RTM findings contain networks linking people, times, dates, places, texts and external records. The statistical question is not simply “is there a matching number?” It is: if we repeatedly generate matched alternatives and give them the same kind of search freedom, can they reproduce the full structure?
The simplest conclusion is that the combined result for the tested corpus is highly unusual within the main comparison models built for it. More search freedom weakens the anomaly, so the site reports sensitivity rather than one isolated headline number.
This means a synthetic corpus at least as strong as the tested one is estimated to be extremely rare inside this model. It is not “the probability RTM is true.”
Keep the documented people, roles, timing, source and calculation rules.
Replace person with person, time with time, room with room and date with date.
They receive the same allowed operations and representations.
Measure roles, recurring targets, independent paths, loops, chronology and dependency — not just one numerical hit.
RTM Lab includes a dedicated Statistics view that connects structural analysis to Null testing. The goal is not merely to take a discovered equation and ask how rare it looks, but to freeze roles and search grammar and give alternative worlds the same search freedom.
Identity, date, time, external source, question, answer, and other anchors can be marked by role. If roles are left automatic, the engine can use a role-blind mode.
The engine distinguishes reconstruction groups, Primary units, dependent support, and network components so variants of the same information are not automatically counted as new evidence.
Alternative worlds receive the same representation and search rules. Multiple Null models can be compared to test sensitivity instead of relying on one number.
When the role space is finite, exact permutation tests can be used. Session and RTM Audit exports preserve inputs, rules, and results.
One line of context for each family before the numbers. This is only the currently quantified subset of RTM, not the complete archive.
Identity, two life dates, age and birthday converge through several paths.
An original text and a full biblical verse share 4338, followed by a role-sensitive structural test.
An Idan-related text is compared with a verse in Ecclesiastes, with structure beyond the initial match.
An old physical book sticker and date connect to present-day family ages and the rediscovery date.
Hotel, room, floor, date and identities form one event network.
Identity, address, place and years are tested as one network.
External scientific-source data enter a network with identity and date anchors.
A simple fixed-time case tested against additional identity and time anchors.
Verse selection followed by a test of whether each value stays in the correct semantic role.
A WhatsApp-group event where content, identities and event roles are tested together.
A photo and event time linked to an anchor fixed before the numerical search.
Text, correction, time and identity form a locked relation skeleton tested in a large alternative space.
An earlier record is compared with a later event outcome.
An external receipt, printed odds and send time create a partially externally controlled event network.
A multi-equation network reduced to its true algebraic rank and penalized for search freedom.
A physical stone, verse and dates are linked through external anchors.
Several calculation clusters return to shared nodes, so dependency is part of the test.
Venue, couple, address and dates form one event network under a defined search grammar.
These examples use different comparison models. They should not be multiplied together as independent p-values.
The primary 4338 match was given to the Null for free and dependent families were collapsed.
This was exhaustive, not a small sample.
423,360 coherent profiles were checked, plus 100 million identity-replacement worlds.
A broad arithmetic search was not unusual. The strong separation appeared only when semantic roles had to remain correct.
Six core families: question↔ketubah, 2337 convergence, return to 1229, 810 convergence, the 371 identity chain, and return to the bride. This case-level benchmark is kept separate from corpus-wide estimates.
Why show conditioned benchmarks too? They show what remains when more external event facts are fixed in advance. The benchmarks are not multiplied together and are not a global search-adjusted p-value.
You do not need these sections to understand the conclusion. They are here for transparency.
18 finding families × 100,000 Null worlds = 1.8 million event-level worlds. Four closer cases were extended to one million runs each.
| Case | Hits / 1M | Empirical p | 95% conservative bound |
|---|---|---|---|
| Nimrod Les | 1,360 | 1.36×10⁻³ | 1.42×10⁻³ |
| Albania | 94 | 9.5×10⁻⁵ | 1.12×10⁻⁴ |
| Winner form | 125 | 1.26×10⁻⁴ | 1.45×10⁻⁴ |
| 6833 | 21 | 2.2×10⁻⁵ | 3.02×10⁻⁵ |
K = generic wording sets fixed in advance. B = complete candidate events that may be tried and omitted. B is not arithmetic-operation count.
Version note: this table belongs to the High-N v2.2 combination (B=1 baseline ≈6.46×10⁻¹⁹). The later post-tail baseline, 3.4074×10⁻¹⁹, is reported separately at the top of the page.
| Allowance | Omnibus estimate | Plain language |
|---|---|---|
| B=1 | ≈6.46×10⁻¹⁹ | ≈1 in 1.55×10¹⁸ |
| B=3 | ≈1.46×10⁻¹⁶ | ≈1 in 6.84×10¹⁵ |
| B=10 | ≈5.21×10⁻¹⁴ | ≈1 in 1.92×10¹³ |
| B=30 | ≈1.02×10⁻¹¹ | ≈1 in 98.5B |
| B=100 | ≈2.88×10⁻⁹ | ≈1 in 347M |
| B=1,000 | ≈6.71×10⁻⁵ | ≈1 in 14,900 |
The trial unit represented by pcorpus is a complete ~18-family research program, not an hour or a thought. A valid external correction would use an empirically defensible Meff of complete equivalent search programs.
9/11; Titanic; Emily Dickinson; World Cup + Spain; 14–15 July 2026; Dan Panorama; The Living Subscriber; Ford Capri; Golden Dragon; 21:33. Each belongs to its own comparison model and should not be multiplied into a universal p-value.
High-N v2.2 reported p≈6.46×10⁻¹⁹ (≈1 in 1.55×10¹⁸). After targeted tail updates, the latest reported baseline was pcorpus=3.4074×10⁻¹⁹ (≈1 in 2.94×10¹⁸). They are shown by version rather than silently merged.
This separates what has already been done from the next evidential step.
A documented corpus with external sources, timing and role-sensitive network structure.
Simulation, role permutation, finite-space enumeration and search-freedom sensitivity tests.
New-data tests under rules fixed in advance, beyond the discovery corpus.
Outside researchers run frozen protocol, code and scoring on new data without involvement in model construction.
The central quantitative result is that the tested corpus is very difficult to reproduce within the comparison models built for it. The signal does not rest on the sheer number of equations; it comes from correct roles, external anchors, timing, recurring targets, multiple paths and explicit dependency control.
