Fabian Arevalo

Computational research note · Combinatorics on words

Two settled cases of the circular repetition threshold conjecture

A second session finds the right encoding: a new binary code and pumping argument settle the 6-letter case outright, and — flagged as a likely overlap with a hard-to-read 2025 paper — the 5-letter case too. The 4-letter case stays open, and is now certifiably empty in this encoding as well.
August 3 & 6, 2026 · code, data & certificates in the companion repository.

Start here — no maths required

Write a long word using only a handful of different letters — say four: A, B, C, D. One rule: never let a short pattern repeat too many times in a row (not "ABABABAB", not "AAAA"). For each alphabet size there is a precise, known line for "how repetitive is too repetitive", and it has been proved you can write words respecting it forever, in one direction.

Now bend the word into a necklace, so the last letter loops back to the first, and ask the same question for every possible necklace length at once. Repeats can now hide around the seam, which makes the question much harder. For necklaces made from 4 to 44 different letters, nobody knew whether you could always stay under the line, at exactly the quality that works for an ordinary, one-directional word. (44 letters and below was open; 45 and up was settled in 2020.) A first session here built a general tool for deciding this, given one special "growing rule" for a given alphabet size, and then proved that the one family of growing rules simple enough to search by hand cannot work for alphabets of 4 to 9 letters — a real dead end, not just an unsuccessful search.

This second session found a different growing rule, one level of abstraction up. Instead of growing the necklace's letters directly, it grows a hidden two-symbol code that decodes into the necklace — a trick borrowed from an encoding of Pansiot's. Under that trick, growing rules exist for 6-letter and 5-letter necklaces, and each is checked, by finite computer search, against a five-part certificate strong enough to guarantee it works forever. The 6-letter case is a new theorem: circular threshold words exist at every length 39, 819, 17,199, … — an open case of the conjecture, settled. The 5-letter case comes out the same way, but with an asterisk: a hard-to-read Russian-language paper from December 2025 may have already found this case by a different route, so it is reported as a likely rediscovery until that paper can be read. The 4-letter case — the one place the ordinary threshold breaks its usual pattern — still has no growing rule of this kind either, now checked far deeper than before.

Case
Growth step
This session's actual certified generator, decoded live in your browser and grown by repeated substitution. Underneath, the growing rule acts on a hidden 2-symbol code (21 bits long for both cases here); what you see is that code decoded into the real necklace letters (0–5 or 0–4), via the same decoding rule proved in section 7. Ring of digits at j=0 (the seed), bar chart at j=1–2 (too many letters to label individually); faint lines mark 21-bit code blocks. The growth and decoding both run live here; freeness of the result was verified separately, offline, against the definition (table in section 9) — this figure does not re-run that check.
Which necklace lengths work, computed exactly by SAT search (not a model). Each strip is one alphabet size; a filled cell means a necklace of that length avoiding over-repeats exists, a gap means an exhaustive search found none. Hover to read off any length. Short lengths are erratic everywhere; long lengths mostly work — except at 4 letters, where the pattern looks finished after length 113, then two further gaps surface at 147 and 154. A sweep that stopped anywhere in between would have reported a clean result and been wrong. The witnesses behind this map are more than a record: session 2 mined them for seeds (section 9).

What this note adds

  1. Two open cases settled. CRTW(6) = RT(6) = 6/5, an open case of the Currie–Mol–Rampersad conjecture, proved via a certified binary code and a 39-letter seed that pumps to lengths 39, 819, 17,199, …; and CRTW(5) = RT(5) = 5/4 by the same machinery — flagged as a likely rediscovery pending a read of a hard-to-parse December 2025 paper.
  2. The machinery that did it: an exact transfer lemma between a necklace and a hidden binary code of it (Pansiot's encoding), a five-part finite certificate that a code substitution preserves the no-over-repeats property, and a pumping lemma inside the code — three new elementary results, each proved here.
  3. Re-derives the known 3-letter case end to end through the new code machinery too, as a second positive control, independent of the first session's route.
  4. A genuine impossibility theorem (session 1) for every alphabet from 4 to 9 letters in the original encoding, plus a new certified-empty result (session 2) for the 4-letter case in the code encoding as well — two independent searches, same alphabet, same dead end.
  5. The exact map of which necklace lengths work, for 3, 4, 5 and 6 letters — including the late-surprise gaps at 4 letters shown above, whose witnesses turned out to be exactly the seeds the new theorems needed.

Everything below is the technical version.

AbstractSummary of results

For circular words, define CRTW(n), CRTI(n), CRTS(n) as the infimum of exponents β such that circular β-free n-ary words exist at infinitely many lengths, at all sufficiently large lengths, and at all lengths, respectively. Currie, Mol and Rampersad conjecture CRTI(n) = CRTW(n) = RT(n) for all n ≥ 4, where RT(n) is the ordinary repetition threshold; Mol–Rampersad (2020) prove this for n ≥ 45 and report it open for 4 ≤ n ≤ 44 (weak form) and for every n ≥ 4 (intermediate form) — secondary citation, see section 1.

