What separates the real from the false?
Six hypotheses, published before any record existed and tested exactly as worded. Only skeptic-reviewed records count; nothing is a result until its threshold is met.
| Hypothesis | Current reading | n / needed | Status | |
|---|---|---|---|---|
| H1 | The size of the leap predicts the fate Chronology anomalies that were later vindicated extended the then-accepted limit modestly; those later refuted claimed much larger leaps. The field accepts deep time in steps.
Test · disproofTest. leap = claimed age (midpoint) ÷ the accepted limit at the time of the claim (`orthodoxy_at_claim.limit_bp`). Compare medians for vindicated vs refuted chronology cases. Would disprove it. once at least 12 resolved chronology cases are catalogued, the vindicated median leap is not lower than the refuted median. Registered 2026-09-23, before any data. |
Vindicated median 1.1×, refuted median 10.8× | 4 / 12 | awaiting data |
| H2 | Independent dating methods predict vindication Anomalies later vindicated more often had two or more independent dating methods, and dates from more than one laboratory, when first claimed.
Test · disproofTest. share of `multiple_dating_methods = yes` (and separately `multiple_labs = yes`) among vindicated vs refuted cases. Would disprove it. at 20 or more resolved cases, the share is not higher among vindicated cases. Registered 2026-09-23, before any data. |
2+ dating methods: vindicated 1/7 · refuted 0/7 | 14 / 20 | awaiting data |
| H3 | One team and closed material predict refutation Refuted anomalies more often rested on a single team and on material critics could not examine.
Test · disproofTest. share of `independent_teams = no` and of `material_accessible = no` among refuted vs vindicated cases. Would disprove it. at 20 or more resolved cases, these features are not more common among refuted cases. Registered 2026-09-23, before any data. |
One team: refuted 4/7 · vindicated 4/7; closed material: refuted 0/7 · vindicated 0/4 | 14 / 20 | awaiting data |
| H4 | The objections that proved wrong were mostly priors Objections later shown to be wrong were disproportionately arguments from the accepted model ("it can't be that old") rather than specific evidential problems such as contamination or disturbance.
Test · disproofTest. share of objections with `kind = prior` among those with `outcome = wrong`, vs among those with `outcome = held`. Would disprove it. at 30 or more objections with known outcomes, the prior-based share among wrong objections is not higher than among objections that held. Registered 2026-09-23, before any data. |
Model-based objections: 25% of those proved wrong vs 9% of those that held (4/16 vs 4/44; Fisher p = 0.19; weak — few cases) | 60 / 30 | consistent |
| H5 | Vindication is slower than refutation The burden of proof is asymmetric: anomalies that turn out real take longer to be accepted than false ones take to be rejected.
Test · disproofTest. median years from `year_claimed` to `year_resolved`, vindicated vs refuted. Would disprove it. at 20 or more resolved cases, the vindicated median is not longer than the refuted median. Registered 2026-09-23, before any data. |
Median years to verdict: vindicated 12 · refuted 2 | 14 / 20 | awaiting data |
| H6 | A proponent's outside stake is a warning sign A documented commercial, religious, nationalist or fame-seeking motive (`proponent_stake = yes`) is concentrated among refuted cases.
Test · disproofTest. count of `proponent_stake = yes` by status. Would disprove it. two or more vindicated cases with a documented proponent stake, or a rate among vindicated cases at least half the refuted rate. Registered 2026-09-23, before any data. |
Outside stake: refuted 2/4 · vindicated 0/6 | 14 / 20 | awaiting data |
| H7 | Many megaliths are dated by association, not directly For most major megalithic monuments, the construction age rests on material found with the monument (charcoal, bone, pottery, inscriptions, style, historical texts) rather than a direct date of the stonework (e.g. rock-surface luminescence, cosmogenic exposure, U-series on carbonate formed after construction). That leaves room for large errors in either direction.
Test · disproofTest. an audit of 40 monuments whose list is fixed and committed before any dating research: widely cited megalithic sites across all inhabited continents. For each: stated construction date, dating basis (direct / associated organic / historical-contextual / stylistic / none), number of independent dates and their spread, and whether a direct method is feasible. Would disprove it. fewer than half of the 40 depend on association alone. Registered 2026-09-23, before any data. |
consistent. After an independent Opus review of the Sonnet audit (19 corrections), dating bases for the 40: direct 1 (Nan Madol, U-Th on coral), associated organic 25, historical-contextual 5, stylistic 6, none 3 — about 35 of 40 rest on association alone (at least 29 even under the most generous definition of direct), far beyond the disproof line of 20. Unverifiable: Almendres, Karahan Tepe, Ġgantija, Ħaġar Qim, and a paywalled 2015 luminescence study of Egyptian monuments. Top leads for a first direct date: Gunung Padang, Almendres, the Great Sphinx, Carnac, Sacsayhuamán. Files: analysis/megaliths/audit_review.json, leads.json. | 40 / 20 | consistent |
H1 · How far each claim reached
Claimed age ÷ the accepted limit when the claim was made. Chronology cases with a sourced limit.
