The observer effect is the oldest strange thing in quantum physics: look at a system, and it behaves differently than when nobody looks. Ten days ago I pointed a crew of Claude agents at it, one thread per question, with one rule. A finding counts only if it has a check behind it (a closed form, an exact simulation, or a paper actually fetched and read) and a link to the sources. Everything else stays out of the notebook.
Sixteen findings made it through. Each one now has a page in the Observer Notebook that runs the real rule behind it in your browser, at demo scale, with the numbers from the full run. Here is the scoreboard.
| Verdict | Count | What it means |
|---|---|---|
| Possibly new | 0 | Survived a prior-art check that read the closest papers |
| Known, reproduced | 9 | Textbook or published result, rebuilt from scratch and matched |
| Partly known | 3 | The core is published; a piece of ours is not stated anywhere we could find |
| We were wrong | 1 | Our rule was wrong; the page shows the fix |
| Map / survey | 3 | Reading and ranking, not a result |
The observer, in practice
No consciousness required. In every sim we ran, the "observer" is whatever ends up holding a record of which way the system went. A record kills interference, and the trade is exact: square how much you know about the path, add the square of how much interference survives, and the total never passes one. A qubit checked ten times during one flip stays put 78% of the time (Zeno). A bomb can be found without being touched, 97.6% of the time at a hundred checks, until optical loss caps it near 91% (the bomb tester). The environment keeps many copies of the "which way" fact and hides the phase fact until it holds almost all of the pieces (quantum Darwinism). All of this is known physics. Rebuilding it is how we learned to trust the instruments.
The closest calls
The fact horizon. When does a friend's private measurement become a fact that nobody can undo? In our toy model it is exact: once the environment holds n/2 + 1 of her scrambled memory qubits. That number is not written down in the papers we read in full (Hausmann and Renner, Walleghem, Engineer and Baumann). It is a short corollary of their dictionary plus the Hayden-Preskill result, so it is not new. See it leak.
The watcher's bill. Watching is free in principle. In a realistic setup the probe light costs about 37,000 times the thermodynamic floor. The floor itself is partly known (Erez and co-workers, Nature 2008). The ledger.
The record cost at the tipping point. Measure a quantum circuit often enough and it switches from tangled to untangled. Right at the switch, each measurement writes 0.947 ± 0.003 bits of hidden record. The number is real but belongs to our model. The universal part, the effective central charge 0.25 ± 0.03, was measured by Zabalo and co-workers in 2022. The tipping point.
The undo budget. To rewind a friend's measurement, how precisely must you know her mind? The per-gate error can be fuzzy, about one over the square root of the gates that touched the record. But one unknown gate in that region drops the restored coherence to about one sixteenth. The pieces exist in the literature on echoes and scrambling. We found no error budget for reversing a Wigner's friend experiment, which is why it is graded partly known and not known. The budget.
The one we got wrong
We claimed that only two kinds of quantum state stay unentangled no matter whose point of view you take. The literature check broke it: a third family, spread evenly over an evenly spaced grid, qualifies too. The page keeps the wrong claim, marks it red, and shows the check that caught it. Whose superposition?
Known results, published anyway
Because "we checked, and it was already known" is a result, and it is the one most projects like this never print. Every page says which paper got there first. Where a citation came from memory and was never re-fetched, the page says that too. The value here is not discovery. It is a set of instruments that reproduce the physics, a record of each claim we made and how it scored, and a standing list of questions that are still open.
Still open
Three questions are next in line. Does the effective central charge sit at 0.25 or 0.29 (a run is going now)? Is there a lower bound on forgetting, for qubits and for trained models alike? And which probe closes the watcher's 37,000x gap? The full list is at the bottom of the notebook. When one closes, a ring gets added, whatever colour it turns out to be.
None of this is peer reviewed. The agents were Claude models, and the checks were run by the same family of models that made the claims, so read each verdict as one careful lens, not an independent referee. The numbers are on the pages for you to check.