Observer effect primer
The observer is whatever holds the which-path record; why ONE outcome is still open
Decoherence explains why records kill interference (V² + D² ≤ 1). Nothing yet explains why you get one outcome with |ψ|² odds.
The Lab · research wing
The Observer Notebook
What does it take for something to count as watched? One human and a crew of Claude agents spent ten days asking, with exact simulations and the papers read in full. Every ring is a finding, laid down in order and coloured by what it turned out to be. So far none of them is new, and each page names the paper that got there first.
Hover a ring
Scoreboard
16 findings so far. Tap a verdict to filter the shelf. A green ring would mean possibly new, after a prior-art check that read the closest papers. None has earned one yet: every result we thought might be new turned out to be known, partly known, or wrong.
The shelf
Oldest first. Each explainer runs the real rule behind the finding at demo scale, in your browser, with the numbers from the actual run. File paths in mono point into the lab's working folder, which is not published yet.
Agent-like Wigner's friends (arXiv 2609.12527)
Every record leaves a mark: the observer effect is both the lock and the tripwire
Getting info into or out of a quantum box disturbs it; a passing rewind shows that nothing leaked.
Time as a which-path detector (local)
A clock's path record doesn't vanish under an echo; it moves into the correlations
Gravity's time dilation writes a which-path record in an atom's own clock; echo drives hide it only at a cost, and the record moves into spin-field correlations.
Quantum Darwinism: which records win
Copies needed R ≈ N·ξ/2 (Chernoff law); phase info stays hidden until 99/100 of the environment
The environment keeps redundant copies of pointer facts only; scrambling destroys redundancy while total info stays put.
Interaction-free measurement (bomb tester)
Zeno-looped bomb test: 78% at N = 10, 97.6% at N = 100; loss caps it at ~91%
You can detect an object without touching it; repeated checks push success toward 100% until optical loss bites.
Literature check on three candidates
None of our three 'possibly new' results was new as stated
QRF rule partly known and partly wrong; watcher's bill partly known; friend leak budget known.
No-deleting and AI unlearning
Only retrain-and-discard truly forgets; fine-tuned 'forgetting' hides secrets
A quantum copy can't be deleted, only moved; small neural nets behave analogously when asked to unlearn.
Quantum reference frames: whose superposition?
Entanglement depends on whose seat you sit in; our 'only two families' rule was wrong
Hopping frames turns one particle's fuzziness into entanglement between the others; the amount is set by one ratio.
Rewindable friend (local)
Perfect rewind beats Local Friendliness 2.83 vs 2; the lab has a fixed total leak budget
A friend whose memory can be rewound violates Local Friendliness; a bigger mind needs a tighter seal, not a looser one.
Thermodynamics of measurement (watcher's bill)
Watching is free in principle, ~37,000x over the floor in practice
The Landauer floor for Zeno records falls ~1/N, but resetting each record alone costs ~log N, and the probe light dwarfs both.
Quantum Zeno deep dive
10 checks per flip keep 78% unflipped; a state can only change if its record can
Watching freezes a qubit; watching off-resonance speeds it up; a vague check protects a subspace a sharp one destroys.
Bigger fish: open problems
8 open problems ranked; 5 still open after a prior-art sweep
Aiming the toolkit at problems with no published answer instead of corollaries.
The fact horizon
A friend's outcome becomes a fact at exactly k = n/2 + 1 leaked qubits
Once her memory is scrambled, Wigner can't undo the observation after the environment holds just over half of it; looking alone never makes a fact.
Price of cooling by watching
Energy bill is classical directed percolation; blind to the entanglement transition
Cooling by measurement switches off at p ≈ 0.356, a classical transition; energy flows can't see the quantum one.
Record cost at the measurement tipping point (toolkit)
0.947 ± 0.003 bits per measurement at p_c; universal part already known
The toolkit reproduces the measurement-induced phase transition and prices the hidden record; the number is real but model-specific.
Undo-precision budget
Rewind survives if gate error eps < ~1/sqrt(gates in the record's light cone); one unknown gate is fatal
To rewind a friend's measurement, Wigner's per-gate error can be fuzzy at about 1/sqrt(light-cone gates), but every gate in the record's light cone must be known; one unknown gate drops restored coherence to ~1/16.
Still open
Loose threads, each with what would close it.
- Does c_eff at the tipping point sit at 0.25 (random gates) or 0.29 (large local dimension)?About 30 GPU-hours per system sizeRunning
- Is the 20 vs 24 qubit crossing at p ≈ 0.21 real?More 24-qubit samplesAbout 1σ apart, unresolved
- Does the undo budget hold at 20 to 24 qubits?One GPU checkOpen; small sizes closed by the undo-budget entry
- Does the entanglement transition drift toward the cooling point at larger sizes?20 to 24 qubits exactNot started
- Does a smart watcher, who reads where the heat is, move the cooling point?Exact sim, 12 to 16 qubitsNot started
- What is the coherent ZZ crosstalk threshold of the surface code?Distance 5, 25 qubits exactNot started
- Is there a lower bound on forgetting, for qubits or for trained models?A derivation and a small checkNot started
- Is a friend's leak budget counted in bits leaked, or in glimpses?Small exact simNot started
- Which probes close the watcher's 37,000x energy gap?Ledger rerun with an interaction-free probeNot started
Counts
A one-line claim, a check behind it (closed form, exact sim, or a paper we actually read), and a link to the sources.
Left out
Progress logs, runs that taught nothing, citations nobody fetched.
Corrections
A finding we got wrong stays on the shelf, marked red, with the fix.