Beyond the two coefficients

In May I recorded a clean obstruction. For every fixed look-ahead depth k at least 2, and every choice of coefficients alpha and beta in a linear energy, there exist infinitely many odd starting points where the score does not fall. Mersenne odds give explicit witnesses. The original plan, find alpha, beta, and a magic k that makes energy descend for all n, dies as a plan.

That was already a real result. Small. Negative. Worth writing carefully. It is still not a proof of Collatz, and it never claimed to be. If you have not read the May entry, read it as the emotional and practical death of the coefficient search. This July entry is the mathematical widening: the problem was never mainly that I had only two knobs.

I kept a linear ansatz because it matched the slogan I cared about. Reward the present collapse. Reward the near future. Search. The slogan named a direction. The ansatz froze the direction into a uniform gadget. Uniform gadgets are vulnerable to streaks longer than their memory. May showed that vulnerability for linear memory. July shows it for arbitrary memory of fixed length.

There is a reason to keep both stamps public. Chronology is part of honesty. Strengthening a theorem should not erase the season when the first form died. Warehouse behavior rewrites the past into inevitability. Notebook behavior keeps the steps.

Setup in plain language. Work with odd positive integers and the Syracuse map: multiply by three, add one, divide out all factors of two. Write nu_i for the number of twos peeled at the i-th upcoming Syracuse step. Fix an integer k at least 1. Let F be any real-valued function of k real inputs. Define energy as log n minus F of the window (nu_0, ..., nu_{k-1}).

Call an odd n an L-bad integer when the first L upcoming valuations are all one. Ugly window. Expanding streak. Mersenne-type odds give infinite families of such n. Broader arithmetic progressions in the two-adic neighborhood of minus one give more.

Theorem, strengthened form: fix k and fix any F of that window. Let n be any (k+1)-bad odd integer. Then the window at n and the window at C(n) are identical strings of ones. The F terms cancel in the energy difference. What remains is log of C(n) over n, which equals log of (3n+1) over (2n), which is strictly larger than log of three over two. Energy rises. Therefore no choice of fixed k and F makes this energy a strict Lyapunov function on all odd positives.

That is the whole kill. Not a vibes argument. Not a statement that Collatz is false. A statement that one proof strategy is closed: pick a fixed look-ahead depth and any reward of that window, then hope energy always falls.

Corollaries follow immediately. The linear May ansatz is a special F. Coefficient search cannot succeed inside that special case. Buying a longer fixed flashlight only raises the streak length needed to break you. The adversary with access to longer L-bad families always has more streak than you have window.

I care about the uniform margin. The rise is not an epsilon that clever F might shrink. Once cancellation happens, the rise is the pure log growth of an expanding step. Cleverness has nowhere to sit.

In 2026 there is also residue-only Lyapunov obstruction work that dies on integers congruent to minus one modulo high powers of two. Same geometric neighborhood. Different ansatz. Cite that lane. One concrete pointer in the public conversation is Hsieh 2026 on Zenodo: residue corrections of the form log n plus a function of n modulo a power of two, obstructed on the hard congruence classes near minus one.

How this note differs: we obstruct look-ahead window rewards F of upcoming valuations, not residue corrections c of n mod 2^m. A publishable contribution here is the clean closure of the look-ahead-window class, with explicit families and reproducible code. It is not a claim of priority over all Lyapunov obstruction ideas. It is not a claim that Omega is a new named object beyond Terras-type exponent data.

I am repeating the citation because ego is sticky. Negative-result culture can become a race to sound like the owner of impossibility. The adult move is narrower. Name the class you closed. Name the parallel class others are closing. Let the shared geography teach you where the hard integers live without demanding a crown.

Terras, Lagarias, Tao, Simons-de Weger, Krasikov-Lagarias: these are not decorative names. They are the prior-art spine any short note must face before arXiv. Phase 0 literature notes are still a gate. This notebook entry is a field stamp, not a submission. If Phase 0 finds the exact window theorem already in print, rewrite as exposition and keep the adaptive and book tracks. Being scooped on a negative result is not a tragedy. Publishing a false priority story is.

Omega inflation dies here again on purpose. Summing valuations over a window is natural. Natural is not novel. The novelty question that matters is whether the obstruction theorem for arbitrary fixed-window F is cleanly written and correctly positioned. That is the bar.

After a fixed-window death, the slogan that arrives first is: let k depend on n. That slogan is the right direction and a common place to reintroduce sloppiness.

Adaptive depth is not automatically a theorem. Equivalence claims need a correct descent threshold. One-step Syracuse descent for large n requires enough valuation to beat the factor three. Roughly, valuation at least 2 is the beginning of one-step contraction in log scale, with lower-order terms. The naive surplus nu minus log base two of three over two is positive even for nu equals 1. So it does not encode descent. It encodes a feel-good score that celebrates expanding steps.

If you define a window surplus with that naive threshold, you can convince yourself that valuation-one steps are already "progress" in a bookkeeping sense while the odd part is growing. That is the surplus trap. The flashlight has to earn its length against a real contraction criterion: a product of terms like 3 over 2^{nu_i} less than 1 over the window, or equivalently a sum of valuations large enough compared to k log base two of 3, up to lower-order corrections you actually control.

Open question that still interests me: for which explicit families can one bound the minimal window length that yields true contraction? Mersenne and L-bad families are the stress tests. If your adaptive story cannot speak to them, it is not ready.

Another open caution: proving that k(n) is finite for every n is, in careful formulations, equivalent to Collatz or nearly so. That means the adaptive reformulation is a lens, not a shortcut. Lenses are valuable. Shortcuts that are secret restatements are how people fool themselves after a negative theorem.

I will not ship an equivalence claim in this entry. I will ship the warning. Adaptive plus real contraction is the research door. Adaptive plus naive surplus is a newly painted warehouse room.

Formal drafts belong in the project proof files. Human surfaces belong where a non-specialist can feel the shape without being lied to. Ugly Windows is that surface: fixed look-ahead, Mersenne streaks, seasons that refuse to let a simple score of progress keep falling.

The essay is not a substitute for the theorem, and the theorem is not a substitute for the essay. Dual surfaces prevent two failure modes of communication. Pure formal notes can be correct and unread by the people who need the ethical lesson about dead plans. Pure essays can become myth if they float free of witnesses. Keep both. Date both. Let them point at each other.

This notebook entry sits between them. More technical than the essay. More lived than the arXiv draft. Status: small-result. Claim: fixed-window look-ahead rewards are obstructed. Non-claim: Collatz.

If you only remember one image, remember a flashlight with a fixed beam length aimed down a hallway of doors that are all expanding steps. While the hallway is longer than the beam, every clever painting on the beam looks the same after you take one step forward. The paint cancels. The hallway still rose. That is the obstruction.

Energies that use adaptive window length k(n), with a contraction test that is actually a contraction test.

Energies that use information outside any fixed finite window of upcoming valuations: global orbit data, total stopping-time structure, hybrid certificates that mix look-ahead with cycle bounds or density machinery.

Statistical questions about minimal contracting windows on large ranges, clearly labeled as computation and heuristics unless they become theorems.

Cycle-exclusion lanes that do not require a uniform Lyapunov function on all odd n.

Literature completion: Phase 0 notes, accurate Hsieh citation details, Terras relation stated without novelty inflation, cold ask to a specialist whether the window theorem is already standard folklore in some survey corner I have not seen.

None of these open items is a promise. Open means open. The May ledger listed remainders. This July stamp updates the mathematical ceiling of the closed part: not only linear F, any F of fixed window.

There is a draft note in the Going Home proof files aimed at math.NT after a literature pass. Tone target: small, precise, humble. No metaphysics in the PDF. No Collatz proof claim. Companion code for reproduction.

Submission checklist, said out loud so I cannot pretend it is optional: Phase 0 lit notes filed. Theorem proof typed without handwaves. Mersenne and abundance lemmas fully proved, not sketched. Parallel residue-obstruction work cited accurately. Terras relation of Omega stated without novelty inflation. Code link durable. Zero spirituality in the abstract. A cold email or equivalent reality check from someone who lives in this literature.

Why wait? Because negative results still deserve prior-art hygiene. Because parallel 2026 work makes haste look like territoriality. Because the book and essay tracks can carry meaning while the math track stays austere. Because warehouse days taught me what happens when every genre crowds one scroll.

If I submit before the lit pass, I am gambling reputation for calendar speed. Calendar speed is not a scientific value. The open notebook already timestamps the strengthening. Timestamp is not the same as arXiv. Timestamp is enough for the climb narrative while the note matures.

Stop searching for alpha and beta. Stop searching for a larger fixed k as if it were a ladder out of the obstruction. The ladder needs more streak, and more streak exists.

Keep the lab scripts. Extend them when the adaptive contraction criteria are sharp enough to test. Do not extend them as a way to postpone accepting Theorem 1.

Keep philosophy in philosophy. The Santana Principle as awareness metaphor can remain a book engine. It cannot be smuggled into the abstract of the obstruction note.

Keep public language disciplined. On the site, small-result means small-result. Hooks must refuse solved-language. If a teaser sounds like arrival at one, rewrite the teaser.

Keep citation of Hsieh-style residue work adjacent whenever we discuss the minus-one neighborhood. Adjacent citation is how we stay in a research community instead of a private myth.

Keep publishing ugly middles. May death. July widening. November warehouse memory. Pressure-testing. The climb is the product until a true theorem about Collatz exists, which this is not.

Beyond the two coefficients means beyond the fantasy that the binary search was the boss fight. The boss fight was fixed finite look-ahead itself. Any scorecard locked to a window chosen in advance can be broken by a streak longer than that window.

Beyond also means beyond novelty theater around Omega. Beyond means into adaptive contraction with adults in the room. Beyond means into literature posture that can survive a specialist's first ten minutes. Beyond means refusing to inflate a closed strategy into a solved conjecture.