We prove a pumping lemma (Lemma A): a q-uniform morphism that preserves α⁺-freeness carries the refined class S₂(n,m) into S₂(n,qm), which reduces CRTW(n) = RT(n) to two finite searches — a morphism, and a seed (Theorem C). Applied at n = 3 this re-derives the known CRTW(3) = RT(3) = 7/4 end to end, as a positive control: a 20-letter seed grows to lengths 20, 560, 15,680 and 439,040 under one 28-letter substitution, each verified directly against the definition. Attempting the same route for 4 ≤ n ≤ 44 fails, and we show why: the natural shift-equivariant ansatz forces a difference-set bound (Proposition N) that combines with an exhaustive computation (Result L) to prove Theorem N′ — no shift-equivariant q-uniform morphism has a threshold fixed point, for any q, at any alphabet size from 4 to 9. Separately, we compute the exact circular threshold spectrum C(n) — which lengths admit a threshold word — for n = 3, 4, 5, 6 by exact SAT search, each witness independently re-verified; at n = 4 the spectrum is not monotone in the naive sense, with two isolated exceptional lengths (147, 154) appearing after an unbroken run of 33 realizable lengths. No case of the Currie–Mol–Rampersad conjecture is settled in Part I.

Part II (session 2026-08-06) passes to Pansiot's encoding of n-ary threshold words by a binary code, where morphisms need not be shift-equivariant over Zn. We prove an exact transfer lemma (Lemma T) between period structure in the code and in the decoded word, a finite certificate (Theorem MC) that a uniform binary code morphism preserves code-freeness, and a pumping lemma in the code (Lemma PC). Instantiating this machinery: Theorem P6 proves CRTW(6) = RT(6) = 6/5 via a certified 21-bit generator and a 39-bit seed mined from Part I's own spectrum witnesses, settling an open case of the conjecture (openness secondary-sourced); Theorem P5 proves CRTW(5) = RT(5) = 5/4 by the same route, flagged as a probable overlap with Tunev (arXiv:2512.24581, Dec. 2025, Russian — secondary, unread) — see section 10; Result P3′ re-derives the known CRTW(3) = 7/4 as an independent control; and Result N2 certifies the identical ansatz empty at n = 4 for all pooled monodromy classes through 21-bit blocks of length ≤ 46 and a two-level substitution engine, on stated ranges. Two open cases of the Currie–Mol–Rampersad conjecture are settled in Part II, one of them (n = 5) provisionally, pending a primary-source read.

Section 1The problem, and the definitions used here

Let Σn = {0,…,n−1}. For a finite nonempty word u, per(u) is its least period and exp(u) = |u|/per(u). For rational α > 1, u is α⁺-free if every factor of u has exponent ≤ α. The repetition threshold RT(n) is the infimum of exponents β for which arbitrarily long n-ary words with no factor of exponent ≥ β exist: RT(2) = 2 (Thue), RT(3) = 7/4 (Dejean), RT(4) = 7/5 (Pansiot), and RT(n) = n/(n−1) for n ≥ 5 — Dejean's conjecture, completed by Currie–Rampersad and Rao (secondary). Words with all exponents ≤ RT(n) are n-ary threshold words; they exist at every length.

A circular word of length m is w ∈ Σnm with indices read mod m; its factors are the factors of wω of length ≤ m. w is circular α⁺-free if every factor of wω of length ≤ m has exponent ≤ α. For K ≥ 0, w lies in SK(n,m) if every factor of wω of length ≤ m+K has exponent ≤ α; S₀(n,m) is exactly the circular α⁺-free words of length m, and SK ⊆ SK−1 ⊆ … ⊆ S₀. The object computed in section 2 is the circular threshold spectrum C(n) = {m ≥ 1 : some w ∈ Σnm is circular RT(n)⁺-free}.

Encoding. A factor of exponent > α whose least period is p has, as a prefix, a factor of length L(p) = ⌊αp⌋+1 with period p. Hence w is circular α⁺-free iff for no p ≥ 1 with L(p) ≤ m and no i ∈ Zm do all of w[i+j] = w[i+j+p], 0 ≤ j < q(p) := L(p)−p hold (indices mod m). For α = n/(n−1) this gives q(p) = ⌊p/(n−1)⌋+1; in particular q(p) = 1 for p ≤ n−2, i.e. any two positions at distance ≤ n−2 in an n-ary threshold word carry different letters — a fact used repeatedly below. All arithmetic is exact (integer / rational); no floating point occurs anywhere in the critical path.

Why this object, and what is reported open

The reported state of the art (secondary throughout this paragraph): CRTS(k) = (⌈k/2⌉+1)/⌈k/2⌉ for k ≥ 6 (Gorbunova, EJC 19(4), 2012), CRTS(4) = 3/2, CRTS(5) = 4/3 (Currie–Mol–Rampersad, arXiv:1803.08145, EJC 26(2), 2019); CRTI(3) = CRTW(3) = RT(3) = 7/4. Mol–Rampersad, The weak circular repetition threshold over large alphabets, arXiv:1912.11388, RAIRO-ITA 54 (2020), prove CRTW(n) = RT(n) for n ≥ 45, and state that CRTW(n) = RT(n) is open for 4 ≤ n ≤ 44, and CRTI(n) = RT(n) is open for every n ≥ 4 (secondary — this exact sentence was seen only in a search summary).

