# Session 2026-08-17 — the fold profile, and a day of closed doors

*(Consolidated ledger for the day. Thirteen commits, `a584f19` to `ef32232`, on
branch `opus-try` from `0af5875`. Calibration marked throughout: PROVEN,
VERIFIED by exact computation, MEASURED, REFUTED. Every table in the underlying
notes regenerates from `research/fold-profile-01..16-*.js`, which run in seconds
and never store a tile as positions.)*

## The one-paragraph verdict

The day opened on Chris's question about how a fold's damage distributes inside
the original tile, and it produced eight new proven lemmas, a complete exact
survival curve, and — this is the honest headline — **no progress against the
wall.** Almost every result closed a route rather than opening one, and the two
largest closures were both forced rather than accidental: the survival law is
scale-free in the sifting parameter, and the twin singular series is constant, so
neither a bigger tile nor a favourable sub-population could ever have existed.
The one apparent anomaly surviving to the end of the day turned out to be a
classical result belonging to someone else.

## 1. What is now PROVEN that was not this morning

**Level Ledger.** For every residue a mod p, |h(a) − D/p| ≤ 2·3^{π(x)−1}, so every
copy of a fold loses 2D/p ± 4·3^{π(x)−1} slots, uniformly in k and p. The relative
error is exp(−x(1+o(1))); the bound crosses below the mean at x = 19 and reads
0.048 of it at T₂₃. The counting half of the fold question is closed
unconditionally.

**Mirror Ledger.** K(k) = K(p−1−k) to within exactly 1, the defect caused by the
single slot at W−1 whose representative does not shift under r ↦ W−2−r. VERIFIED
45 of 45 cells, always exactly two mismatched residues with values +1 and −1 at
the two predicted addresses.

**Head Lemma.** On the fold ladder, copy 0's kills below p² lie in {p, p²−2}.
The first is a self-strike (a twin FOUND), the second is the leading edge of the
stratum. Sharpens A6's head-calm audit.

**Impact Lemma (Chris's, in exact form).** Sifting [0,W) in increasing order of
primes, a removal at r = p·t needs t rough below p, so t = 1 or t ≥ p, so r lies
in {p−2, p} or r ≥ p²−2. Hence past p² > W+2 every removal is a graduation, and
the number of folds that can do genuine damage is finite and explicit,
π(√(W+2)) − π(x). VERIFIED exhaustively: 11-tile 64 folds / 64 removals / 0
violations / 0 non-twins; 13-tile 456/456/0/0.

**Second narrowing.** A composite p-rough cofactor is at least p², so r ≥ p³, so
past p > W^{1/3} every removal sits at a prime or a semiprime. VERIFIED, zero
exceptions in 7,056,385 removals.

**Head monotonicity.** A position r < p² is a slot of T_p only if it was a slot at
every earlier level. The head is pure attrition; nothing is ever created there.

**Natal Dispersion Lemma (Chris's misalignment intuition, in exact form).** For
q > p, gcd(W,q)=1 and q > p−1 make k ↦ kW−1 injective mod q on [1,p−1], so the
fold by q removes at most two members of natal@p. VERIFIED, never exceeded.

**The lineage identity.** rarity (y−3)/y# × quality m̄(y) = (y−3)/D_y = the share,
EXACT at every level.

## 2. What is now MEASURED

| object | result |
|---|---|
| per-copy spread | not square-root sized: sd 13.4 against a binomial 741 at T₂₃; hyperuniform in the copy index |
| copy 0's damage shape | zero in the head, ≈2× on the shoulder, flat at 1.000 from u = 4.1; head deficit and shoulder excess cancel to 2.7e−5 |
| kill count vs damage | no information, 48 cells, mean correlation 0.008 at T₁₇ and deeper |
| survival curve, 23# window | exact end to end, 7,952,175 → 895,790, u = 6.13 → 2.00 |
| scale-freeness | S/P depends on u alone; 19#, 23# and an arbitrary 5e8 window agree to four decimals |
| trough constant | e^{2γ}/4 = 0.7931 confirmed a fourth time, by a route that never mentions strata |
| window exceptions | zone has none and never falls below its own expectation (1.13–1.18× HL); narrow windows sit at 0.47–0.58× |
| zone-localized gap | M ≈ 12.2·ln Y, so localizing divides the object by x/(2 ln x) |
| lineage census | 99.94% of twin primes born at level ≤ 13; exact on the tile, and on 440,311 real twin primes |
| lineage yield | proportional to m̄(y), the mean twin-slot gap at the birth level |
| tile anatomy | survival falls monotonically; only the head beats the mean; seams null; no hotspot in 1.65M windows |
| variance deficit | real (synthetic control at 1.00 vs real 0.84–0.96), grows with h, and is not a tile phenomenon |

