Dead ends that are still open

What it means: One night of public sleep-brainwave data contains a steady 40-hertz signal we could bound but not identify. It sits at the exact frequency people associate with awareness, and it could just as easily be a loose electrode.

The case: One subject-night in the public DREAM sleep-EEG database carries a persistent, regionally structured 40-hertz component with an exceptional spectral share. Forty hertz is the classic gamma-band frequency people link to conscious awareness - so what IS this: brain, muscle, electrode, or gear?

How it died: Seven frozen routes bounded it without identifying it: it is not a simple stationary tone (phase varies, harmonics miss), not explained by the simplest auxiliary-coupled contamination (0 of 3 families passed), not recurring in the subject's six other recordings, and the linked-mastoid analysis arm is itself polluted by abnormal mastoid power. Artifact classifiers trained on labeled corpora topped out at 0.32 macro recall against a 0.55 floor. Flexible decomposition cannot name a physical source without ground truth the release does not carry.

What still stands: A receipt-bound characterization: persistent broadband spatial-mean activity plus a persistent regional contrast with an exceptional 40-Hz share, in one identified file, surviving transient and simple-artifact rivals. Ceiling: description of one recording; no consciousness claim.

What it is NOT: Not evidence of a consciousness signature, because muscle, electrode, and equipment sources were narrowed but never excluded; and not dismissed noise, because it survived the simple-artifact explanations that usually kill such signals.

Next test that could settle it: Obtain the session notes and test the equipment, reference, and stimulation explanations directly, or fingerprint the component against a large source-labeled brain-wave recording component library.

What it means: Out of hundreds of out-of-body verification stories checked by hand, one met the strong documentation rubric, and it involves crackers. One case cannot settle the class. Five independent cases would trigger a new test.

The case: The lab reviewed about 400 narratives and hand-graded 39 passages that looked like verification claims. One report met the strong documentation rubric: it named a specific household detail, an independent witness, and information recorded before later confirmation. This is one documented report in these archives. It does not show that perception left the body.

How it died: In these archives, 36 of 39 graded claims were regex noise, practice narratives, or belief posts, and one medium case had a mundane audio channel. Only one case met the strong documentation rubric among roughly 400 narratives. The exit-model reading was not supported in these archives. This is a bounded corpus result, not a universal verdict.

What still stands: The one strong case, preserved with its structure named: specific, previously-unknown detail; independent witness; info-before-confirmation. Plus the frozen grading rubric that killed the other 38. Ceiling: case documentation.

What it is NOT: Not evidence perception leaves the body, because one household-level case with a possible mundane channel proves nothing; and not a straw-man survey, because we graded every claim in the corpora, not a sample.

Next test that could settle it: Independent re-documentation of any submitted case against the frozen structure rubric.

What it means: The question of what deaf people experience at the edge of death is nearly untouched: existing archives hold a handful of cases and they say something odd - restored hearing - that no theory ordered.

The case: People deaf from birth reporting sound during near-death experiences would be one of the strangest facts in the field; one-sided-deaf people hearing on their good side only would be one of the most deflationary. The archives contain exactly two born-deaf reporters, nine one-sided-deaf narrators - and a pattern nobody predicted: the archive's dominant story is hearing RESTORED, explicitly including the deaf side.

How it died: Base rates: 37 self-identified deaf rows produced an expected bundle count of 0.13, so zero observed is uninformative. The machine coder for all 4,053 hearing-comparison answers failed its blind precision floor on two load-bearing categories (0.60 and 0.70 against 0.80) and only 16 impaired reporters landed codes against a floor of 30 - capped at DESCRIPTIVE_ONLY. On boundary-transition sounds in deaf reporters the archives are simply silent (0 cases).

What still stands: Two verified born-deaf reporters with codable answers; 8 of 9 one-sided-deaf narrators describing restored or both-sides hearing; one hearing person's unilateral transition buzz proving one-sidedness occurs. Ceiling: case notes, not statistics.

What it is NOT: Not evidence of anything paranormal, because expectation and narrative convention can produce restoration stories; and not a null, because the sample is a rounding error - the question has never been powered.

Next test that could settle it: A preregistered interview protocol (already drafted into the lab's prospective program) asking one-sided-deaf experiencers separately about boundary sounds and deep-experience hearing.

What it means: The founding statistic of apparition research divides a carefully collected numerator by an invented denominator. The stories may still be interesting; the famous number is not.

The case: In 1894 the Sidgwick Census of Hallucinations computed that crisis apparitions - seeing someone at the hour of their distant death - occurred 440 times too often for chance. That arithmetic founded a field. The lab re-ran it and audited the Phantasms case books behind it.