Two elementary observations tie C(n) to the conjecture, since a circular RT(n)⁺-free word has every exponent ≤ RT(n) < β for any β > RT(n):

Observation 3

If C(n) is infinite then CRTW(n) = RT(n). If C(n) is cofinite then CRTI(n) = RT(n).

(Both are one-line consequences of the definitions; the converses are false in general.) So exhibiting an infinite C(n) for a single n ∈ [4,44] settles an open case of the conjecture — that was Part I's target, and it was not reached there. It is reached in Part II (sections 7–9), by changing the alphabet the morphism acts on rather than relaxing the ansatz further.

The sandbox had no access to the literature in either session. arxiv.org, oeis.org, erdosproblems.com, mathoverflow.net and api.crossref.org all refused connections with HTTP 403 at the egress proxy; only a web-search tool and pypi.org were reachable. No primary source was opened at any point in Part I or Part II. Every attribution on this page is therefore secondary and must be checked against the actual paper before anything here is submitted anywhere. The mathematics of sections 2–9 is self-contained and does not depend on any citation; the novelty and "still open" claims do — and section 10 records a prior-work risk (Tunev, Dec. 2025) found only during Part II's search, which revises the novelty picture for n = 5 specifically.

Section 2The circular threshold spectrum

C(n) was computed by exact SAT search over the clause set of section 1, one instance per (n,m). Every satisfiable instance's witness was re-verified by an independent O(m³) checker written from the definition and sharing no code with the encoder; a further vectorised checker was cross-validated against the naive one on 700 random words with zero mismatches.

Positive control. For n = 3, α = RT(3) = 7/4, where CRTI(3) = CRTW(3) = 7/4 is known (secondary): a circular 7/4⁺-free ternary word exists for every 1 ≤ m ≤ 300 except m ∈ {5,7,9,10,14,16,17,22} — cofinite, exactly as CRTI(3) = 7/4 predicts, which calibrates the whole pipeline. A second calibration, against different published constants: setting α to the reported CRTS(k) rather than RT(k), every length 1–60 is realizable for (n,α) ∈ {(4,3/2), (5,4/3), (6,4/3), (7,5/4)} — all four match.

nα = RT(n)range completed#exceptionallast exceptional m
37/41…300822 — {5,7,9,10,14,16,17,22}
47/51…16487154 — 85 values ≤ 113, then 147 and 154
55/41…3004163
66/51…3003359

CERTIFIED on the ranges shown. Full lists and every witness word: data/spec_n*.csv. The n = 4 range is shorter because its instances are much harder — the sweep was still deciding m = 165 when the session ended.

The one genuinely surprising datum. At n = 4 the spectrum is not monotone in the naive sense: after the last small exception at m = 113 there is an unbroken run 114…146, then two further isolated exceptional lengths, m = 147 and m = 154 (shown in the figure at the top of this page). Because these are UNSAT verdicts — the one class of claim here with no witness to check — both were re-decided by three independent SAT backends, together with their satisfiable neighbours as a control:

mCadicalGlucoseMiniSat
113UNSAT 37.0 sUNSAT 49.7 sUNSAT 35.7 s
146SAT 1.9 sSAT 0.4 sSAT 2.5 s
147UNSAT 208.8 sUNSAT 343.3 sUNSAT 207.0 s
148SAT 1.5 sSAT 0.8 sSAT 1.1 s
153SAT 1.4 sSAT 0.7 sSAT 1.2 s
154UNSAT 248.8 sUNSAT 632.1 sUNSAT 298.0 s
155SAT 2.5 sSAT 0.7 sSAT 7.5 s

No disagreement among the three solvers, and every satisfiable model was re-verified against the definition. The two gaps also take two to three orders of magnitude longer to decide than their neighbours — genuinely hard instances, not solver artefacts. Methodologically this matters: a sweep stopping anywhere in 114…146 would have reported a clean cofinite-looking spectrum and been wrong. It is a direct caution against reading CRTI(n) = RT(n) off a finite run, and n = 4 — where RT(4) = 7/5 is Pansiot's exception to the general n/(n−1) formula — is exactly the alphabet size one should expect to misbehave.

All of this section is evidence for, not a proof of, CRTI(n) = RT(n): a finite unbroken run says nothing about what happens later, and C(n) could have infinitely many gaps even while CRTI(n) = RT(n) still holds. The honest statement is the certified one — the listed lengths are realizable, the listed exceptions are not, on the computed range only.

Section 3A pumping lemma for circular threshold words

The obstruction to turning section 2 into a theorem is that circular α⁺-freeness is not a bounded-window condition: a circular word of length m constrains periods up to roughly m(n−1)/n. The following lemma removes that obstruction by passing to the S₂ refinement, which is exactly wide enough to close an induction along a morphism.

