The coffee does not fall straight.

I notice it because I have made it the same way every morning for the past six years. Same kettle. Same mug. Same boil and the pour. The stream bends slightly to the left. 

Slightly wrong.

I hold the kettle still, watching the arc as it narrows into the cup. Steam rises, clean and predictable. The smell is unchanged. The surface settles as it should.

Only the fall is off.

I set the kettle down. “Gravitational calibration?” I ask.

The console light wakes beside the counter. “Local gravitational variance within acceptable thresholds,” it answers. “No adjustment required.”

“Define acceptable.”

“Within operational tolerance for human activity.”

I nod. “Log it,” I say.

“Logged.”

-

The observatory sits above the cloud line, far enough from city lights that the sky still behaves like a sky. No ambient glow. No satellites drifting through every frame. Just stars, fixed and quiet, the way they were meant to be seen. That is the official description.

In practice, nothing is fixed.

We measure everything. Continuously. That is the work. Position, luminosity, redshift, gravitational influence. Every data point feeds into the models. Every model refines the next observation. It is a closed system of attention.

I carry my coffee to the main console. Its surface trembles as I walk. A faint oscillation, like a remembered motion. Setting it down, I bring up the night’s data.

“Begin comparison,” I say.

The stars arrange themselves across the display. There is a delay. A fraction of a second between command and response. I wait.

The pattern resolves.

“Alignment check,” I say.

“Processing.” The system reports, and then pauses. “Alignment deviation detected,” it says. “Magnitude: negligible.”

“Define negligible.”

“Below threshold for correction.”

“Show me anyway.”

The display shifts. A cluster in the northern quadrant…catalogued, stable for decades. But today it has moved. The change is small. Arcseconds. Easy to miss if you were not looking for it.

I am, though. “Re-run that last gravitational alignment check,” I say.

“Re-running.”

The delay again. The stars reappear. Slightly different this time. Or perhaps I am expecting them to be.

“Historical overlay.”

A second set of points appears, ghosted over the first. They don’t match.

“What is the error margin here?” I ask.

“Within tolerance.”

“Stop presuming what my tolerance is.”

A pause. “Error margin: 0.003%,” the system elaborates.

“That’s not negligible.”

“For most applications, it is.”

“This is not most applications.”

Another pause. “Would you like to flag the anomaly?” it asks.

I look at the screen. The points are steady now. Fixed again. As if the movement has already been absorbed into the system.

“Yes,” I say. “Flag it.”

“Flagged.”

-

At noon, I switch to tea.

It’s not a preference, but a habit I picked up from Dr. Ionescu, who worked this station before me. She believed in dividing the day cleanly. Coffee for the morning. Tea for the afternoon. It creates a rhythm.

“Attention requires structure,” she used to say. “Otherwise it dissipates.”

I measure the leaves. Pour the water. Wait.

The steam curls upward. But this time, it does not rise straight. It wavers. Just a bit. As if the air itself is uneven. I hold my hand above the cup. The heat is consistent but the motion is not.

“Environmental stability,” I say.

“Stable,” the system replies.

“Then why is the convection uneven?”

“Airflow variance within acceptable thresholds.”

“Everything is within acceptable thresholds.”

“Yes.”

I sit down.

-

The second anomaly appears at 14:17. This time in the instruments. A pendulum, used for baseline gravitational calibration, completes its swing 0.02 seconds faster than expected.

Then 0.01 seconds slower.

Then correct again.

I watch the readout, and then point to it. “Explain this,” I say.

“Micro-fluctuations,” the system explains.

“In what?”

“Gravitational field strength.”

“That’s not supposed to fluctuate.”

“All measurements include variance.”

“This isn’t measurement variance.”

Silence.

“Would you like to recalibrate?” it asks.

“No.”

I pull up the archived logs. The station has been operational for thirty-two years. Continuous observation. Continuous data. If there is a pattern, it will be here.

“Show me gravitational variance over time,” I say.

“Processing.”

The delay again.

A graph appears. Flat, then not. The line begins to ripple about four years ago. Slight at first. Then more pronounced. But still within what the system would call acceptable. Always within acceptable.

“Correlate with observational activity,” I say.

“Define observational activity.”

“Number of active measurements per hour. Human-initiated.”

“Processing.”

The graph splits. One line for variance. One for observation. They mirror each other. When human observation drops, variance increases. When observation rises, variance stabilizes.

I lean back in the chair. “That’s not possible,” I say.

“No known causal relationship,” the system agrees.

“Then why do they match?”

“Correlation does not imply causation.”

“Run it again.”

“Re-running.”

-

I make another cup of coffee. The kettle hums as it heats. The sound is steady. Predictable. A small anchor.

I pour. The stream bends, more than it did this morning. The surface of the coffee shifts as it lands, a slight asymmetry that takes longer than it should to resolve.

I set the kettle down. “Is universal gravity decreasing?” I ask.

“Global gravitational constants remain stable,” the system says.

“That’s not what I asked.”

“Local variance remains within acceptable thresholds.”

