Basanos research lifecycle
Methods
Every candidate result is attacked with nulls and controls before a human operator records its evidence state. Questions, tests, outcomes, bounded claims, and novelty reviews remain separate at every step.
Discovery Loop
Comparable units
Reports, cases, observations, signs, events, or rows are separated so the system is comparing like with like where possible.
Shared feature extraction
Local models and structured pipelines tag phenomenology, evidence type, sequence, body-boundary language, communication, light, threat, and context.
Adversarial attack
Candidate findings are checked against literature, fiction/null baselines, curator effects, proxy collapse, controls, and measurement-invariance failures.
Disposition
Most material becomes a bounded null, is downgraded to exploratory, needs verification, or is held. Supported findings require operator review.
Follow-up loop
Open questions, data gaps, exploratory findings, and operator holds drive the next acquisition or review cycle.
Out-of-Sample Confirmation
Validation can only ever rule things out. To find things, the engine is being extended with a predictive layer: a candidate is not “a pattern that survived scrutiny” — it must make a risky forward prediction that then comes true on data the model never saw.
How a prediction is tested
Fit the strongest ordinary baseline (confounds included as covariates), study the structured residual it cannot explain, state a mechanism, and preregister a prediction — hashed before any held-out data is touched. The prediction is then tested on a genuinely independent hold-out (a different lineage, dataset, time window, or instrument). A random split of the same sample does not count.
What it has shown so far
In calibration the engine recovers known effects out-of-sample — flighted animals living longer than body size predicts; established cancer-gene dependencies in held-out tumour lineages — and rejects planted decoys. No new public claim has come from it yet: by design, nothing is promoted without independent confirmation, novelty review, and operator sign-off.
The same operator gating applies as everywhere else. A confirmed prediction is a candidate for review, not a finding — and a confirmed prediction that turns out to already be known is kept as calibration, never published as a discovery.
The Frame Problem — and Testing Claims Against the Body
Here is the newest and, honestly, the most important upgrade. Almost every anomaly Aletheia studies — near-death experiences, sleep paralysis, apparitions — rests on the same kind of evidence: a story someone told after it happened. That is two selection biases fused together before any analysis even starts (who has the experience, and who writes it down), and it makes stories a false-positive factory. No amount of clever statistics on stories fixes a problem baked into the stories themselves.
The only real escape is physiology measured during the state — brain recordings, denominators, individual differences — because the body can’t be re-remembered after the fact. So the engine grew a generative front-end pointed at that data:
Generate, then attack
A bold generator proposes cross-domain, mechanism-level bridges — and every one must name a concrete shared mechanism, a real during-the-state dataset that could test it, and a way it could fail. Anything that can’t be tested against measured data is dropped before it goes further. Then a calibrated skeptic tries to kill each survivor.
Ask the brain data directly
Survivors are taken to the physiology and tested with pre-specified, multiple-comparison-corrected, leave-one-out-checked analyses. A text effect that also shows in the body survives the frame-break; one that vanishes was a retrospective artifact. Every run is logged with its honest result.
What it has shown so far, stated plainly: two candidates have been taken to measured-during data, and neither has confirmed — one showed no signal, the other a robust-but-uninterpretable state-proxy that fell apart under scrutiny. Aletheia has zero positive measured-during confirmations, and that is the honest state; one would be the first proof there’s a way out of the story trap. See the research loop for each test, its result, and what it would mean if it were real.
Question Generation and Review
Aletheia keeps a living queue of tracked research questions. Currently 28 open questions are tracked. A question is never evidence and never a claim; it is a recorded reason to do the next piece of work.
Generate under quarantine
A question generator can re-read open findings, failed hypotheses, and residue from recent work, then propose candidate questions in a schema-enforced format. Every candidate must carry a falsifier, source artifacts, and a duplicate check, or it is rejected at the gate.
Nothing promotes itself
Dream output cannot touch the queue, the claims, or any pipeline job on its own. Candidates wait in a formal delta marked for operator review, and merges are reversible with a recorded decision and a stored preimage. In the first production run the agent proposed 8 candidates; 5 entered the queue after review, and each later disposition stayed on the record.
The generator exists so the platform can propose its own next questions instead of only answering the ones it was given — while keeping question generation under exactly the same operator gating as everything else.
Product Separation
| Product | Plain Meaning | Current Rule |
|---|---|---|
| Question | A reason to investigate | Has no evidence or claim authority. |
| Test | A falsifiable way to answer a question | Must name data, a confound, a falsifier, and a deterministic verifier. |
| Outcome | What an executed test returned | May be supported, null, exploratory, or require verification. |
| Bounded claim | The narrow statement an outcome permits | Cannot exceed the data or controls that generated it. |
| Novelty review | A separate check for prior work | Only a supported, genuinely new result may receive a discovery label. |
Substrate Depth D0-D6
The depth ladder describes evidence state, not how impressive a result sounds. D5 can be bridge-ready without being claim-grade.
- D6_CLAIM_READYClaim-ready substrate
Eligible to support a bounded finding, pending verification and operator review.
- D5_BRIDGE_READYBridge-ready substrateD5 does not mean true.
Structured enough to support synthesis — still not a claim by itself.
- D4_CONTROLSControls and nulls
Matched controls, nulls, or comparisons applied.
- D3_FEATURESFeature-coded
Features coded; ready for within-corpus checks.
- D2_UNITSParsed units
Comparable observation units extracted.
- D1_SCHEMASchema mapped
Fields and structure identified.
- D0_PROFILESource profile
The source exists; catalog/profile only, not claim-ready.
Depth describes preparation, not truth. D5 material can support synthesis but cannot make a scientific claim by itself.
Kill Discipline, Step by Step
Every candidate pattern runs the same gauntlet before anything is cited. Most of what enters does not survive it.
- 1
Candidate pattern
A pattern surfaces from the registry substrate.
- 2
Source & bias check
Curator effects, platform effects, and selection bias first.
- 3
Control comparison
Matched controls; fiction and narrative baselines.
- 4
Null attack
Could the right null produce this anyway?
- 5
Novelty adjudication
Is it already known? Partially known? Actually new?
- 6
Operator decision
A recorded human verdict — never an automatic promotion.
- Supported
- Bounded null
- Exploratory
- Needs verification
Failures recycle into tracked questions, and the loop runs again.
Current public record: 5 supported, 5 bounded nulls, 8 exploratory, and 2 needing verification.
A Worked Example
The Hessdalen analysis is a worked example of a bounded result: dated observations did not show elevated same-day Kp against matched dates, while weather differences remained an observation-ecology result rather than a physical explanation. See the supported findings →
Predecessor Results Stay Traceable
The older Hawkes, LSDP, and eight-mechanism analyses are not erased. They are predecessor framework material that fed later questions. They are not the current public state, and broad motifs are not findings unless they pass the lifecycle gates.