H5 · Years from claim to verdict
Open cases show the years so far.
H2 · H3 · H6 · Evidence at the claim filled yes · half partly · ring no · dot unknown
H4 · Objections by kind and fate
| Kind | Held | Wrong | Open |
|---|---|---|---|
| Layers or association | 10 | 2 | 22 |
| Dating or samples | 9 | 1 | 25 |
| Natural, not human | 11 | 8 | 23 |
| Fraud or provenance | 7 | · | 4 |
| Other | 3 | 1 | 1 |
| Contradicts the accepted model | 4 | 4 | 13 |
Discovery track
Scans of open data for patterns nobody has proposed. Methods are pre-registered; nulls are reported.
| R1 | Coherent residue: do dismissed dates line up? Dates set aside as "too old" are usually noise (contamination, old wood, reservoir effects, lab error), and noise should scatter. If, within a region, isolated too-old dates from independent sites (different sites, labs and materials) cluster at a common age more than chance allows, that coherent residue is a candidate signal of an unrecognized earlier occupation. Method · disproofTest. (1) Data: open radiocarbon databases, p3k14c first. (2) Outliers: at sites with ≥ 5 dates, a date older than every other date at the site by > 3 combined standard errors, with no other date at that site within 2σ of it (an isolated old date, the kind excavators dismiss). (3) Mundane filters, applied before any clustering: marine shell and other reservoir-prone materials; bone dated before 1990 without stated collagen quality control; dates beyond 40,000 radiocarbon years; duplicate or split samples. (4) Coherence: per region, the largest number of independent sites whose outliers overlap in one 2σ age window, compared with a permutation null that shuffles outlier ages across regions within the same continent (10,000 permutations). (5) Threshold: p < 0.01 after Bonferroni correction across regions. (6) Every surviving cluster goes to a researcher and a skeptic: is it already known, and is there a mundane cause? Would disprove it. (the method's value) no cluster survives, or every survivor is already known or mundanely explained — a publishable null for this approach on this database. | null. p3k14c v2022.06 (173,946 dates) → 2,122 isolated old outliers after the mundane filters, in 89 testable regions; none passed the Bonferroni threshold (closest: Russian Federation, corrected p = 0.68 — what chance predicts across 89 tests). Blind spots: compilations may omit dates excavators rejected; a continent-wide pattern would not stand out; Australia was untestable as a single region. Report: analysis/residue/r1_report.md. | tested — null |
| R2 | The frontier scan: dates older than the accepted first arrival If people reached a region earlier than accepted, dates older than the accepted first arrival should recur across independent sites, laboratories and materials — the pattern that preceded the acceptance of pre-Clovis sites. Method · disproofTest. (1) Regions and their accepted first-arrival limits, each with a source, are fixed in analysis/frontier/limits.json and committed before any scan — at least the Americas, Sahul, Japan, Remote Oceania, New Zealand, Iceland, Madagascar and the Caribbean. (2) From p3k14c, select dates older than the limit by more than 2 standard errors (after converting between radiocarbon and calendar years consistently), applying R1's mundane filters. (3) A region is a lead if at least 3 independent sites, with at least 2 laboratories and 2 material classes between them, have pre-limit dates. (4) Every lead goes to a researcher and a skeptic: is the site already debated, and is there a mundane cause? Would disprove it. (the method's value) no region qualifies, or every qualifying site is already known and debated — the scan adds nothing new. | limits fixed and pushed before the scan (be225f4). 77,098 dates in the tested regions → 487 older than their region's limit → 273 dates at 89 sites after R1's filters. Two regions met the rule: Iceland (3 sites; one is a mis-coded record — without it, no lead) and the Americas (85 sites: 56 fossil-fauna localities with bone or unspecified dates; known controversies such as Bluefish Caves, Cactus Hill, Meadowcroft and Old Crow; and 10 unfamiliar sites with old charcoal or wood dates, sent to literature triage). Sahul untestable (limit beyond the 40,000-year filter); no p3k14c rows for New Zealand, Madagascar, the Caribbean or western Remote Oceania. Report: analysis/frontier/r2_report.md. | tested — no genuine leads (4 weak, unpublished-context cases open) |
Notes on our own method
| M1 | Known weakness of this first test In the first catalog round, features are coded by researcher agents who know each case's outcome. Even with the rule to code "as things stood at the time", hindsight can leak into the coding and inflate H1–H6. Test · disproofTest. re-code a random subset blind: a separate agent sees only sources published before resolution, never the outcome. Compare agreement. Would disprove it. (that this bias matters) blind and unblinded codings agree on at least 90% of feature values. | planned for after the first 24 records |