Repository navigation
conformance: promote gf256 -> strict SW-bitexact (last rung, 75/0/8) - #1399
Merged
Merged
Conversation
… rung, 75/0/8) gf256 (GoldenFloat256: S1 E97 M158, BIAS=2^96-1, u256_software) promoted from bitexact_selfconsistent to strict bitexact. This closes the last selfconsistent rung of the GoldenFloat ladder; horizon-A SW ceiling reached (75 bit-precise; 8 structural remain terminal, no single decode law; 83/83 SW-bitexact NOT achievable). Bias hold lifted: the 2026-07-05 bias audit resolved the earlier 'open bias R&D' hold (#1386). Decode uses ONLY the closed-form interchange bias 2^(E-1)-1 = 2^96-1 (identical rule to gf128/gf512). The descriptive PHI_BIAS spec metadata is NOT part of the decode path; no decoded value depends on it. Witness chain [verified SW], abs_error=0: - dyadic normalizer (gf_wide_independent_witness.py): 2021/2021 bit-exact - Fraction oracle (conformance/witness/gf256/gf256_decode_ref.py): 2021/2021 exact - cross-check dyadic==Fraction over 201512 representative codes (seed=256): agree - analytic separation-bound (SEPARATION_BOUND.md): M=158>>52 -> no FP lowering, every finite value is an exact dyadic odd*2^k, no rounding. INDEX totals: bitexact 74->75, selfconsistent 1->0, structural 8 (sum=83). NOT on-silicon Tier-E: gf256 is u256_software, has NO RTL; Tier-E ceiling 71/83 (trinity-fpga #199) unaffected. Closes #1397 Refs #1386
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
This was referenced Jul 5, 2026
Closed
This was referenced Sep 28, 2026
gHashTag
added a commit
that referenced
this pull request
Oct 4, 2026
Nothing in bootstrap/src or bootstrap/build.rs uses candle. The two lines came back when bootstrap/Cargo.toml was re-created in #1399, after #980 had already flagged them as dead weight, and a one-sided dependabot bump (#1470) left two candle-core versions in the graph. Clean `cargo build --release -p t27c` on an M1 Pro: 142.7 s -> 97.9 s (110.9 s on a second clean run), 386 -> 279 units. Cargo.lock loses 74 packages; no version changes. `t27c suite` failure set is identical before and after (TOTAL 662, PRIMARY 95, BLOCKED 138). Closes #5899 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 4, 2026
Merged
dmitrii-f-t27
pushed a commit
to dmitrii-f-t27/t27
that referenced
this pull request
Oct 4, 2026
…g#5904) An on_comb parameter becomes a module port only when its width can be derived exactly, and the port is as wide as the input the generated function declares for it (FR-001). One definition decides this -- VerilogCodegen::entry_port_width -- and the entry-points census calls it instead of keeping a private is_sized_primitive list. - Struct ports: a struct whose fields are all strictly sized is a port of the summed field width (no `_ => 32` default). The width is checked against packed_width_in on the same struct map, so port == function input. Nested structs, slices and f32/f64 stay refused. - uN / iN for 1 <= N <= 128 are N bits in ports and type_to_width; type_is_signed accepts iN. - A refusal is reported on stderr as `t27c gen-verilog: ENTRY POINT REFUSED -- <param>: <type>`; the in-file comment and the exit status are unchanged. - A parameterless on_comb is called with W530's 1'b0 placeholder (`on_comb()` failed iverilog). Reachable now that `-> u1` is sized; also fixes d_full_path.t27 and d_static_one.t27, which already failed iverilog on master. Measurements (release t27c, 1174 specs): - ternary_mac + on_comb(acc: i32, a: i8, w: TernaryWeight): NO DATA PORTS 0, `input wire [7:0] w` 1, yosys 222 cells. - led_off_test.t27: ENTRY POINT REFUSED 0, `function [31:0] on_comb` 0; iverilog rc 0, yosys 0 cells (result tied to 1). - slice spec: stderr refusal lines 1, exit 0. - Corpus gen-verilog diff vs master: 8 specs, stderr and exit codes identical for every spec. - Odd widths: igla/training/low_bit_ternary (i2 -> signed [1:0]), numeric/formats (u5, u4 parameters), numeric/gf4 (u4), tri/crypto/base64 (u6 locals), tri/utils/utf8 (u21). - led_off_test no longer refused (u1); d_full_path and d_static_one call on_comb(1'b0). - No corpus on_comb takes a struct. Corpus specs refused: 1 -> 0. - t27c entry-points: FORCED_ROOT 3 -> 7, FORCED_ROOT_WIDE 33 -> 29, FORCED_SCALAR 22 -> 25, FORCED_WIDE 55 -> 52; named-type blockers 105 -> 84; HAS_ENTRY 95, NOPARSE 12, NO_CANDIDATE 469 unchanged. - Seals: 9 files for the 5 odd-width specs resealed with `t27c seal <spec> --save`; check_seal_coverage.py OK (1314 hold). - FROZEN_HASH resealed with `t27c frozen-digest`. - cargo test -p t27c: 2797 passed, 2 failed (the 2 known baseline failures), vs baseline 2789 passed; +6 tests in tests/verilog_struct_entry_ports.rs, +2 entry_points unit tests. gen/verilog/ is a stale snapshot (last regenerated at gHashTag#1399), so it was not regenerated here. Closes gHashTag#5904 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 6, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Promote gf256 -> strict SW-bitexact (74/1/8 -> 75/0/8)
Promotes
gf256(GoldenFloat256: S1 E97 M158, BIAS=2^96-1, u256_software) frombitexact_selfconsistentto strict SW-bitexact. Last selfconsistent rung ofthe GoldenFloat ladder -> horizon-A SW ceiling reached (75 bit-precise; the 8
structural formats are terminal with no single decode law; 83/83 SW-bitexact is NOT
achievable).
Resolves the bias hold (#1386)
Research issue #1386 required resolving the bias-definition before any promotion.
The 2026-07-05 bias audit resolved it: the decode uses ONLY the closed-form
interchange bias
2^(E-1)-1 = 2^96-1(identical rule already applied to gf128 =2^48-1 and gf512 = 2^194-1). The descriptive
PHI_BIASspec metadata is NOT part ofthe decode path; no decoded value depends on it. Decode-definition is definitive.
Witness chain — [verified SW], abs_error=0
gf_wide_independent_witness.pyconformance/witness/gf256/gf256_decode_ref.pySEPARATION_BOUND.mdconformance-integrity-gate(wp18)Two structurally independent exact decode paths (integer normalizer vs
Fraction-significand + symbolic shift) -> genuine second witness. OOM-safe (±2^96
exponent kept symbolic).
Changes
conformance/vectors/INDEX_all_formats.json: gf256bitexact_selfconsistent->bitexact; counts 74/1/8 -> 75/0/8; sha256 + n_vectors updated.conformance/vectors/gf256_conformance_v0.json: promoted pack (2021 dyadic vectors, promoted schema withwitnessesarray).conformance/witness/gf256/{README.md, SEPARATION_BOUND.md, gf256_decode_ref.py, cross_check_representative.py}: new witness dir.docs/NOW.md: promotion entry.Honesty scope
[verified SW]only. gf256 isu256_software(NO RTL) -> no decode-HW/compute-HWcell exists for it; the Tier-E ceiling 71/83 (trinity-fpga #199) is unaffected.
Catalog remains 83 formats. No categorical claims.
Closes #1397
Refs #1386