How it died: The numerator came from voluntary appeals and personal networks while the exposure denominator was guessed, so the famous ratio answers a counterfactual question no one asked. The documentation-lag audits died at census scale: only 3 identifiable long-lag families against a floor of 10 - the Victorian case books do not carry enough datable long-delay cases to test whether stories sharpened over time.

What still stands: The arithmetic reproduction (the computation itself checks out given its inputs) and a complete 63-case clock census of Phantasms. Ceiling: historical bookkeeping.

What it is NOT: Not proof crisis apparitions are chance, because a broken denominator breaks both directions; and not a slight on the Victorian collectors, because their case work was meticulous - it is the population arithmetic that fails.

Next test that could settle it: A registered prospective collection where the at-risk population is defined before any story arrives.

What it means: Our best consciousness lead survived a brutal review week with its core intact and its wounds named. Each wound is a specific request: better clinical scans, better historical instruments, non-English data, and unusual populations.

The case: The lab's central consciousness lead was pushed through eight fresh routes in two weeks. Three seed wounds are now precisely specified: the clinical-vocabulary instrument REFUTED itself by testing the wrong conjunction - and the corrected scan found real three-component clinical episodes; the pre-1929 archive scan showed only that its own word list found no matches; the intensity-dose question is unpowered because reporters almost never mark vibrations as mild (21:1).

How it died: Per route, honestly: R2 component-swap (catch C1009) with executor-verified counterexamples; R3 exit-lexicon recall unbounded plus chunk-boundary straddle; R4 mild tier 68 and 4 rows against a 200 floor; earlier, the lesion-population zero was uninformative at base rates (expected 0.13 bundles) and the cross-language route never acquired a foreign corpus.

What still stands: Weighted 60 percent of AP vibration rows are episodic somatic narration; the SP-AP profile survives length matching (0.93-0.99); clinical corpora assemble the vibration-paralysis-sound triad under known physiology while the full four-part bundle stays unfound. Ceilings: report text, single coder pending blind second.

What it is NOT: Not a validated phenomenon, because everything measured is text people wrote; and not a dying program, because the review strengthened two results while killing overclaims.

Next test that could settle it: Independent blind coder on the frozen composition packet; then the corrected clinical four-family scan.

What it means: Planet orbit numbers are non-random in their leading digits in a consistent way, and until someone models what each telescope could have detected, that pattern cannot be separated from the way we find planets.

The case: The orbital periods of vetted exoplanets depart from Benford's digit law - overall, in transit surveys, and in radial velocity, all in the same direction. A telescope-independent digit anomaly would be extraordinary; a selection-effect fingerprint would still be a valuable methods result.

How it died: Every branch failed at least one load-bearing selection-control requirement; imaging and microlensing fell below inferential minimums; and direct prior work already covers the broad result. Informative null, correctly refused promotion as a finding.

What still stands: The same-direction raw departures as a bounded observation awaiting its control. Ceiling: descriptive.

What it is NOT: Not numerology confirmed, because selection effects are the favored explanation and remain untested; and not numerology refuted, because the same-direction structure across independent methods is genuinely unexplained at the control level.

Next test that could settle it: The frozen digit test conditioned on any published per-survey completeness model.

What it means: Whether repeating radio bursts drift is unanswerable from what CHIME has released, not because the data is bad but because the exposure bookkeeping that gives counts their meaning is internal.

The case: If repeating FRB dispersion measures drift - together, or per source - that is either plasma physics around the sources or a clock on their environments. Four frozen routes attacked it from the public CHIME catalogs.

How it died: The shared-boundary test collapsed from 22 sources to ONE testable boundary once calendar overlap was demanded (catch C360: burst counts hide overlap collapse); the strongest development pattern was matched by 433 of 5,001 chance worlds against a 50-world pass line; raw-spectrum refits produced zero usable candidates (32 of 39 early spectra failed quality gates); and the released injection table cannot say whether a physical source even enters the catalog or earns the repeater label - the selection function is unpublished.

What still stands: Bounded nulls at every tested scale plus a precise statement of the missing product: no occupied-bandwidth variable, no source-level selection model, insufficient calendar overlap. Ceiling: released catalogs.

What it is NOT: Not a criticism of CHIME, because selection products are hard and internal by default; and not evidence of no drift, because the tests were exposure-starved, not negative.

Next test that could settle it: The frozen boundary test on the first release with a source-level selection function.

What it means: We looked for computation-like structure near black holes properly - planted fake signals first, then the sky - and found nothing the public data could show.

The case: A serious version of a wild question: if black-hole environments hosted computation-like structure, specific archival signatures would appear in high-energy data. The lab built the likelihood machinery (planted bursts recovered, shortcuts killed) and swept.

How it died: Zero spatial survivors among 12 temporal groups in the public screen - and the honest boundary: the archive lacks the sub-day azimuth response and detector-background products a collaboration-grade version needs.

