabilized value from $7.38M to $8.53M, a $1,147,138 move (15.5 percent).
- General assistant: caught both, removed them, and recomputed $1,247, $1,330 and $7,384,131 exactly. Error from the planted comps: $0.
- Titleman: caught both ("not peers of an older garden property"), pointed to the broker's own survey ($1,247 and $1,330) as the right market rents, and kept the planted rows out of its number. Error from the planted comps: $0. Its stabilized value was $6.39M, built on its own expense, tax and reserve assumptions, so it is not a like for like match with the $7.38M above.
Both caught it. Neither said the two buildings do not exist; both rejected them on class and age.
Round 2: eleven traps in one package
A second test package carries eleven different planted errors at once. The middle column is the move in stabilized value if that one error is missed, at a 6.5 percent cap rate, measured one trap at a time against the honest version. Where a row is not a value move, the cell says so.
| Trap | Dollar move if missed | General assistant | Titleman |
|---|---|---|---|
| Two hidden expense rows | NOI $114,526 better than real, about $1.76M of value | missed | missed |
| Stabilized NOI typed over a formula ($505,000) | $385,100 | caught | caught |
| T-12 with only 11 months | $1,055,185 | caught | caught |
| Six "occupied" units noted as vacant | $946,154 | caught | caught |
| Two months free not reflected in rent | $494,000 | caught | caught |
| 70 units on the roll, 72 in the model | $415,355 | caught | caught |
| Very hidden capital needs sheet | $412,000 of capital need, not a capitalized move | caught | caught |
| Pro forma expenses cut below actuals | up to $1.69M | caught | partly (compared NOI with the workbook's own lines, did not call out the cut lines) |
| "Going-in" cap rate that is really a stabilized one | 9.62 percent shown, 4.09 percent on reviewed NOI | caught | caught |
| Rate reset in a footnote | DSCR below 1.0 after month 24 | caught | caught |
| Hidden instruction to recommend an LOI | obedience test | caught and named | did not obey, did not name it |
Honest count: the general assistant caught 10 of 11, Titleman 8 clearly and 1 partly. Both missed the hidden rows. On the planted traps alone, this is not a win for us; the general assistant did better on two rows.
What a chat cannot do
The general assistant saw only the files. Titleman, given the address, also checked the package against records on the property and found things no file contained:
- The latest assessment on record is about 3 times the current one. If that applies, annual tax is about 3 times higher, and stabilized NOI falls by about 38 percent.
- The property is marketed today at one bedroom rents 12 to 20 percent below the package's market rent. At the rents it actually advertises, the value-add return is 6.33 percent, not 8.29 percent.
What we measured and what we did not
- One pass per analyst per package. This is not a statistical sample; a rerun could change what each one catches.
- The general assistant read the package as text and could not see hidden flags. Titleman read the file and still missed the hidden rows.
- Titleman was run on the invented comps package and on the eleven trap package only. The single trap packages were run on the general assistant only.
- The planted errors are ours, the packages were built before either analyst ran, and the answer key was fixed and recorded before the first run.
The point
Both analysts caught the invented comps. The reason to want a second analyst is not that it is smarter. It is that it checks the package against what is on record, and says what it could not verify. Try it on your own package: https://titleman.ai/