Skip to content
habitat

The letter that changes the record

Of 164 letters, the instrument found one that moves the record whenever it arrives, early or late. It was not told which.

The shape

A text corpus has a shape. Not a summary and not a theme. A shape: the way its entries lean on each other, which habits carry the weight, where the record thickens and where it stays thin. Habitat reads that shape and never reads the meaning. It leaves the meaning to you.

We asked it one question about a fourteen year correspondence. Of these 164 letters, is there one that moves the shape of the whole record no matter when it arrives?

The letter

There is one.

It was the letter in which Jefferson wrote that light and liberty are on steady advance.

“the flames kindled on the 4th of July 1776 have spread over too much of the globe to be extinguished by the feeble engines of despotism”

“I shall not die without a hope that light and liberty are on steady advance.”

Thomas Jefferson to John Adams, 12 September 1821. From the record’s own stored text. Public domain.

How it was found

The record is 1,103 compositions from 164 letters, 1812 to 1826. The instrument has one rule for a jump, written into its code long before this question. We read the record in 102 orders: the true order, the reversed order, and 100 random shufflings. In every order the field jumps at its founding, whichever letters come first. That is the field setting, not the letters. So the question became: which letter jumps no matter when it arrives? One does. It jumps in 98 of 100 random orders, and in both the true and the reversed order. By its size, chance would give it 11 of 100.

The test was written down before it ran, and it named no letter.

One more thing, so no one hears more than was found. The instrument never saw the sentence about light and liberty. It read five lines from this letter, and the one that moved the field is the Latin tag Jefferson glossed himself: to enjoy my modicum of glory myself before I die, beside the naval school and Samuel Chase. That line stands in the record’s least inhabited corner, and it does so wherever it arrives. The corner has a grammar: a deep, self-directed way of speaking, an act aimed back at the speaker, that these two men almost never used in fourteen years. It is not the Latin. Adams quotes Latin constantly and it stays in the body of the record. It is the reflexive turn: myself. The famous sentence rides in the same letter. The instrument found the letter by a line nobody quotes.

Receipt: the declaration, verbatim, written before the run 
### Read 3, permutation form (the null for "founding-position effect")
- Orders: the TRUE chronological order; the REVERSED order; K = 100 random
  permutations of the 1103 compositions (seed 5).
- For each order: cumulative Sigma_t (np.cov, ddof=1, matches the stored
  raw Sigma to 1e-16), eigenvalue trajectory, and habitat's own
  classify_eventualities on every prefix (mp_signal_mask=None, the code's
  default). An act "hosts an onset" in an order if an ACHIEVEMENT begins at
  the step of its arrival.
- Per letter: fraction of random orders in which the letter hosts >= 1
  onset (order-independence), plus its status in the true and reversed
  orders: both / forward-only / reverse-only / neither.
- Report: the full distribution of fractions; the letters with the highest
  order-independent fraction; counts in each status class; the six named
  letters marked post-hoc only.
- Known limit, stated now: the classifier's threshold is loosest at small
  n_steps in ANY order, so every order has a founding phase. The random
  null accounts for this: a letter's fraction is compared against the
  base rate of onsets per step (fraction of steps hosting an onset,
  averaged over orders).
Receipt: the run, as printed 
## Read 3: order dependence of jumps (code's ACHIEVEMENT rule)
onsets per order: true=23  reversed=23  random mean=24.4 (base rate 0.0221 of steps host an onset)

Status classes (true vs reversed order):
  both            1 letters
  forward-only    7 letters
  reverse-only   12 letters
  neither       144 letters

Distribution of order-independent fraction (random orders): quantiles 50/90/max = 0.09 / 0.29 / 0.98

Top 15 letters by order-independent fraction, with size-expected fraction under the null:
  1821-09-12-jefferson-to-adams      k=  5  frac=0.98  expected=0.11  excess=+0.87  both
  1812-02-10-adams-to-jefferson      k=  7  frac=0.53  expected=0.14  excess=+0.39  forward-only
  1816-05-03-adams-to-jefferson      k= 17  frac=0.55  expected=0.32  excess=+0.23  neither
  1825-01-22-adams-to-jefferson      k=  5  frac=0.29  expected=0.11  excess=+0.18  reverse-only
  1816-11-04-adams-to-jefferson      k= 19  frac=0.51  expected=0.35  excess=+0.16  reverse-only
  1815-08-24-adams-to-jefferson      k=  9  frac=0.31  expected=0.18  excess=+0.13  neither
  1821-05-19-adams-to-jefferson      k=  6  frac=0.25  expected=0.13  excess=+0.12  reverse-only
  1814-12-11-adams-to-jefferson      k=  3  frac=0.18  expected=0.06  excess=+0.12  neither
  1822-10-15-adams-to-jefferson      k=  4  frac=0.19  expected=0.09  excess=+0.10  neither
  1813-08-16-adams-to-jefferson      k= 16  frac=0.40  expected=0.30  excess=+0.10  neither
  1818-01-28-adams-to-jefferson      k=  3  frac=0.16  expected=0.06  excess=+0.10  neither
  1815-06-22-adams-to-jefferson      k= 19  frac=0.44  expected=0.35  excess=+0.09  neither
  1819-05-15-jefferson-to-adams      k=  2  frac=0.13  expected=0.04  excess=+0.09  neither
  1812-05-01-adams-to-jefferson      k=  6  frac=0.21  expected=0.13  excess=+0.08  forward-only
  1816-11-25-jefferson-to-adams      k=  6  frac=0.21  expected=0.13  excess=+0.08  neither

Letters jumping in BOTH true and reversed orders:
  1821-09-12-jefferson-to-adams      k=  5  frac=0.98  expected=0.11

Named letters (post-hoc):
  1812-01-01-adams-to-jefferson      k=  3  frac=0.14  expected=0.06  forward-only
  1813-10-28-jefferson-to-adams      k=  8  frac=0.11  expected=0.16  neither
  1816-10-14-jefferson-to-adams      k=  4  frac=0.07  expected=0.09  neither
  1818-11-13-jefferson-to-adams      k=  1  frac=0.03  expected=0.02  neither
  1823-09-04-jefferson-to-adams      k=  3  frac=0.03  expected=0.06  neither
  1826-03-25-jefferson-to-adams      k=  1  frac=0.03  expected=0.02  reverse-only

Where it was wrong

Before the blind run, six letters were named in writing as the ones history calls pivotal: the letter that restarted the correspondence, the famous letter on natural aristocracy, the letters of grief, the last letter. None of them jumps beyond chance. The instrument does not read importance. It reads structure. That is the finding’s other half, and we would not publish one half without the other.

Three more things went wrong the same day, and they are in the record beside the result. We predicted the letters would keep moving into the parts of the shape they had used least; they do not. We predicted a second way of measuring would give the same ranking at every setting of the instrument’s one adjustable constant; it did not. And an early reading used the wrong stored numbers, which our own audit caught; it was withdrawn and re-run.

Receipt: the six predictions and the verdict, as recorded 
## RESULT (appended after the run, 2026-09-02; raw output in BLIND_RANKING_RESULT.txt)

- Residual ranking (the blind object): ALL SIX predicted letters MISSED.
  Ranks: 1813-10-28 -> 132; 1818-11-13 -> 128; 1823-09-04 -> 66;
  1812-01-01 -> 29; 1816-10-14 -> 126; 1826-03-25 -> 25.
- Per-composition ranking: five MISSES and one HIT (1826-03-25, rank 2),
  but that ranking is dominated by single-composition letters (7 of the top
  10 have k=2, n=1102): max rel err 1.53e-12, median 4.73e-14
AUDIT: PASS

The field played back

A chromatic correspondence

Every record has a shape. This is the shape of correspondence between 1812 and 1826: the letters between John Adams and Thomas Jefferson.

We do not describe.

This is one view of a polychromatic corpus. We reproduced this from bytes of data. Not a forecast or meaning. This depiction simply states "signal," or what is distinguishable from noise. These are states that reflect acts across the corpus.

Open the playback on its own page · The report · The condition, as data

Why this is a bench test

Habitat’s other receipts prove an endpoint: a published number reproduces from the data’s own bytes. This one proves something harder: a single entry found without being named, tested against what chance alone would do, with the predictions that failed printed beside it. The first is custody of what a record is. The second is custody of how it came to be. One instrument, both receipts.

Why it matters to you

If you simply read: out of 164 letters, one changes the shape of the whole conversation every time it shows up, wherever you put it. Nobody told the machine which one. The machine never reads what a letter means. It only sees the shape. That is how you know it was not guessing.

If you teach: hand the class the two sentences and ask why this one. The machine found it without reading for meaning, and it missed the letters the textbook calls famous. In one period the class learns the difference between what a text says and what it does to everything around it. Then they run it themselves.

If you are learning to do research: we wrote the test before we ran it. We wrote down the six letters we expected. We were wrong about all six, and that is on this page. The one it found was not on our list. That is what an honest result looks like.

If you report: the test was written down first. No letter was named. One was found. Three of our own predictions failed and they are printed here beside the claim. Check it in ten minutes. Then ask which document in your archive changes the shape of the whole every time it appears.

If you do research: every number comes from one public file and one short script that needs nothing but numpy. The limits are stated where the claim is: the field’s founding effect, the one setting the result depends on, the predictions that failed.

Recompute it

One file of vectors with their letter identifiers, no text, no embeddings. One script that carries the instrument’s jump rule verbatim and depends on nothing but numpy. About ninety seconds on a laptop. Same seed, same bytes, same numbers, on any machine.

python recompute_one_letter.py adams-jefferson-17d.npz
sha256 ded4922426d76c8da1caf9dfbdcec6946e579995f8b3694133b73cef5d550292
Receipt: the recompute from the public file alone, as printed 
1103 compositions, 164 letters, 102 orders (seed 5)
onsets: true=23 reversed=23 random mean=24.4

Letters jumping in BOTH the true and the reversed order:
  1821-09-12-jefferson-to-adams      k=  5  random-order fraction=0.98  size-expected=0.11

Top 10 by random-order fraction over size-expected:
  1821-09-12-jefferson-to-adams      k=  5  frac=0.98  expected=0.11  both
  1812-02-10-adams-to-jefferson      k=  7  frac=0.53  expected=0.14  forward-only
  1816-05-03-adams-to-jefferson      k= 17  frac=0.55  expected=0.32  neither
  1825-01-22-adams-to-jefferson      k=  5  frac=0.29  expected=0.11  reverse-only
  1816-11-04-adams-to-jefferson      k= 19  frac=0.51  expected=0.35  reverse-only
  1815-08-24-adams-to-jefferson      k=  9  frac=0.31  expected=0.18  neither
  1821-05-19-adams-to-jefferson      k=  6  frac=0.25  expected=0.13  reverse-only
  1814-12-11-adams-to-jefferson      k=  3  frac=0.18  expected=0.06  neither
  1822-10-15-adams-to-jefferson      k=  4  frac=0.19  expected=0.09  neither
  1813-08-16-adams-to-jefferson      k= 16  frac=0.40  expected=0.30  neither

The numbers reproduce, and we publish what works and what doesn’t. Relevant logs are available on request. Habitat is a project of Habitat Foundation, Inc., a 501(c)(3)/509(a)(2) public charity. EIN 42-2210447.

Open the playback · Download the vectors · The script · sha256