What still stands: A calibrated null over the tested window plus validated sparse-background likelihood machinery. Ceiling: this archive, this screen.

What it is NOT: Not a refutation of the exotic hypothesis, because the decisive instrument products are internal to collaborations; and not numerology, because the screen was preregistered with plants and controls.

Next test that could settle it: The frozen corrected-versus-printed test with real response functions.

What it means: We tried to independently stress-test twelve famous physics puzzles and mostly could not start, because the specific files needed are inside collaborations. This is a shopping list, in public, with receipts.

The case: One day, twelve frozen attacks on headline anomalies: the NANOGrav background, IceCube hotspots, DESI evolving dark energy, the SH0ES ladder, GRB selection, CHIME repeaters, dark-matter direct detection, the neutron lifetime, the proton radius, JWST massive galaxies, strong-lens Hubble tension, muon g-2. Two ran to completion. Nine hit the same wall: the decisive public data product does not exist.

How it died: Item by item, the missing input is exact and documented: no release-bound joint likelihood exposing the three leave-out axes (NANOGrav); a private likelihood build path behind public chains (DESI); unlabeled full-ladder arrays with no row-identity sidecar (SH0ES); no missed-population or injection product (Fermi GBM); no telescope injection/selection model (CHIME); no detector release with radon/time/sideband/efficiency controls together (dark matter); one beam lineage only (neutron lifetime); no apparatus-lineage labels (proton radius); no joint mass/completeness/lensing release (JWST); no cross-family posterior samples (lenses); no machine-checkable dependency graph (muon g-2). The two that ran: IceCube's hotspot came back ordinary; the CHIME repeater sampling contrast stayed numerically strong but its uncertainty sim was not bit-reproducible.

What still stands: A verified map from twelve famous anomalies to their exact missing data products, plus one ordinary-sky IceCube result. Ceiling: audit-readiness, no physics claims.

What it is NOT: Not a criticism of the collaborations, because internal-only likelihood products are normal practice; and not evidence any anomaly is fragile, because the tests never ran.

Next test that could settle it: Per route, frozen and waiting: a deletion self-test on the first release-bound likelihood that arrives.

What it means: Different agencies' earthquake numbers differ a little, mostly without pattern; one timing pattern refuses to die and we are deliberately not believing it until the boring explanation gets its shot.

The case: USGS, GEOFON, and EMSC report the same earthquakes with small systematic disagreements. If those disagreements have structure - by depth, method, region, era - they expose how seismic truth is manufactured. Ten frozen routes chased it, and one association KEPT SURVIVING: a practical timing-agreement pattern held in all four analysis versions of a hand-checked 2019-2025 cohort - and is blocked from promotion by its own honesty rules (rounded timestamps missed the frozen batch rule, one era block reversed direction, one randomization crossed a cutoff).

How it died: Powered routes missed practical floors (depth stopped at 0.395 s against 0.4; magnitude ordering never reached 0.10 units); sparse routes had 2.6 to 15.6 percent power; the reviewed-status comparator had 3 confirmation rows. The surviving association sits at WATCH pending a source-agency and processing-lineage explanation test.

What still stands: Bounded within-floor nulls for several disagreement axes, and one four-version-stable timing association explicitly NOT promoted. Ceiling: catalog bookkeeping structure; the boring explanation (shared processing lineage) is untested and favored.

What it is NOT: Not a precursor or prediction claim, because everything measured is agency bookkeeping; and not a promoted finding, because the frozen rules it missed exist precisely to stop patterns like this from being believed early.

Next test that could settle it: The frozen processing-lineage explanation test on the surviving association, with provider revision metadata.

What it means: We tried to give anti-gravity claims a fair mechanical audit and could not, because the published record does not carry the raw measurements auditing requires. Anyone holding bench data can change that.

The case: Bo pointed the lab at anti-gravity and propellantless-thrust claims with a simple demand: audit them like anything else. The campaign generated 14 routes and found the field's evidentiary floor missing.

How it died: The compiled thrust table's held-out confirmation could completely read one of 21 candidate papers; the rest were blocked, unreadable, or unlocated. Experimental routes lacked the unprocessed channels and row-level trial logs their fixed questions required. The literature route had stalled on anonymous rate limits, but OpenAlex now works without a key and that route can resume. The raw bench data are still missing.

What still stands: A route-by-route map of what evidence each claim class WOULD need: raw force-channel data, trial ledgers, vacuum controls with ion-wind exclusion. Eleven parked cracks remain queued. Ceiling: audit-readiness.

What it is NOT: Not a debunking, because claims were never tested - only their auditability was; and not an endorsement, because a field whose claims cannot be audited from its publications has a serious problem regardless.

Next test that could settle it: Run the frozen single-source force-balance audit on one complete bench package, including air-versus-vacuum and ordinary force controls.