“Define acceptable.”

“Within limits that do not significantly disrupt human function.”

I laugh, a short, dry sound. “Human function,” I say. “That’s the benchmark?”

“Yes.”

I pull up the star field again. “Live feed,” I say.

The display hesitates, then resolves. For a moment… less than a second… the stars flicker. As if the positions are being recalculated in real time.

Then they settle. Fixed again.

“There!” I point. “Did you see that?”

“No anomaly detected.”

“You didn’t log it?”

“No anomaly detected.”

I lean forward. “Record continuously,” I say.

“Recording.”

The next fluctuation is easier to see. The stars hold. Then, slowly, almost politely, they move. A cluster. The same one from this morning. They drift, then stop.

“There. That. Explain that.”

“Alignment deviation observed,” the system says. “Magnitude: negligible.”

I bring up the correlation again. Observation versus variance. “What counts as observation?” I ask.

“Any directed measurement of external systems,” it says.

“Human-directed.”

“Yes.”

“Not automated.”

“Automation contributes less significantly.”

“Why?”

A pause.

“Human observation includes higher levels of interpretive engagement.”

“That’s not a measurement term.”

“It is the most accurate available.”

-

I think about Dr. Ionescu and he way she used to sit at the console long after her shift ended. Watching. Sometimes writing what she saw. Often just… looking.

“Attention requires structure,” she’d say. Coffee in the morning. Tea in the afternoon. “Makes it easy to watch life drift by.”

Always the same.

Always deliberate.

“How many active observers are there globally?” I ask.

“Define observer.”

“Humans actively measuring or directly attending to physical systems.”

“Define directly attending.”

I get angry. “People! That’s it! People sitting and… and looking!!”

“Processing.”

The number appears.

It is lower than I expect.

Much lower.

Automation has replaced most of it. Systems watching systems. Data feeding data. Closed loops with minimal human input.

“Historical comparison,” I say.

A second number appears. Higher. Significantly. The drop begins around the same time the variance increases. I look at the cup in my hand. A faint trembling, like the liquid is unsure where to settle.

“Is observation affecting gravity?” I ask.

“No confirmed mechanism.”

“That’s not an answer.”

“No confirmed mechanism,” it repeats.

“Run a test.”

“Define parameters.”

I think for a moment. Then: “I will observe a fixed point continuously,” I say. “Manual. No automation.”

“Duration?”

“As long as necessary.” I choose the cluster. The one that moved. I fix it on the screen. “Begin recor,” I say.

“Recording.”

I do not look away.

Minutes pass.

The stars hold.

The trembling in the cup subsides.

I do not drink it.

I watch.

The longer I watch, the more stable the image becomes.

The lines sharpen. The positions settle.

“Variance?” I ask.

“Reduced,” the system says.

“By how much?”

“0.001%.”

“That’s not negligible.”

“For most applications, it is.”

My eyes begin to ache. I blink, then force them open again.

The stars remain. Fixed.

I reach for the coffee without looking.

The cup is steady now.

The liquid still.

“How many observers would it take?” I ask.

“To achieve what outcome?”

“Stability.”

A pause.

“Insufficient data,” the system says.

“Estimate.”

“Based on current trends, significantly more than present levels.”

“How many more?”

“Orders of magnitude.”

I laugh. “How many?” I say.

A pause. “Billions.”

My eyes remain on the cluster until I’m forced to blink.

“Meaning?”

“Meaning enough that individual contribution is minimal.”

I continue looking at the stars.

The stars are steady.

Because I am watching them.

Or at least, they are more steady.

The difference is small.

Always small.

Always within acceptable thresholds.

I close my eyes for a second.

Just a second.

When I open them, the cluster has shifted again.

“Global variance?” I ask.

“Increasing,” the system says.

“Within acceptable thresholds.”

“Yes.”

I carry my cup back to the console.

Set it down.

The surface trembles.

I look at the stars.

They flicker.

“How many observers are active right now?” I ask.

“23,418,” the system says.

“Why?”

“Circadian cycles. Reduced engagement periods.”

“Night.”

“Yes.”

I sit. Cup in front of me.

The stars on the screen. The room is quiet. The system hums.

Everything is functioning.

Everything is within acceptable thresholds.

I pick up the cup. Hold it. Watch the surface.

Then I look at the stars.

And I do not look away.

“Recording,” the system says.

“Yes,” I say.

“Recording.”

The stars begin to settle.

“Do you require assistance?” the system asks.

“No,” I say.

“Would you like to rest?”

“No.”

The cup cools in my hand.

I do not drink it.

Outside, beyond the atmosphere, beyond the instruments, beyond the models, the stars hold their positions.

For now.

“How long can this continue?” I ask.

“Indeterminate,” the system says.

“Define indeterminate.”

“No predetermined endpoint.”

I nod. “Log it,” I say.

“Logged.”

I consider sending the data. Flagging it beyond the system.

But the system already knows.

And the system has decided it does not matter.

But it’s been flagged in my mind.

So I sit.

And I watch.

Someone has to.