I want the last paragraph to be usable as a quote without becoming a lie. Here it is. We closed the fixed look-ahead window class of Lyapunov rewards for the Syracuse map: every fixed k and every F of that window fails on infinitely many odd n, with explicit L-bad witnesses near minus one in the two-adics; parallel residue-only obstruction work finds hard integers in the same neighborhood by a different ansatz; Collatz remains open; the research door we respect is adaptive depth with a real contraction test, not a naive surplus and not a coefficient grid.

That is the July stamp. Checkable. Narrow. Dated. Not home as arrival. Home as return to a true description of what the flashlight can and cannot do.

If you need a still shorter version for a stranger: fixed look-ahead scores die on ugly windows; we proved that for any score of a fixed window; others are killing residue scores nearby; nobody here solved Collatz.

If you need a version for me on a tired night: stop retuning the two knobs. The knobs were never the door. The fixed window was. Walk toward adaptive contraction or walk toward another problem. Do not build another warehouse in the hallway you already measured.

Take a window length k equals 2 for concreteness, though the argument does not need smallness. Let F be any function of two inputs. Let n be 3-bad, so the first three valuations are one. Then the window at n is (1, 1). After one Syracuse step, the window at C(n) is (nu_1, nu_2) which is again (1, 1). F sees the same pair. Subtracting F cannot create a decrease. The log term increases by more than log of three over two because an expanding step grew the odd part.

Now lengthen the flashlight to k equals 8 because it feels safer. The adversary chooses a 9-bad n. The window of eight ones slides by one step and remains eight ones. Same cancellation. The feeling of safety was a feeling about your beam, not a fact about the hallway.

Replace linear F with a neural net over the window if you want to be theatrical. On the all-ones window the net outputs some number c. After the step it outputs the same c. Theater cancels. This is why the strengthening matters psychologically. It blocks the escape hatch that says my two coefficients were too primitive.

Replace F with a function that tries to special-case the all-ones tuple. It still outputs a single number for that tuple. Before and after, same tuple, same number. Special casing the villain signature does not help when the signature is stable along the step that breaks you.

That stability is the geometric content. L-bad streaks are not only bad because they expand. They are bad for look-ahead certificates because they look like themselves under a shift until the streak ends. Fixed windows are shift-local. Shift-local scores are blind on long constant streaks.

Residue certificates fail for a related geographic reason with a different mechanism: congruence information near minus one cannot manufacture enough descent reward when the dynamics force expansion. I am not restating Hsieh's proofs here. I am marking the kinship and the difference. Kinship: minus-one neighborhood. Difference: residue ansatz versus look-ahead-window ansatz.

If a future ansatz wants to survive, it must either change window length with n in a controlled way, or reach outside pure finite look-ahead of nu's, or accept that it only certifies a subclass and say so. Subclass theorems can be honorable. Quietly claiming a subclass certificate as a universal Lyapunov function is how we got here.

End of the slow example. If it felt repetitive, good. Repetition is the pedagogic form of cancellation. The math is repeating ones. The mistake is believing a new costume changes the ones.

Before I touch another candidate energy in this lane, I want a preflight that does not depend on mood. First: is the reward a function of a fixed finite window of upcoming valuations? If yes, stop. Theorem 1 already killed that class. Second: if the window length depends on n, what is the contraction predicate in inequalities I can falsify on Mersenne odds? If I cannot state the predicate, I do not have a method. Third: does any surplus I write stay positive on valuation one by design? If yes, I am in the naive surplus trap until I prove that positivity means descent, which it does not.

Fourth: have I checked the minus-one neighborhood and L-bad families before celebrating a numerical win on random samples? Random samples can hide the villains that matter. Fifth: have I written the prior-art paragraph naming Terras-type look-ahead data and residue-obstruction parallels such as Hsieh 2026? If the paragraph feels embarrassing, good. Embarrassment is often the adult form of citation. Sixth: am I about to put awareness language into a lemma slot? Move it to the book track immediately.

Seventh: is the public sentence I want to publish shorter than the theorem and louder than the theorem? If louder, rewrite until the volume matches the math. Eighth: is there a reproducible script? No script, no stamp. Ninth: am I calling this a Collatz proof in any sidebar, talk, or late-night message? Then I am the unreliable narrator of my own project, and the notebook should correct me in public.

This checklist is not bureaucracy. It is anti-warehouse machinery. The November day taught me how rooms multiply when verdicts are expensive. The May day taught me that verdicts can be negative and still worth the cost. The July day taught me that widening a negative verdict is also worth the cost when it closes escape hatches. Escape hatches are how smart people stay wrong elegantly.

I will also keep a tenth item, smaller and more personal. After writing a strengthening, do not immediately invent a new brand name for the adaptive restatement. Naming is packaging. Packaging is how Omega got overclaimed. Let the adaptive question stay a question until a contraction theorem exists that does not assume Collatz in disguise.

If someone quotes only the flashlight metaphor and skips the non-claim, they will misunderstand on purpose or by haste. The non-claim is load-bearing. Fixed-window look-ahead Lyapunov rewards are obstructed. Collatz is open. Those two sentences must travel together in every serious summary of Going Home math as of 2026-07-31.

That is enough checklist. The work now is literature, LaTeX care, and honest adaptive experiments that can fail out loud. Failure out loud is the house style. Solved conjecture fanfare is not.