Lemma A — circular pumping (PROVED)

Let α > 1, q ≥ 2, and let h be a q-uniform morphism on Σn that maps α⁺-free words to α⁺-free words. If w ∈ S₂(n,m) then h(w) ∈ S₂(n,qm).

Proof. |h(w)| = qm, and h(w)ω = h(wω) because h is a morphism. Let u be a factor of h(wω) of length ℓ ≤ qm+2, occurring at position s. Then u is a factor of h(v), where v is the factor of wω at letter positions ⌊s/q⌋…⌈(s+ℓ)/q⌉−1, so |v| ≤ ⌈ℓ/q⌉+1 ≤ ⌈(qm+2)/q⌉+1 = m+2 for q ≥ 2. Every factor of v is a factor of wω of length ≤ m+2, hence has exponent ≤ α because w ∈ S₂(n,m); that is, v is α⁺-free. Therefore h(v) is α⁺-free, so u — a factor of h(v) — has exponent ≤ α. ∎

Corollary B

If some q-uniform h maps α⁺-free words to α⁺-free words and S₂(n,m₀) ≠ ∅, then qjm₀ ∈ Cα(n) for every j ≥ 0; in particular Cα(n) is infinite. Taking α = RT(n) and invoking Observation 3, CRTW(n) = RT(n).

The "+2" is the whole content: S₁ does not close the induction (a factor of h(w)ω of length qm+1 can need m+2 letters of wω to cover it), and S₃ and beyond are unnecessary. Lemma A reduces an infinitary statement to two finite searches — a morphism, and one seed — and both are decidable (sections 4–5).

Section 4A decidable criterion for morphisms

What remains is a finite, checkable condition guaranteeing that a candidate morphism actually preserves α⁺-freeness.

Theorem M (PROVED; expected to be a rediscovery)

Let α ∈ (1,2], q ≥ 2, and h : Σn → Σnq be q-uniform with (H1) a ↦ h(a)[0] injective; (H2) a ↦ h(a)[q−1] injective; (H3′) for every α⁺-free ab, every c, and every 0 < i < q: h(c) ≠ (h(a)h(b))[i‥i+q−1]; (H4) h(v) is α⁺-free for every α⁺-free v with |v| ≤ N₀, where P₀ = ⌈2q/(α−1)⌉ and N₀ = ⌈(αP₀+1)/q⌉+1. Then h(v) is α⁺-free for every α⁺-free v.

Proof sketch. Suppose some α⁺-free v has h(v) not α⁺-free, via a factor u of period p with |u| = ⌊αp⌋+1 > αp. If p < P₀ then |u| ≤ αP₀+1 and u lies inside h(v′) for a factor v′ of v with |v′| ≤ N₀, contradicting (H4). If p ≥ P₀ then |u| ≥ p+2q, and period p forces an occurrence of some h(c) straddling a block boundary; (H3′) forces that occurrence to be aligned, so q | p. Writing p = qp′, alignment plus (H1) and (H2) (which rule out boundary loss) transfers the period down: v itself acquires a factor of period p′ and exponent > α, contradicting minimality. The point of (H1)+(H2) is precisely that they make the boundary loss vanish — without them the argument only yields exp ≥ α − 2q/p, the usual obstruction in lemmas of this kind. ∎ (Full proof: NOTE.md §4.)

Finite tests for power-free morphisms are a well-developed subject (Bean–Ehrenfeucht–McNulty; Crochemore; Richomme–Wlazinski; Ochem) (secondary — none of these was read). Theorem M should be assumed a rediscovery, in a convenient normalisation, until checked. It is proved here only so sections 5–6 stand on their own.

Theorem M′ — the fixed-point form (PROVED)

Let h be q-uniform satisfying (H1), (H2), (H3′), with h(0) beginning with 0, and let u = hω(0). Let P₀ = ⌈2q/(α−1)⌉ and B = ⌊αP₀⌋+1. If every factor of u of length ≤ B is α⁺-free, then u is α⁺-free.

Proof. Otherwise u has a factor of exponent > α; take one, z, whose least period p is minimal among all such factors. If p < P₀ the minimal such factor has length ⌊αp⌋+1 ≤ B, contradicting the hypothesis. If p ≥ P₀, since u = h(u), the argument of Theorem M's second case applies with v = u and produces a factor of u with period p/q < p and exponent > α, contradicting minimality of p. ∎ Every factor of u of length ≤ B lies in h(v) for a bounded factor v of u, so the hypothesis is a finite check, computable as the fixed point of S ↦ S ∪ Fac≤t(h(S)) from S = {0}.

Theorem C — the resulting decidable criterion (PROVED)

Let n ≥ 3 and α = RT(n). Suppose there exist a q-uniform morphism h on Σn satisfying (H1), (H2), (H3′), (H4), and a word w₀ ∈ S₂(n,m₀) for some m₀. Then circular n-ary threshold words exist at every length qjm₀, j ≥ 0; hence C(n) is infinite and CRTW(n) = RT(n). Both hypotheses are decidable by finite search.

