Skip to content

Blog: the golden ratio in this format is a scale factor, not information - #830

Merged
gHashTag merged 2 commits into
mainfrom
post/phi-is-a-scale
Aug 18, 2026
Merged

gHashTag merged 2 commits into
mainfrom
post/phi-is-a-scale

Conversation

@gHashTag

@gHashTag gHashTag commented Aug 18, 2026 •

Copy link
Copy Markdown
Owner

Bilingual post (EN + RU) on what was measured today, and the first post to carry a diagram — adds a figure block kind holding inline SVG plus its renderer case.

What it reports

where phi² = phi + 1 ?
the oracle, exact ring Z[phi] yes, symbolically — 14641/14641 sign decisions agree with 60-digit arithmetic
the closed 2-bit format not measurable — 1, 2, π, 0.001 and 10¹² all give identical tables
the RTL, fp32 no, off by 1 ulp

And the theorem: ⟨x,y⟩ = phi²·⟨x̃,ỹ⟩ exactly, so phi is a per-tensor scale. The fp32 multiplier the board runs for that product costs 1654 cells against 13 for the exact unit, 16/16 input pairs equivalent, on Yosys 0.65.

The diagram

A difference diagram, not an illustration: the same constant in two placements, and the one edge that exists on the left (phi factors out) and cannot exist on the right (Stakhov's positional weights). Inline SVG in currentColor, so it themes with the page.

The SVG lives in this repository's own data files and is never supplied by a reader, so the dangerouslySetInnerHTML on that path sees no untrusted markup.

Verification note, stated honestly

The post builds, and the figure is present in the DOM with all 30 labels and the right viewBox. I could not verify its rendered geometry: the browser pane in this session reports window.innerWidth === 0, so every element measures 0×0 — including body. I nearly attributed that to my own SVG and started "fixing" it. The intrinsic width/height attributes I added are a real improvement, but I am not claiming they fixed a bug I never demonstrated.

Receipts in the post: trinity-fpga #582, #584, t27 #2177, and the two prior-art DOIs (Stakhov 2002, Bergman 1957), both verified through Crossref.

Bilingual, and the first post to carry a diagram: adds a 'figure' block
kind holding inline SVG, plus its renderer case. The SVG is authored in
this repository's data files, never supplied by a reader, so no
untrusted markup reaches that path.

The post reports what was measured today: phi is not observable at all
in the closed 2-bit format (1, 2, pi, 0.001 and 1e12 give identical
tables), a GFTernary dot product is exactly phi^2 times the same codes
read as balanced ternary, and the fp32 multiplier the board uses for
that product costs 1654 cells against 13 for the exact unit -- 16/16
input pairs equivalent, on Yosys 0.65.

The diagram is a difference diagram: the same constant in two
placements, and the one edge that exists on the left and cannot exist
on the right.
The body file was generated by concatenating two blocks that each
carried the import line, so 'Block' was declared twice. My own check
missed it: I ran tsc --noEmit -p tsconfig.json, and the project's
typecheck is tsc -b, which is what the ratchet runs.

Ratchet now: 179 errors across 26 files against a baseline of 184
across 27, and no file gained.
@gHashTag
gHashTag merged commit 92617fb into main Aug 18, 2026
14 of 17 checks passed
@gHashTag
gHashTag deleted the post/phi-is-a-scale branch August 18, 2026 05:29
github-actions Bot added a commit that referenced this pull request Aug 18, 2026
Blog: the golden ratio in this format is a scale factor, not information (#830)

* Blog: the golden ratio in this format is a scale factor, not information

Bilingual, and the first post to carry a diagram: adds a 'figure' block
kind holding inline SVG, plus its renderer case. The SVG is authored in
this repository's data files, never supplied by a reader, so no
untrusted markup reaches that path.

The post reports what was measured today: phi is not observable at all
in the closed 2-bit format (1, 2, pi, 0.001 and 1e12 give identical
tables), a GFTernary dot product is exactly phi^2 times the same codes
read as balanced ternary, and the fp32 multiplier the board uses for
that product costs 1654 cells against 13 for the exact unit -- 16/16
input pairs equivalent, on Yosys 0.65.

The diagram is a difference diagram: the same constant in two
placements, and the one edge that exists on the left and cannot exist
on the right.

* Fix a duplicate type import the ratchet caught

The body file was generated by concatenating two blocks that each
carried the import line, so 'Block' was declared twice. My own check
missed it: I ran tsc --noEmit -p tsconfig.json, and the project's
typecheck is tsc -b, which is what the ratchet runs.

Ratchet now: 179 errors across 26 files against a baseline of 184
across 27, and no file gained.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant