Project Aletheia
Independent research lab · reconciled 23 July 2026
Aletheia builds Basanos — a verification engine that makes cheap, ungameable ground-truth checks (locks), runs a guesser→checker→lock loop, and calls something a discovery only when it survives every honest test. The guesser is a commodity; the lock is the moat.
We point it at the hardest evidence there is — the anomaly corpus (near-death experiences, sleep paralysis, apparitions, UAP reports) — because if the checks hold where wishful thinking usually wins, they hold anywhere. How Basanos works →
These are lifecycle counts from the reconciled public catalog. A supported finding passed its evidence checks; a discovery also needs verified novelty.
The most important thing the lab does
Almost every anomaly — near-death experiences, sleep paralysis, apparitions — rests on stories told afterward, and stories are a false-positive factory. The only escape is physiology measured during the state. The discovery trail is where Aletheia takes a story-based claim and asks the brain data directly: what survived, what died, what it’s blocked on, and what it would mean if it’s real.
Three findings currently pass their bounded evidence checks, but none is a verified discovery. The measured-during physiology branch still has zero positive confirmations.
See the checked research loop →Every result on this site travels the same loop — the funnel above. Sources arrive immutable, get an evidence contract, and supply candidates with their boring twins attached. Tests are frozen before data is read, verdicts are deterministic and LLM-free, and nothing closes without calibration, an operator decision, and an honest ending — claim, useful null, or kill.
The loop can also feed itself: a quarantined question generator re-reads open findings and failed hypotheses and proposes new research questions. None enters the tracked queue without operator review. Surviving questions are deduplicated and ranked, then published as an open list. Read the open research questions →
The registry is preserved source material, not a public claim list. The reconciled catalog separates supported findings, nulls, exploratory outcomes, questions, and blocked tests. Nothing receives a discovery label without a dated novelty review.
Evidence checks passed within the stated claim boundary.
Supported findings with a dated review showing the result is genuinely new.
The test found no supported effect inside its stated boundary.
Interesting signals that do not have claim authority.
Prior claims that must be checked again before public support.
Tracked reasons to run future tests — never evidence and never findings.
Honest disposition is the credibility signal. These rules hold for every result the platform publishes.
A failed prediction remains in the findings ledger as a bounded null instead of disappearing.
An interesting pattern can stay available for follow-up without receiving claim authority.
The Hessdalen claim is about observation ecology under matched controls. It is not a physical explanation of the lights.
A supported finding is not called a discovery until a dated prior-art review permits it.
Three outcomes currently clear the evidence gate. None has cleared the separate novelty gate, so none is labeled a discovery.
In the full-date Hessdalen event corpus, dated observations behave as an observation-ecology result rather than a physical explanation: event days are not elevated on same-day Kp against matched +/-7/14/21/28 day controls, while precipitation and wind-gust metrics are lower than controls. This claim does not explain what the Hessdalen lights are, does not claim weather causes the phenomenon, and does not refute geomagnetic hypotheses generally; it only says same-day Kp is not elevated in this bridge. Weather evidence is Open-Meteo reanalysis-grid data, not direct station instrument logs for every event.
In two raw psychedelic-EEG datasets (ketamine and DMT), apparent increases in signal-diversity complexity (Lempel-Ziv) are largely explained by flattening of the 1/f aperiodic slope; after controlling for the aperiodic exponent, no separable complexity effect remains in these datasets.
Closed-set pseudonymous linkage within DreamBank after obvious content clues were removed.
Validated markers in 2,192 OBERF narratives follow a coarse order in the telling: onset/body markers early, tunnel and light mid-narrative, felt presence after them, return-to-body last; p=1e-4 in both corpus halves under an event-set-preserving null. Early-cluster internal order is not stable, so the supported claim is a coarse gradient, not a step-by-step sequence.
In the public 128-study ganzfeld compilation: pooled excess +376.2 hits over the ~5,459-trial chance expectation (p<1e-5), robust to greedy removal of 46 of 128 studies; decline over time absent at the frozen 0.01 threshold; small-study association suggestive only (p=0.019-0.047). Structure of the compiled file; upstream selection untested; no psi conclusion either way.
Aletheia runs its discovery infrastructure against the hardest evidence domain available. If kill discipline holds here, it holds anywhere.
A domain where selection bias, cultural priming, and wishful thinking usually win.
Nothing on this list is specific to anomalies. It ports to any contested evidence domain.
The corpus is the stress test. The method is the product.
Methods, controls, null results, and the evidence trail behind every finding.
Read the methods →Adversarial verification infrastructure applied to a hard test corpus.
About the lab →Open-source website: public lifecycle catalog, provenance links, and validation gates.
View on GitHub →Findings, questions, tests, and failures stay separate by design. Have a dataset or replication target? Send it for review.