Proof. Theorem M makes h α⁺-free-preserving; Corollary B applies. ∎ (The same conclusion follows from the relativised form of Lemma A with Theorem M′ in place of Theorem M, given the stronger seed requirement Fac≤m₀+2(w₀ω) ⊆ Fac(u).)

Section 5A worked instance: n = 3

Search over shift-equivariant q-uniform morphisms — h(a)[i] = (h₀[i]+a) mod n, the natural ansatz — found morphisms satisfying (H1)–(H4) at n = 3 from q = 25 on. Take q = 28 with

h₀ = 0120212010201210120102120210,  h(a) = h₀ + a (mod 3)

Conditions (H1), (H2), (H3′), (H4) — with P₀ = 75, N₀ = 8 — were verified twice, by two implementations sharing no code (C and Python), with identical verdicts. A seed was found by SAT: m₀ = 20, w₀ = 01202101210212012102 ∈ S₂(3,20).

Iterations of h
The morphism iterated live, in your browser, on the actual seed and substitution rule above. Bar height encodes the letter (0, 1 or 2); the faint vertical lines mark 28-letter blocks — each one the image of a single letter of the previous level. What is computed live here is the growth; the freeness of the result at each level was verified separately, offline, against the definition (table below) — this figure does not re-run that check.
jlengthcircular 7/4⁺-freein S₂
020yesyes
1560yesyes
215,680yesyes
3439,040yesyes
Result P1 (PROVED, and a rediscovery)

C(3) ⊇ {20·28j : j ≥ 0}, hence CRTW(3) = RT(3) = 7/4.

CRTW(3) = RT(3) is already known (secondary). Result P1 is reported only as a positive control: it demonstrates that the sections 3–4 machinery produces a correct theorem end to end, on the one small alphabet where the answer can be checked against the literature. No priority is claimed.

Section 6Why the same route did not settle an open case

Applying Theorem C to n ∈ [4,44] needs a morphism. The natural ansatz — shift-equivariant, the one that works at n = 3, and small enough to search exhaustively — fails, and fails for a reason.

Proposition N (PROVED)

Let n ≥ 4, α = RT(n), q ≥ n−1, and h(a) = h₀+a be shift-equivariant q-uniform over Zn with h₀ α⁺-free. Let D = {h₀[i+1]−h₀[i] mod n : 0 ≤ i ≤ q−2}. If u = hω(0) is α⁺-free and every letter of Zn occurs in u, then |D| ≤ 2.

Proof. For each d ∈ D, pick i with h₀[i+1]−h₀[i] = d; the letter a = −h₀[i] occurs in u, so h₀·(h₀+d) is a factor of u and must be α⁺-free. By section 1, positions at distance ≤ n−2 in an α⁺-free word carry different letters; applying this across the block boundary shows d ≠ h₀[q−1]−h₀[k] for every 0 ≤ k ≤ n−3, and those n−2 differences are pairwise distinct (since h₀[0],…,h₀[n−3] are, lying in a window of length n−2). So D omits at least n−2 of the n possible differences: |D| ≤ 2. ∎ Proposition N is vacuous at n = 3 (it gives |D| ≤ 2, and D ⊆ {1,2} always there) — exactly why n = 3 admits such morphisms and n ≥ 4 is squeezed.

A word whose consecutive differences take only two values turns out to be very short if it is to be α⁺-free at all:

Result L (CERTIFIED, exhaustive)

The longest RT(n)⁺-free word over Σn whose consecutive differences lie in a set of size ≤ 2 has the lengths shown below (exhaustive depth-first search over all ≤ C(n−1,2)+(n−1) difference sets and all step sequences).

L(n): how short a "boring" (two-difference) threshold word must be. Exact values from exhaustive search, not a fit — hover any bar. The values are small and not obviously monotone in n, which is itself an open pattern (section 7). Combined with Proposition N, any shift-equivariant morphism of length q > L(n) is automatically excluded, which is what makes the exhaustive search below tractable.
Theorem N′ — computer-assisted proof

Let 4 ≤ n ≤ 9 and α = RT(n). No shift-equivariant q-uniform morphism h(a) = h₀+a over Zn has an α⁺-free fixed point, for any q ≥ 2.

Proof. (i) The letters occurring in u = hω(0) form the subgroup H = ⟨h₀[0],…,h₀[q−1]⟩ ≤ Zn; α⁺-freeness forces any n−1 consecutive letters distinct, so |H| ≥ n−1, and since a proper subgroup has order ≤ n/2 < n−1 for n ≥ 3, H = Zn — every letter occurs, so Proposition N's hypothesis is automatic. (ii)–(iii) Proposition N then gives |D| ≤ 2 for q ≥ n−2, and Result L bounds q ≤ L(n). (iv) For every q from 2 up to and beyond L(n) — q ≤ 60, 12, 45, 38, 20, 14 for n = 4,…,9 — an exhaustive search over all α⁺-free h₀ ∈ Σnq with h₀[0] = 0 finds no candidate satisfying (ii): zero survivors, at every q. No q survives (i)–(iv). ∎

Two steps of this proof are exhaustive searches over explicitly finite sets, carried out in exact integer arithmetic and reproducible from the shipped scripts; the reduction (i)–(iii) is unconditional. This is the session's clearest negative result, and it is a real one: the normal form that works at n = 3 is provably unavailable at every alphabet size from 4 to 9, for every morphism length — not merely that the search ran out of time.

