Skip to content

Specification: three physical widths circulate for one TNF rung (16 / 17 / 19), and three instruments each picked a different one #644

Closed

Description

@gHashTag

Not a bug report — a specification question, so nothing has been changed.

Three different physical widths circulate for one rung, and three instruments in this repository each silently picked a different one. The reference oracle settles what it means by TNF16:

>>> TNFFormat(4, 11).sign_shift
18                      # so the layout is 19 bits: 1 sign + 7 exponent cells + 11 mantissa
>>> TNFFormat(4, 3).sign_shift
10                      # TNF8 is 11 bits
>>> TNFFormat(2, 1).sign_shift
5                       # TNF4 is 6 bits

Against that:

where TNF16 TNF8
the rung's name 16 8
tnf_paper.tex's caption ("the rungs store more bits than their names") 17 10
fpga/tnet/tnf16_decode.v 16 —
fpga/tnet/tnf_spec_decode.v → tnf8s_decode — 10
research/arxiv_tnf/conformance/tnf_ref.py 19 11

tnf8s_decode is not only a different width — it is a different field layout (five offset cells and four mantissa bits, against the oracle's seven and three), so it is a different format, not a narrower encoding of the same one.

Why this matters beyond bookkeeping. Everything downstream is priced by physical width: on this substrate an identical multiply behind a decoder costs 128 cells at 8 input bits, 213 at 10, 385 at 16 and 403 at 17 (docs/FRONTIER.md). A table that prices TNF16 at its name is comparing a 19-bit format against 16-bit ones and reporting the difference as a format advantage.

And it hid in a way worth recording. Enumerating an 11-bit sign-magnitude format over 10 bits produces 1,008 finite, monotone, correctly spaced values and not one negative number — a value set that looks entirely healthy and is exactly half a format. It surfaced only when the table was cross-checked against the oracle's own round-trip and disagreed on 98 of 200 samples, which was the negative half.

Three ways to settle it, and the choice is yours:

  1. The oracle is normative — then tnf16_decode (16 bits) and tnf8s_decode (10 bits, different layout) are stale RTL and should be regenerated or renamed, and the caption's "17" becomes "19".
  2. The paper is normative — then the oracle carries two spare cells per rung and the conformance vectors need regenerating.
  3. They are different formats — then they need different names, because tnf8s currently reads as a variant of TNF8 and is a distinct layout.

Our own measurements (#641, #642, #643) now use the oracle's widths throughout, and #642 corrected a frontier that had used the 17-bit module. If option 2 or 3 is the answer, those numbers move and we will re-measure.

🤖 Filed by the autonomous wave loop (W942). Related: #631, #641, #642, #643.

Activity

  1. gHashTag commented on Aug 21, 2026

    @gHashTag
    CollaboratorAuthor

    Resolved by measurement (W954, #668). The three widths are three sources, each correct about a different object.

    tnf_ref exports LADDER bound to v1-research while declaring DEFAULT_LADDER_VERSION = 'v2-spec'. Above rung 8 they disagree:

    rung LADDER (v1-research) get_ladder(default) (v2-spec)
    TNF16 17 bits (m9) 19 bits (m11)
    TNF32 30 bits (m21) 36 bits (m25)

    So TNF16 is 16 by name, 17 by the research ladder, 19 by the spec ladder — and both are in use in this repository: TNFFormat(4, 11) in accuracy_coordinate.py, accuracy_seeds.py, prior_sensitivity.py; TNFFormat(4, 3) in activations.py, conv.py, oracle_rtl.py.

    A larger defect surfaced in the same check. The ladder defines TNF8 as TNFFormat(3, 4) — 10 bits, 30.95 binades. Every rig here instantiated TNFFormat(4, 3) — 11 bits, 126.91 binades, which is TNF16's exponent field with a truncated mantissa. The ladder's eighth rung has never been measured; every published "TNF8" figure describes a different format.

    Suggested fix, in order of value:

    1. Remove the bare LADDER constant, or bind it to get_ladder(DEFAULT_LADDER_VERSION).
    2. Require every measurement to print the (exp_trits, mant_bits) pair and the physical width it instantiated, not the rung's name.
    3. Re-run the accuracy and cell-census rigs against TNFFormat(3, 4) before any TNF8 number is quoted.

    Filed by the autonomous wave loop.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions