constant-recognition
Discipline for integer-relation searches (PSLQ, LLL) that turn high-precision digits into exact constants. Use whenever recognizing a numerical value as a closed form, fitting boundary constants, or reporting that no closed form exists.
BootLoops-ai · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add BootLoops-ai/skills --skill constant-recognitionAbout this skill
name: constant-recognition description: Discipline for integer-relation searches (PSLQ, LLL) that turn high-precision digits into exact constants. Use whenever recognizing a numerical value as a closed form, fitting boundary constants, or reporting that no closed form exists.
constant-recognition — naming numbers without fooling yourself
An integer-relation algorithm is a fitting machine. Hand PSLQ a vector of reals and enough coefficient freedom and it returns a relation, every time; that is what lattice reduction does. Whether the relation means anything is decided entirely by choices made before the search — which constants are allowed, how large the integers may be, how many digits pay for it all. The same choices made after seeing the digits are curve fitting with a number theorist's vocabulary. Everything in this skill exists to keep the decisions on the correct side of the search.
The asymmetry to internalize: a hit is cheap — noise plus freedom produces hits on demand — while a null is informative only against a declared ring. "PSLQ found nothing" means nothing by itself. "Not a Z-linear combination of {1, ζ(3), π² log 2, Li₃(1/2)} with coefficients below the declared height at the declared precision" is an exact, reusable, citable statement. The whole protocol is arranged so that both outcomes mean something.
The digit budget
Before anything else, the arithmetic that governs the whole exercise. A candidate relation among n constants with integer coefficients up to height H consumes roughly n·log₁₀H digits of precision just to be expressible, and a search run with only that much precision can always satisfy itself. The digits left over — working precision minus that cost — are the only evidence. If the surplus is thin, the search is a coin flip; if the ring and heights cost more digits than you have, the search is theater, and it will still return a relation. Budget first: choose the ring, choose the largest coefficients you would believe, compute the cost, and demand a surplus of many tens of digits before running at all. When the budget does not close, the correct moves are to compute more digits or to shrink the ring. Running anyway is the original sin from which every failure below descends.
Procedure
-
Declare the ring, in writing, before the first search. List the constants the answer is allowed to draw on and the reason each one is on the list: a weight grading, the arithmetic of the geometry the problem lives on, the closed forms of solved neighboring cases. The reason matters — "it shows up in related problems" admits everything eventually. Date the list and keep it with the run, so the record shows the declaration preceded the search.
-
Fix the height bound and pay for it. State the maximum coefficient size you would accept in a believable answer — informed by the heights that appear in solved relatives of the problem — then verify the digit budget above. Record both numbers with the declaration.
-
Plant the controls before the real value goes in. Two of them, both pushed through the identical code path — same script, same precision, same ring, same height bound:
- a positive control: a value whose closed form in the declared ring is independently known. The search must recover exactly that form.
- a negative control: a value constructed to lie outside the ring — the digits of a random real at the same precision serve. The search must return null.
A search whose controls have not run is uncalibrated, and any hit it produces is unreviewable.
-
Run, and record everything. Precision, ring, height bound, the exact input vector, the algorithm and its settings. A relation whose search parameters were not recorded cannot be distinguished, later, from one found by dredging.
-
Rerun at genuinely higher precision. Not a handful of extra digits — enough to move the noise floor. A true relation persists with the same integers; a spurious one dissolves or reshuffles its coefficients. Identical integers at two well-separated precisions is the minimum bar for taking a hit seriously. It is not the final bar.
-
Certify on digits the search never touched. The digits that suggested the relation may not also certify it — that is testing a fit on its training data. Evaluate the value and the proposed closed form by an independent route — a different method or a different implementation — at precision beyond everything the search consumed, and demand many digits of agreement past the fit. These held-out digits are the acceptance gate; nothing moves forward without them.
-
Report one of two things. Either the closed form with its full certification record (ring, heights, precisions, control outcomes, held-out agreement), or the null with its parameters plus the value itself: a convergent series or integral representation with an arbitrary-precision evaluator, so the number is usable without a name.
Failure modes
These are the classes the discipline exists to catch. Each one produces output that looks like success.
Ring enlargement after seeing the digits. The declared search fails; a plausible extra constant goes in "because it appears in related problems"; the enlarged search fails; another constant goes in; the third search closes, with large coefficients. The closure was manufactured by the enlargement process — each added constant is added freedom, and enough freedom always closes. An enlargement is a new experiment: it needs a justification that does not mention the failed search, and the digit budget must be paid again for the bigger ring. Repeated ring changes against the same digits are dredging, whatever the log calls them.
Height creep. The same failure on the other axis. The search at the declared bound returns nothing; the bound is raised "just to see"; a hit appears. Each raise
Read the complete source ↗