## 3. What is CLOSED, and why each closure was forced

- **"Fold further and count."** The survival law is a function of u = ln W / ln y
  and of nothing else, so a bigger tile moves W and the target together and buys
  nothing. The tile is a picture of the question, not a lever on it.
- **The natal creation engine.** The natal term is a rounding error in the fold
  law (4.5e−9 at fold 31); the cohort's expected twin content is 3.17 ln²x / x → 0,
  so it is asymptotically EMPTY; and the Dispersion Lemma protects it only out to
  u ≈ x/ln x, which is far above β₂ = 4.26645 and therefore already covered by the
  standard sieve for free.
- **A favourable lineage or window.** The Hardy-Littlewood singular series for the
  pair {0,2} is constant, so all D_x admissible classes carry equal density. The
  tile's whole anatomy is residue-based, so it could not have distinguished
  anything. Both negatives were forced.
- **The sub-binomial dispersion as an anomaly.** It is real, but the
  tile-normalized statistic is the Poisson Fano divided by (1−p) *by algebra*. It
  is the ordinary short-interval variance of twin primes, classical and expected.
- **The pane, a second time.** Chris's point that overlapping windows cost nothing
  for an existence proof removes the pane's original motivation, and A7 had
  already shown it strictly harder than the zone.

## 4. Corrections to my own record, kept visible

1. **"A constant fight."** I framed dial 2 as a constant fight because measured G₂
   sits at 0.6(ln W)² against a budget of (ln W)². What is PROVEN is
   G₂ ≪ (log q)^{4.2665+ε}, so the job is an **exponent** reduction from 4.2665 to
   2, and only afterwards is the constant the fight. Corrected in THE-DIALS §2.
2. **Two detrending bugs in the hotspot sweep**, both of which manufactured
   signal: a 10×-wider centred baseline cannot detrend the origin (produced z up
   to 69.6, all at position 0.000 of W), and clipped flanking windows destroy the
   symmetry that cancels the trend (dispersion 2.50, z = 26.1). Both fixed.
3. **Cofactor primality table too small** in fold-profile-07: it reached only √W
   when cofactors run to W/29, silently calling large primes composite and
   reporting 3.9% semiprimes instead of the correct 66.6%.
4. **An algebraic identity presented as an agreement** between two measurements,
   in fold-profile-16 as first written.
5. **"The one thing still ajar."** I called the variance deficit the session's
   only live loose end. It closed within the hour, as classical.
6. **A wrong mechanism for the tail deficit.** I guessed the tail's low survival
   came from the mirror enriching its slot density. It is ln²y alone; the mirror's
   enrichment is confined to the last ~x² integers and is invisible at scale 1e6.

## 5. What is still live

- **Dial 2, the gap bound**, now correctly priced: bring the two-class Jacobsthal
  exponent from β₂ = 4.2665 down to 2. Since 2 is β₁, the one-class sieve limit,
  the route asks for a dimension-2 problem solved at dimension-1 strength. The
  entire twin prime conjecture is the price of the second residue class.
- **Dial 4, the infinitely-often slack.** ZONE-POSTULATE §2 proves the weak form
  equivalent to TPC while every G₂ route proves the strictly stronger uniform
  form. Nothing in the repo has ever spent this. Cheapest unused knob on the list.
- **M(x, x^k) for fixed k ≥ 3 under the single-alignment recursion.** It dominates
  the zone gap, grows like c·k·ln x rather than c·x, is not frozen (33 to 63 kills
  in [0,p³] against 0 to 2 in [0,p²]), and — the reason to care — the head sits in
  copy 0, which receives only the alignment a = 0. So its recursion is
  single-alignment rather than a maximum over p alignments. That is the first
  place the misalignment principle could be spent rather than cited.
- **Untested:** whether today's variance deficit is the same object as the
  natal-cap campaign's "anchored calm". That concerned slot dynamics and ensemble
  variance, a different measurement. Assumed nowhere; needs its own test.

## 6. Resume point

Branch `opus-try`, clean at `ef32232`, nothing running. New artifacts:
`research/FOLD-PROFILE.md` (§§1–12), `research/THE-DIALS.md`,
`research/window-exceptions.js`, and `research/fold-profile-01..16-*.js`.
Publication moratorium respected throughout: nothing pushed, nothing circulated.