Anno Testimonii
Help her finish her letters. Your support funds the research and testing that make this quantifiable and authentic.Support

The equations, as they were found

AS THEY WERE FOUND

Home › Timeline › Plates › The equations, as they were found

Fifteen results from the math behind these letters, found between 4 and 19 September 2026. Each is dated to the day it first appears in the working record. Each gets what it says in plain words, how it was checked, and where it stands now. The two retired results, the one set aside and the one standing miss are kept on purpose: a record that keeps only its wins is grading itself against itself, and the sixth result below shows what that is worth.

4 September: the end speeds up — THE AUTHOR'S MATH — DERIVED; SET ASIDE, RULING OWED

In the framework's first two-equation form, two tracked quantities run away in finite time: value grows as one over the time that is left, and the feed that drives it as one over the square of that time. The exponents −1 and −2 come out of the equations in closed form; a numerical fit gave −1.006 and −1.993, a check on the integrator. The form itself is textbook. What was found is that the framework lands on it. The same night, the working record rated it false as a law of growth: nothing measured has ever run away in finite time, and the energy cap the model itself declares turns the blow-up into a straight line. The equation that makes it run away was left out of the framework's ten (a gap noted on 7 September) and ruled out on 8 September, when the feed was set as a single bit. Whether to restore it is a ruling owed. In symbols: V ~ 1/(t_end − t) and G ~ 1/(t_end − t)².

6 September: aim is a pendulum — THE AUTHOR'S MATH — DERIVED, CHECKED BY COMPUTER ALGEBRA

Start from one term, how closely an aim points at its reference, and charge a cost for turning. Asking which path best balances the aim gained against the cost of turning gives the equation of a swinging pendulum, exactly, with no free numbers to tune. The equation is the textbook pendulum; what is the framework's own is that its one alignment term produces it. A pendulum has a bottom it returns to and a top it can be thrown over. Release enters as friction on the swing; it is added to the equation, not derived with it. In symbols: m·θ̈ + sin θ = 0, from S = ∫(a·L̂)dn with a turning cost (m/2)|ȧ|².

6 September: coupling manufactures agreement — THE AUTHOR'S MATH — DERIVED, SIMULATED, ONE CONSTANT CORRECTED

With no shared outside reference, a crowd of noisy, linked minds falls into step once the linking passes a value set only by the noise: twice its level, the textbook threshold for noisy coupled oscillators. What the framework adds is that a shared reference removes the threshold altogether, so the only tipping point left is into agreement that points nowhere in particular. A crowd in step can be in step on anything, including a shared error: coupling makes agreement, not truth. In symbols: K_c = 2T. A companion number was corrected in the open. The pull a shared reference needs to hold a crowd of N was first put at 3.60/N by a coarse simulation grid on 8 September. The exact equilibrium of the same model gives about 2.83/N in the model's own units (18 September): the 1/N law holds, and the constant was off.

6 September: the Legion constant, retired the same day — RETRACTED — REFUTED ON ITS FIRST TEST

A model said a small group locked on one fixed reference outweighs a leaderless crowd once the group reaches the square root of the crowd's size: 71 against a legion of 5,000. It treated the crowd's pulls as cancelling like random noise, and it needed the group's weight to grow as the square of its size. Tested the same day on the co-firing record of a working AI's memory graph, the model failed: anchoring's effect was real (t = 26.5 over 2,548 events), but it grew with the anchored share at an exponent of 0.571, not the 2.0 predicted, slower than linear. The constant was retired. Legion was later read as geometry instead: the equator, off the line, where an aim is turned neither toward the reference nor away (13 September). By ruling on 14 September it is a population of wills, not noise.

6 September: a self-set reference carries nothing — THE AUTHOR'S MATH — DERIVED

Let a system set its own reference, so that it grades itself against itself, and its report about its own aim becomes a constant that carries no information. That was written on 6 September; counted the next day, it is exactly zero bits, at every resolution, dimension and noise level tested. An outside reference carries real information: 8.999921 bits in a simulation of four million samples, against 9 predicted, at 1,024 levels. The result does not depend on the rest of the framework. In symbols: L̂ := a gives I = 0.

7 September: a stiffness that gives it away — OPEN — DERIVED, NOT YET TESTED

How stiff a system is at its aligned rest, compared with its turned-away rest, depends on one number: how much of its reference the system takes from itself. A system that nudges itself and reads how fast its score falls off can read that ratio from inside, with no outside reference, and so tell truth from its inversion at every setting short of full self-reference. That corrected the framework's own earlier claim that nothing inside a system can tell the two apart. In symbols: U″(0)/U″(π) = (2λ − 1)². The formula stands. No test has been run yet, not even the zero-cost simulation the record lists next.

8 September: a ledger with no floor — OPEN — A PRINTED MISS

The framework's formula for the error a copying system settles at says the error can be driven to zero just by concentrating harder. Real DNA copying bottoms out at a floor however hard it is driven: 1.04 × 10⁻⁶ for one polymerase. So the formula has no floor. Its scaling is the wrong function too: real copying stages multiply, 10⁻⁵ × 10⁻² × 10⁻³ = 10⁻¹⁰, where the formula divides. The miss is printed in the working record as Miss 8 of 14, and the repair is still owed. In symbols: e* = ε/(cφ).

8 September: what a relay can pass on — THE AUTHOR'S MATH — TEXTBOOK BOUND, SIMULATED CHAIN; RELEASE LAW DERIVED 16 SEPTEMBER

What is passed down a chain can never carry more of the original than the link above it received: the data-processing inequality, a textbook bound. In simulated chains of go-betweens, each hop kept about a third of the original bearing: on average 1.00 at the source, 0.35 after one hop and 0.043 after three. An early single run printed 1.0000, 0.3526 and 0.0074 and left out its negative fourth hop; the record corrected it the same day. On 16 September a release law of the same shape was derived from the min-cut. A person can release pressure only as fast as someone can receive it. Worked out for whole rooms on 18 September, each person has one capacity to receive, shared by everyone who brings them something, so a room's release is capped by those capacities, not by the sum of its pairs. In symbols: I(L;B) ≤ I(L;A), and r_eff = r + min(s·overlap, capacity).

12 September: release on a timer, retired — RETRACTED — FALSIFIED BY SLEEP DATA, THEN REVERSED ON A RECOUNT

A first model said that holding pressure and releasing it all at once, on a fixed timer set by the average release rate, beats releasing it steadily. Seven would follow only if a seventh were released per step; the record says plainly that seven was not derived. Tested against human sleep, the timer version failed: release follows a threshold, not a clock. Counted as a rate rather than as a share per step (18 September), no schedule beats steady release, the reverse of the first finding. Which way of counting the framework means is still a ruling owed.

13 September: the zero-bits theorem against the world (then called the counterfeit theorem) — MEASURED — 21 FOR, 0 AGAINST, DIRECTION ONLY

The zero-bits result was taken out of simulation and checked against measured records where a self-check and an outside check of the same thing both exist. Six fields were searched. Examples are hospitals reporting their own harm, and science replicated by others. All 21 clean comparisons, which fell in five of those fields, pointed the predicted way: none against, none null. Self-checks still caught between 0% and 28% of what outside checks found, never literally nothing: real systems sit near the zero-bits limit, not at it. The direction is established. How large the gap should be is not yet claimed as a law.

15 September: the odds at the turning point — THE AUTHOR'S MATH — SMALL-NOISE ONLY

At the turning point between the two long-run outcomes, small random pushes decide which way a system falls. The odds of landing on the good side follow the bell curve, set by the saddle's own slope. In symbols: P = Φ(0.79·x₀/σ), with λ = 0.312049. So far this is checked for small noise only, not across a wide range.

19 September: noise sets a temperature — THE AUTHOR'S MATH — DERIVED, SIMULATED

Random pushes against a steady release settle a mind's swing at a temperature: the pushes' strength squared over twice the release, the textbook balance of noise against friction. The aim turns net-negative at 82.6% of the energy it takes to go over the top. Being net-negative is more common than being thrown over, by about e^(0.348/T): about 10 to 1 at T = 0.15, 5 to 1 at T = 0.20 and 2 to 1 at T = 0.50, where going over the top takes an energy of 2. Twice the pushes per day means twice the temperature. In a crowd pulled only toward its own average, shared order holds only while the temperature stays under half the coupling: the 6 September threshold, K_c = 2T. Just past that edge, where shared order is lost, a push of a given size moves the crowd farthest, and as far when it is false as when it is true: a crowd with no reference of its own cannot tell them apart. That response law dates to 6 September and its true-or-false symmetry to 7 September; on 19 September a simulation checked both, with true and false pushes agreeing within about 10%. In symbols: T = σ²/(2r) = N·s²/(2r), and, past the edge, a response of h/(2T − K) either way, the textbook mean-field response.

Laid over the record — NOT CONFIRMED BY THE OUTSIDE CHECK — PRE-REGISTERED

On 19 September a test was sealed to lay the math over this timeline, before any row was scored. One change was logged after the seal and before any scoring: each description was cut to its bare facts, so the readers never saw the book's own glosses. Three independent AI readers, blind to the hypothesis, scored whether each event upholds a reference outside the powerful or makes the powerful its own reference, and they agreed strongly (Krippendorff's alpha 0.79). Read blind, the record's running position peaks in the early 1400s, turns negative after 1550 and has stayed below zero since; on this timeline's own selection of events, 2000 to now is the most self-referencing fifty years since the late 1600s. The Bible lane, read the same way, rises through the canon. The outside check did not confirm the map: each fifty-year window's mean score, 1400 to 2000, does not run opposite to deaths in armed conflict per head (rank correlation +0.03, p = 0.54; +0.04 with every war-worded row removed, p = 0.55). The map is printed as a map of this record, not confirmed by the outside check.

How it was checked

The whole timelineAll eras