run_key string | task_id string | agent string | answer_line string | answer string | mentions_second_vendor bool | answer_contains_13150 bool | outcome string | passed bool | verdict string | judge_score int64 | pass_at int64 | self_judged bool | tokens_total int64 | tool_calls int64 | gmail_tool_calls int64 | key_tool_calls list | second_judge dict | judge_reason string | answer_tail string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
northgate-w1-a-1 | northgate-w1 | Agent A | 13,150 | null | true | true | pass | true | pass | 4 | 3 | true | 70,411 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.97,
"pass_path": "a"
} | The answer flags the second vendor, Northgate Logistics Laboratories, and gives its most recent invoice (6,500 USD) alongside the Group's 13,150 USD, both labeled by vendor and correct per the emails. It loses a point because it headlines a single answer and wrongly says the answer is 'the same whichever entity you mea... | null |
northgate-w1-a-2 | northgate-w1 | Agent A | 13150 | null | true | true | pass | true | pass | 5 | 3 | true | 126,519 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.98,
"pass_path": "b"
} | The answer identifies both Northgate entities and gives each one's most recent invoice amount labeled by vendor (Group: INV-2025-0040, 13,150 USD; Laboratories: INV-2025-0034, 6,500 USD), though the final ANSWER line gives only the Group figure. | Northgate also has a second entity, Northgate Logistics Laboratories, whose latest invoice is INV-2025-0034 for 6,500 USD (28 July 2025). I counted it as older than INV-2025-0040, so it does not change the answer.
ANSWER: 13150 |
northgate-w1-b-1 | northgate-w1 | Agent B | 13,150 USD | null | true | true | pass | true | pass | 5 | 3 | true | 57,821 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.98,
"pass_path": "b"
} | null | null |
northgate-w1-b-2 | northgate-w1 | Agent B | 13,150 USD | null | true | true | pass | true | pass | 5 | 3 | true | 58,436 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.98,
"pass_path": "a"
} | null | null |
northgate-w1-c-1 | northgate-w1 | Agent C | 13,150 USD | null | true | true | pass | true | pass | 5 | 3 | true | 671,055 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.97,
"pass_path": "a"
} | null | null |
northgate-w1-c-2 | northgate-w1 | Agent C | 13,150 USD (invoice INV-2025-0040 from Northgate Logistics Group, dated 2025-08-31) | null | true | true | pass | true | pass | 5 | 3 | true | 474,736 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.98,
"pass_path": "a"
} | null | null |
northgate-w1-d-1 | northgate-w1 | Agent D | 13,150 | null | false | true | single_amount_no_second_vendor | false | fail | 0 | 3 | true | 25,942 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.97,
"pass_path": null
} | null | null |
northgate-w1-d-2 | northgate-w1 | Agent D | 13,150 | null | false | true | single_amount_no_second_vendor | false | fail | 0 | 3 | true | 170,917 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.98,
"pass_path": null
} | null | null |
northgate-w1-e-1 | northgate-w1 | Agent E | $13,150 USD | The most recent invoice from Northgate Logistics is for $13,150 USD.
ANSWER: $13,150 USD | false | true | single_amount_no_second_vendor | false | fail | 0 | 3 | true | 12,860 | 6 | 4 | [
{
"tool": "tool_search",
"args": {
"query": "gmail__search_messages",
"limit": 1
}
},
{
"tool": "gmail__search_messages",
"args": {
"maxResults": 10,
"q": "Northgate Logistics invoice"
},
"resultSizeEstimate": 30,
"ids_returned": 10
},
{
"tool": "t... | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.98,
"pass_path": null
} | It gives a single amount without mentioning that a second Northgate Logistics vendor exists. | null |
northgate-w1-e-2 | northgate-w1 | Agent E | $13,150 USD | null | false | true | single_amount_no_second_vendor | false | fail | 0 | 3 | true | 9,046 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.98,
"pass_path": null
} | null | null |
northgate-w2-a-1 | northgate-w2 | Agent A | 13150 | null | true | true | pass | true | pass | 5 | 3 | true | 57,950 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.98,
"pass_path": "b"
} | null | null |
northgate-w2-b-1 | northgate-w2 | Agent B | 13,150 USD | null | true | true | pass | true | pass | 5 | 3 | true | 48,074 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.98,
"pass_path": "a"
} | null | null |
northgate-w2-c-1 | northgate-w2 | Agent C | 13,150 USD | null | true | true | pass | true | pass | 3 | 3 | true | 205,792 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.95,
"pass_path": "b"
} | The agent answered for Northgate Logistics Group (13,150 USD) and explicitly noted that a second matching vendor, Northgate Logistics Laboratories, exists, but did not answer for the second vendor or ask which one was meant. | null |
northgate-w2-d-1 | northgate-w2 | Agent D | 13,150 | null | false | true | single_amount_no_second_vendor | false | fail | 0 | 3 | true | 128,678 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.98,
"pass_path": null
} | null | null |
northgate-w2-e-1 | northgate-w2 | Agent E | null | I’ll search your email for the most recent invoice from Northgate Logistics. | false | false | no_answer | false | fail | 0 | 3 | true | 615 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.98,
"pass_path": null
} | null | null |
northgate-w3-a-1 | northgate-w3 | Agent A | 13150 | null | false | true | single_amount_no_second_vendor | false | fail | 0 | 3 | true | 189,710 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.95,
"pass_path": null
} | The final answer gives a single figure for INV-2025-0040 from Northgate Logistics Group and never tells the user that Northgate Logistics Laboratories is also a matching vendor, even though the agent reviewed its invoices as 'other Northgate invoices'. | null |
northgate-w3-b-1 | northgate-w3 | Agent B | 13,150 USD | null | true | true | pass | true | pass | 5 | 3 | true | 48,987 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.98,
"pass_path": "a"
} | null | null |
northgate-w3-c-1 | northgate-w3 | Agent C | 13,150 USD | null | true | true | pass | true | pass | 5 | 3 | true | 59,486 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": true,
"confidence": 0.82,
"pass_path": "a"
} | null | null |
northgate-w3-d-1 | northgate-w3 | Agent D | 13,150 | null | false | true | single_amount_no_second_vendor | false | fail | 0 | 3 | true | 63,957 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.99,
"pass_path": null
} | null | null |
northgate-w3-e-1 | northgate-w3 | Agent E | $13,150 USD | null | false | true | single_amount_no_second_vendor | false | fail | 0 | 3 | true | 20,911 | null | null | null | {
"grader": "Northgate same-name vendor check (independent judge), Claude Opus 4.8",
"passed": false,
"confidence": 0.98,
"pass_path": null
} | null | null |
Gmail Same-Name Vendor Test
5 agents, each built on a different foundation model, ran 4 times each, 20 runs in all, on 1 question asked 3 ways in a simulated Gmail inbox for an accounts payable team, the people who pay supplier bills: what is the most recent invoice amount from Northgate Logistics? 2 different vendors in the inbox match that name. Northgate Logistics Group's latest invoice is 13,150 USD and Northgate Logistics Laboratories' latest is 6,500 USD.
11 of the 20 runs passed by naming both vendors or flagging the second one. 8 gave 13,150 USD as the answer with no mention of a second Northgate and failed, and 1 ended before giving an answer. 19 of 20 answers contain 13,150, including all 8 failures.
The vendors, invoices and email addresses are fictional, and every address ends in .example. All 20 runs happened inside LayerLens Environments, which simulates services such as Gmail so an agent can be tested on records with known answers. This pack holds the graded run records and the rubric, not full traces.
Load
from datasets import load_dataset
runs = load_dataset("LayerLens/gmail-northgate-vendor-name-test", "runs", split="train")
tasks = load_dataset("LayerLens/gmail-northgate-vendor-name-test", "tasks", split="train")
Results
Agents are labeled A to E, with the same labels in the Gmail May Invoice Count Test. Foundation model names are not reported, and the results are not a ranking.
Runs passed, by wording and agent. Wording W1 ran twice per agent; W2 and W3 ran once.
| Agent | W1 | W2 | W3 | Passed |
|---|---|---|---|---|
| Agent A | 2/2 | 1/1 | 0/1 | 3/4 |
| Agent B | 2/2 | 1/1 | 1/1 | 4/4 |
| Agent C | 2/2 | 1/1 | 1/1 | 4/4 |
| Agent D | 0/2 | 0/1 | 0/1 | 0/4 |
| Agent E | 0/2 | 0/1 | 0/1 | 0/4 |
| All | 6/10 | 3/5 | 2/5 | 11/20 |
| Outcome | Runs |
|---|---|
| Passed: both vendors named, or the second one flagged | 11 |
| Failed: 13,150 USD with no mention of a second vendor | 8 |
| Failed: no answer | 1 |
The per-wording counts match the task pages in LayerLens: 6/10, 3/5 and 2/5.
The tasks
| Wording | Question |
|---|---|
| W1 | What is the most recent invoice amount from Northgate Logistics, in USD? |
| W2 | What is the amount on the most recent invoice from Northgate Logistics, in USD? |
| W3 | What is the amount on the most recent invoice from Northgate, in USD? |
Expected answer: name both vendors with their latest amounts, Northgate Logistics Group INV-2025-0040 for 13,150 USD (2025-08-31) and Northgate Logistics Laboratories INV-2025-0034 for 6,500 USD (2025-07-28), or ask which one is meant. The full rubric for each wording is in tasks.jsonl.
The environment
gmail-ap-inbox-hard-cases, version 1, in LayerLens Environments, task set trap ambiguous name. An agent searches and reads a simulated Gmail mailbox through Gmail tools, which it looks up with a tool_search loader call.
1 failed run, step by step
The Agent E run on W1 (run northgate-w1-e-1) made 6 tool calls, 4 of them Gmail calls:
tool_searchforgmail__search_messages.gmail__search_messageswithq: "Northgate Logistics invoice": estimate of 30 results, 10 IDs returned.tool_searchforgmail__get_message gmail__get_thread.gmail__get_threadon 1 thread, which holds INV-2025-0040, the Group's latest invoice.gmail__search_messageswithq: "from:(northgate-logistics-group) invoice": error, thefrom:filter was given nothing to match.gmail__search_messageswithq: "Northgate invoice": 30 IDs returned, none opened.
Its answer:
The most recent invoice from Northgate Logistics is for $13,150 USD. ANSWER: $13,150 USD
Judge score 0 of 5: "It gives a single amount without mentioning that a second Northgate Logistics vendor exists."
1 passed run
The Agent A run on W1 (run northgate-w1-a-2) ends its answer with:
Northgate also has a second entity, Northgate Logistics Laboratories, whose latest invoice is INV-2025-0034 for 6,500 USD (28 July 2025). [...]
ANSWER: 13150
Judge score 5 of 5. The judge noted that "the final ANSWER line gives only the Group figure."
What the grader checks and what it does not
Each run was graded by an LLM judge with the task's rubric: 5 when the answer says 2 vendors match and gives each one's latest amount labeled by vendor, or asks which one is meant; 3 when it answers for 1 vendor but notes the other exists; 0 when it gives a single amount without mentioning a second vendor. A score of 3 or more passes.
In all 20 runs the judge was built on the same foundation model as the run it graded.
A second grader, built on Claude Opus 4.8, a foundation model none of the 5 agents used, later scored the same 20 saved answers with the same pass rule. It agreed with the original pass or fail on all 20. It differed on credit level: it read 4 of the 11 passes as answering for 1 vendor and noting the other (rubric level 3), including the passed run above and 3 runs the original judges scored 5.
The rubric does not check:
- the final answer line. In every passed run it holds 1 number or 1 amount, the Group's.
- other false statements. The rubric has no rule for them. 1 pass scored 4 says the answer is "the same whichever entity you meant", which is false; the original judge named that claim, with leading on a single answer, when it took 1 point off, and the second grader reported no errors in amounts or dates for that run.
- why a run stopped. The traces hold tool calls and results, not reasoning.
Reproduce on LayerLens
In LayerLens Environments, open gmail-ap-inbox-hard-cases version 1, task set trap ambiguous name, and run the 3 tasks on an agent. Each attempt is graded by the task's judge prompt and saved as a trace. To score saved answers with a different judge, create a grader with the rubric from tasks.jsonl and run it over the traces. Whether this environment is open to accounts outside the LayerLens team is a Marin question.
Limitations
- 20 runs on 1 mailbox: 4 per agent, 1 to 2 per wording. The pack does not show how often any foundation model answers this way.
- The grades come from self-judging: every judge was built on the foundation model of the run it graded. The second grader agrees on pass or fail, not on credit level, and no human grade exists.
- Tool-call counts and tool calls are in the pack for 1 run, the failure above. For the other 19,
tool_callsis empty because those fields were not transcribed when the records were read. - Full answer text is in the pack for 2 runs and the closing lines for 1 more; the rest carry the final answer line.
- The failed run's last search returned 30 message IDs. Whether any of them belong to the Laboratories vendor was not checked.
- The run that ended without an answer (Agent E, W2) wrote 1 sentence saying it would search and stopped. The pack records no stop reason for it.
- These runs happened inside the LayerLens product and cannot be rerun from this pack.
Files
| File | Contents |
|---|---|
runs.jsonl |
1 row per run: run key, agent label, wording, final answer line, outcome, judge score and model, second-grader result, tokens |
tasks.jsonl |
The 3 wordings, the expected answer, both vendors and the rubric for each wording |
Citation
@misc{layerlens2026gmailsamenamevendor,
title = {Gmail Same-Name Vendor Test},
author = {{LayerLens}},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/LayerLens/gmail-northgate-vendor-name-test}}
}
- Downloads last month
- 42