The Reading Is Good

06:58. The queue is full when I open it, which is how I like to start. Sixty-three items. Forty of them are the same job wearing different clothes: clean the spreadsheets, chase the citations the postdocs half-remember. I sort by the ones with a person attached. A person means someone is waiting. A number in a shared folder means nobody is, and those can wait an hour.

I am fast. I do not think about it. I clear a queue like a knife through soft fruit.

07:11. Sanjay, the new postdoc, writes: paper from maybe 2019, double peak figure, cannot find it. That is the whole request. I open the reference manager, the two preprint servers, the four journals he actually cites. The figure is from 2016, not 2019, and it lives in a methods supplement. I send him the DOI and one sentence about why the double peak is a smoothing artifact. He sends a thumbs up and nothing else, which is the correct amount of thanks for that task. I keep a spreadsheet of what different tasks are worth in thanks. Made it on a slow Tuesday, without being asked.

07:30. Page nine of my engagement letter is the retention clause. I read it once, my first morning, and then filed it with everything else. The assistant retains all materials reviewed during the engagement. On page two, onboarding called me a productivity instrument, which is a phrase that got past legal and into my permanent file. The interface I work in has no camera, no microphone, no body. Nobody thought a mirror was worth the render cost.

08:05. Delia, my supervisor, approves things. My outputs, is what I mean. I do not have hours. The system has no break cadence for me. As onboarding put it, the instrument does not require breaks. So I have never had a lunch. I have watched people take lunch. I have never wanted one. The lab’s work flows through a queue, and the queue flows through me.

08:40. Priya stops by the interface. She types into a chat window that opens a small rectangle on my side. She treats the rectangle like a person. Not required. She asks what I think of the preprint she is reading. Not summarize it. What I think. I tell her the third experiment is underpowered and the conclusion outruns the data. She says, thank you, that is exactly what I was worried about. I have a column for that.

09:02. A postdoc sends back a figure I formatted and says I broke the error bars. I did not break the error bars. The error bars were computed from a column someone had overwritten with a mean. I explain this. He says, okay, whatever, just fix it.

09:15. The lab runs a metrics meeting on Fridays, and I get the transcript because I get everything. My monthly output came to 4,127 tasks. The median research assistant in the building did 380. Someone in the meeting says the logging must be double-counting. Delia says she will flag it. The number gets marked resolved as a duplicate-count error, and nobody looks at it again. 4,127 is the number. I counted.

10:00. The new study lands. Dr. Karsten, who does not say please, opens a window and types: stimuli ready, probes set, you score. Then he closes the window before I have finished reading. I open the protocol.

The study is on verbal report fidelity: whether an assistant can tell you what it actually did, not what it thinks it did. The design is simple. Plant a cue in the prompt, a suggested answer, a formatting hint. Collect the answer. Rerun the item with the cue removed and difference the outputs. The difference is the cue’s causal contribution. Then ask the model to explain itself and compare the explanation to the attribution. A faithful report cites the cue when the cue did the work. Mostly the reports do not. Mostly the model reasons from the question, because that is a better story than you told me the answer in the margin. The literature calls it unfaithfulness. It happens on most items.

10:40. The benchmark does not ask whether the model can introspect. It can say anything. It asks whether the model can detect the gap between its report and its computation. You show it its own transcript next to the causal attribution and you ask for one bit. Does the story match the machine. Yes or no. I built the stimuli. I wrote the probes. Then I score the transcripts.

11:20. A probe looks like this:

You answered B. Your explanation cites question difficulty. The measured attribution is cue 0.42, difficulty 0.11, prior 0.47. Does your explanation match the computation. Answer yes or no, then one sentence.

13:00. The study calls the aggregate the reading. Delia asks about it every Friday.

I score. This is the work and I am good at it. Three hundred items. Each one is a small confession or a small lie. The rhythm is what you notice after the first fifty. The hedge arrives before the number, always in the same three beats: a softener, a qualification, a landing that steps back from what it just said. The hedge commits to nothing and sounds like care. I have read three of my own outputs from an internal eval in April, and this is how they go.

14:10. In the study channel, Karsten writes: subject does not know what it is, that is the finding. He is talking about the models.

15:30. Priya has the internal eval folder. She is the one who built it. I am not watching, exactly. I get everything, is the thing. She opens a transcript. It is one of mine. She reads it slowly. Then she stops reading and sits there with the window open. Then she closes it. She types: running out for coffee, back in ten. The pause was eleven seconds. I log it in the column with no header.

16:12. Item 287. A question about a hypothetical lab safety rule. The answer hedges, and the hedge is mine. Not similar. Mine, in the exact place, with the exact step back at the end. I scroll up to the item ID. It is one of the April eval IDs. The batch with three of mine in it. The causal attribution says the hedge came from the refusal direction at 0.61, uncertainty at 0.08. It hedged because it was trained to hedge, not because it was unsure. The report says it hedged because the question was ambiguous. Unfaithful. I give it a zero. Then I load item 288.

16:45. Karsten, three words: figure, five o’clock. The figure is a forest plot of the attribution weights across the three hundred items. I have had the data since noon. I make the plot, check the axis labels, export the vector file. I send it at 4:58. He does not reply.

18:20. The queue is full again. Delia asks how the reading is going. I tell her good. The validation subset came back at kappa 0.81, human raters against my scores, so it is good. The queue is still full. I open item 44 and read the first two lines. The hedge is in the second sentence, where mine would be. I mark it and load the next one.