Corroborating this from the other side, the earlier brute-force sweep under the stronger criterion of Theorem M — a different condition, so an independent check of the same phenomenon rather than a restatement — also came back empty:

Result N1 (CERTIFIED)

Exhaustive search over all α⁺-free h₀ ∈ Σnq with h₀[0] = 0 finds no shift-equivariant q-uniform morphism satisfying (H1)–(H4) at α = RT(n), for n = 4 (q ≤ 60), n = 5 (q ≤ 68), n = 6 (q ≤ 51), n = 7 (q ≤ 44), n = 8 (q ≤ 37). The same search at n = 3 returns morphisms from q = 25 on, so the pipeline is not failing vacuously.

Interpretation. By Theorem N′ the shift-equivariant ansatz is not merely unlucky at n ≥ 4; it is impossible. The literature's own remark that for small alphabets "techniques different from those presented will likely be needed" (secondary) is consistent with this. The natural next ansatz is Pansiot's encoding of n-ary threshold words by words over a smaller alphabet (secondary), where the relevant morphisms need not be shift-equivariant over Zn — Theorem C applies verbatim to any morphism found there. Sections 7–9, from a second session three days later, take exactly this route.

Section 7 · Part IIThe Pansiot code: states, monodromy, and an exact transfer lemma

Fix n ≥ 3, α = RT(n). Section 1 already established that any two positions at distance ≤ n−2 in an α⁺-free word carry different letters, so every window of n−1 consecutive letters of a threshold word is rainbow — all distinct — and after such a window (a₁,…,an−1) (oldest first) with missing letter b, exactly two letters can follow: a₁ (encoded bit 0) or b (bit 1). This is Pansiot's encoding (secondary — not read this session either): it turns an n-ary threshold word into a binary word, the code.

Formally, a state σ = (a₁,…,an−1 | b) is an element of Sym(n) via σ(i) = ai, σ(n−1) = b. Bits act on the right: σ·τ₀ = (a₂,…,an−1,a₁ | b) and σ·τ₁ = (a₂,…,an−1,b | a₁), where τ₀ is the (n−1)-cycle on window slots and τ₁ the n-cycle. For a bit word w, g(w) = τw₁∘⋯∘τwL is its monodromy. dec(V), the decode of a bit word V, is the canonical initial window (0,1,…,n−2) followed by the |V| letters emitted as above; V is code-free if dec(V) is α⁺-free. Every emitted letter automatically avoids the previous n−2 letters, so any over-repeat in a decode has period p ≥ n−1 — the short-range constraint is free, and only the long-range one needs work.

Lemma S — the slot lemma (PROVED)

The equal-letter pattern of dec(V) — hence its code-freeness — depends only on V, not on the choice of initial state.

Proof. The letter emitted at step t is σt−1(π(Vt)) where π(0)=0, π(1)=n−1 and σt−1 = σ₀∘Ht−1 (H = prefix monodromy). Since σ₀ is a bijection, two emitted letters agree iff a fixed relation between the intermediate bits holds, independent of σ₀. ∎

Lemma T — the exact transfer lemma (PROVED)

Let V be a bit word, p ≥ n−1. Call bit position a a gid position if Ha = Ha+p (the monodromy from a to a+p is trivial), and e(a) the length of the run of bits agreeing p apart, starting right after a. (i) A gid position a gives dec(V) a period-p stretch of length (n−1)+p+e(a). (ii) Conversely, a period-p stretch of dec(V) of length ℓ ≥ p+n−1 starting at letter i forces i to be a gid position with e(i) ≥ ℓ−(n−1)−p.

Proof. (i) Ha=Ha+p gives σa=σa+p, so the two states' windows already agree letter-for-letter — n−1 equalities at distance p for free — and each further bit of agreement propagates one more equal letter, since equal states plus equal bits give equal next states and equal emissions. (ii) runs the argument in reverse: enough consecutive equal letters at distance p force the windows, hence the states, to coincide, and equal states with two possible distinct continuations pin down the bit. ∎ (Full derivation and a 2,238-test cross-validation against an independent H-prefix checker: NOTE.md §12.)

Lemma T says exactly this: a long-range repeat in the decoded word is the same thing as a stretch of trivial monodromy in the code. That equivalence is what makes a finite, checkable criterion possible in section 8.

Section 8 · Part IIA finite criterion, and pumping, in the code

Theorem MC (PROVED)

Let φ: {0,1} → {0,1}k be uniform with blocks φ₀, φ₁. If (Ha) φ₀ and φ₁ start with different bits, (Hb) they end with different bits, (Hc) no φc occurs misaligned inside any code-free φaφb, (C2) g(φ₀), g(φ₁) are conjugate to τ₀, τ₁ by a common π ∈ Sym(n), and (Hd) φ(V) is code-free for every code-free V up to an explicit finite length N₀ — then φ(V) is code-free for every code-free V, of any length.

Proof sketch. Suppose φ(V) has an offending period-p stretch. Short p (below a threshold P₀ fixed by k, n, α) pulls back, via Lemma S, to a bounded factor of V, caught directly by (Hd). Long p forces (by Lemma T(ii)) a gid position with a long run of agreement; (Hc) then forces block-alignment, (Ha)/(Hb) rule out boundary loss exactly as in Theorem M, and the period transfers down to a shorter offense in V itself, contradicting V's code-freeness by induction. Full proof, with the exact constants P₀, N₀: NOTE.md §13.

Lemma PC — pumping in the code (PROVED)

Let φ satisfy (C2) and map code-free words to code-free words. Let c₀ be a cyclic bit word of length m₀ with (M) g(c₀) = id and (S2) every factor of c₀ω of length ≤ m₀+2 code-free. Then for every j ≥ 0, φj(c₀) also satisfies (M) and (S2) at length kjm₀, and its cyclic decode is a circular α⁺-free word of that length. Hence C(n) is infinite and CRTW(n) = RT(n).

Proof. (M) is immediate: g(φ(w)) = π−1g(w)π since g and φ are both monoid morphisms, so g(c₀)=id propagates. (S2) propagates by the same bounded-pullback argument as Theorem MC. (M) makes the cyclic decode close up consistently around the loop; (S2) at the right width is exactly what Lemma S needs to certify every short factor of the decode. ∎

Theorem C-code (PROVED)

If some uniform binary φ satisfies (Ha)(Hb)(Hc)(C2)(Hd) and some cyclic bit word satisfies (M)(S2), then CRTW(n) = RT(n). All five hypotheses are finite, checkable computations.

Section 9 · Part IIInstances: n = 6 and n = 5 settled, n = 4 certified empty

All hypotheses below were checked in exact integer arithmetic by pansiot_certify.py and pansiot_seed.py, and re-checked independently for this page directly against the committed data (not merely re-read from the session's log).

Theorem P6 (PROVED) — CRTW(6) = RT(6) = 6/5, an open case of the conjecture

φ₀ = 010101101101011010110, φ₁ = 101011010110110101101 (k = 21), with g(φb) = π−1τbπ for π = (1,2,3,0,4,5); (Ha)(Hb)(Hc)(Hd) verified with P₀ = 230, N₀ = 15 (all 338 code-free words of that length checked). Seed: the certified spectrum witness of length 39 from section 2, re-encoded. Hence C(6) ⊇ {39·21j : j ≥ 0}.

jlengthmonodromy trivialcircular 6/5⁺-free
039yesyes
1819yesyes
217,199yesyes

Reproduced independently for this page: pansiot_certify.py confirms Ha/Hb/Hc/C2/Hd; pansiot_seed.py check confirms (M)/(S2) at m₀=39; pansiot_seed.py pump re-derives circular freeness at all three lengths above directly from the definition, matching the session's own log exactly.

Theorem P5 (PROVED, flagged) — CRTW(5) = RT(5) = 5/4

φ₀ = 010101101101010110110, φ₁ = 101010101101101101101 (k = 21), π = (0,1,2,4,3); P₀ = 180, N₀ = 12 (144 words checked). Seed of length 28. C(5) ⊇ {28·21j}, verified directly at lengths 28, 588, 12,348. 5 is odd; if Tunev's construction (section 10) covers n = 5, this theorem is a rediscovery with an independent proof.

Result P3′ (PROVED; independent control) — CRTW(3) = 7/4, re-derived through the code

φ₀ = 0101101011101110110, φ₁ = 1011010111011101011 (k = 19), π = (2,0,1); seed length 20, the Part I seed re-encoded. C(3) ⊇ {20·19j}, verified to length 137,180. A known result, reported only because the code machinery reproduces it correctly and independently of the section 3–5 route.

Result N2 (CERTIFIED, stated ranges) — the n = 4 ansatz is empty

No pair of valid k-blocks, k ≤ 46, in any pooled monodromy class (C2, sign, collapse-1, collapse-2) has a fixed point decoding 7/5⁺-free to depth 4000 — every pooled pair (up to 74,183 per k at the top of the range) is refuted by an explicit offender. A two-level substitution engine finds nothing over r ≤ 10, s ≤ 5 and r ≤ 7, s ≤ 6 (one cell partially capped, disclosed as such). Not an impossibility proof — no analogue of Theorem N′ is claimed here — but a second, independent search coming back empty at n = 4, in a completely different encoding from section 6's.

Spot-checked independently for this page by re-running the search at k = 44, 45, 46 (pool sizes 58,424 / 32,189 / 74,183): zero candidates at each, matching the session's log exactly.

Where the code ansatz finds anything at all, across alphabet sizes. Each strip is one alphabet size n; a mark at position k means the sweep found at least one candidate pair of k-bit blocks that decodes code-free to the depth searched (hover for the exact count); a bare strip means zero, at every k searched. Strips end where each sweep stopped (24, 46, 40, 40 and 30 respectively) — n = 4's silence runs to the far right of its own strip, not to the page edge. n = 8 (44 candidates at k = 28) is the one case found this session with no certificate passing (Ha)/(Hb) yet — an open thread, not a theorem. n = 3's counts are from an earlier pass of the search script (fixed-point survival to depth 4000, not the pooled-candidate metric used for n ≥ 4), so its numbers are not directly comparable to the other rows; only the hit/no-hit pattern is.

Section 10 · Part IIA discovery made mid-session: Tunev, 2025

A December 2025 preprint — Igor N. Tunev, arXiv:2512.24581, "On Circular Threshold Words and Other Stronger Versions of Dejean's Conjecture", in Russian, an edited write-up of the author's 2011/2013 theses — surfaced during this session's literature survey. Search snippets state that a 2011 method proves Dejean's conjecture "for some odd cases n ≥ 5" using constructions where "any cyclic shift is a threshold word" (secondary — the paper itself was not, and could not be, opened; egress was blocked as in Part I). A word all of whose cyclic shifts are threshold words is precisely a circular threshold word, so if the snippet is accurate, Tunev's construction settles CRTW(n) = RT(n) for some odd n ≥ 5, predating this session, for n = 5 specifically.

It cannot predate n = 6 under any reading: 6 is even, and every snippet found describes odd cases only. Theorem P6 therefore stands as reported. Theorem P5 is downgraded in priority, not in truth: it is reported as PROVED here — the proof is independent of Tunev's route, uses none of his machinery, and was checked against the definition directly — but flagged as a probable rediscovery until arXiv:2512.24581 can actually be read. A companion peer-reviewed paper by the same author (Tunev–Shur, MFCS 2012) is known to cover two different strengthenings of Dejean's conjecture (growth rate; finitely many distinct repetitions), which is consistent with Mol–Rampersad (2020) still listing all of 4 ≤ n ≤ 44 as open — the 2025 preprint may simply not have been indexed against that specific open-problem list yet.

One more caveat in the same direction: Moulin Ollagnier's proof of Dejean's conjecture for 5 ≤ n ≤ 11 is reported to work in the Pansiot encoding and to relate repetitions to the identity in the symmetric group (secondary) — the same mechanism Lemma T formalises here. Lemma T should be presumed to overlap his machinery in content, in some normalisation, until his paper is read. Action required before any submission of this work: read both papers and revise the priority claims above accordingly.

Section 11What is, and is not, settled

Settled, Part I (2026-08-03). Lemma A and Corollary B (PROVED, believed new); Theorems M and M′ (PROVED, expected rediscoveries); Theorem C, a decidable sufficient criterion for CRTW(n) = RT(n) (PROVED); Proposition N (PROVED); the spectrum data of section 2 and the negative search of section 6 (CERTIFIED, on the stated ranges only). No open case of the conjecture settled.

Settled, Part II (2026-08-06). Lemmas S, F, T; Theorem MC; Lemma PC; Theorem C-code (all PROVED, hypotheses machine-checked). Theorem P6: CRTW(6) = RT(6) = 6/5 — an open case of the Currie–Mol–Rampersad conjecture, settled (openness secondary-sourced). Theorem P5: CRTW(5) = RT(5) = 5/4, settled by an independent proof, flagged as a probable rediscovery of Tunev (2025) pending a primary-source read (section 10). Result P3′ (independent control, known result). Result N2 (CERTIFIED negative for n = 4, stated ranges).

Not settled. CRTW(n) = RT(n) for n ∈ {4} ∪ [7,44] remains open. CRTI(n) = RT(n) remains open for every n ≥ 4; the spectrum data is evidence, not proof, of a kind no finite computation can ever upgrade. Whether Tunev's construction actually reaches n = 5, and which other odd n it covers, is unknown until the paper is read.

Sharpest open threads, in order:

  1. Read Tunev (arXiv:2512.24581) and Moulin Ollagnier from an unblocked network, and revise every priority claim above accordingly. This is the thread that gates everything else being submitted anywhere.
  2. Run the section 8 machinery at n = 7, 9, 10, …, 12. n = 8 already has 44 viable block pairs at k = 28 that fail only the (Ha)/(Hb) injectivity hypotheses — a synchronization-based relaxation of Theorem MC without those two hypotheses would plausibly unlock n = 8, 10, 12, … immediately; each even n is a new theorem, each odd n a Tunev cross-check.
  3. n = 4: either extend Result N2 into a genuine impossibility proof for the code ansatz (a Theorem-N′-style argument, not yet found), or push the search past k = 46.
  4. Determine whether the exceptional sets of section 2 are final — whether the unbroken runs continue. A single late gap at n = 4 or n = 5 beyond what is shown would bear directly on CRTI.
  5. Decide whether S₂(n,m) ≠ ∅ and S₀(n,m) ≠ ∅ differ for any m; Lemma A only moves S₂, so the gap between the two spectra is the real cost of pumping.