Skip to content

The reference evaluator reaches totality: 611/611 programs, abstains 52 to ZERO - #54

Merged
O6lvl4 merged 21 commits into
mainfrom
ref-fan
Aug 27, 2026
Merged

The reference evaluator reaches totality: 611/611 programs, abstains 52 to ZERO#54
O6lvl4 merged 21 commits into
mainfrom
ref-fan

Conversation

@O6lvl4

@O6lvl4 O6lvl4 commented Aug 27, 2026

Copy link
Copy Markdown
Contributor

Stacked on #53. Five arcs finish the abstain burn-down — every one of the
611 corpus programs now evaluates
, byte-identical to almide 0.59.1 on the
full (exit, stdout, stderr) tuple, with whole-corpus snapshot diffs proving
zero regressions at every step. The abstain ledger is EMPTY (ceiling 0).

Arc 1 — deterministic time / fan (52 → 34). ALS-DT2 gains the CM-1 v0.3
charge table as prose; the meter (EIP-150 min-cap, C-320 exit bookkeeping),
DT1 time algebra, fan.bounded/race/timeout — unit-exact at every pinned
boundary. fan_race_mapper (native walls, @xt-allow) now has its normative
values pinned by the ref.

Arc 2 — sized numerics (34 → 22). C-038 typed-position narrowing, C-179
UInt64 upper half, C-180 width wrap + per-width MIN/-1 abort, C-182 Float32
(the widened-f64 spelling — measured), C-190 captures, C-195 checked matrix,
T24 exactness, C-210 NaN canonicalization.

Arc 3 — the long tail (22 → 3). Bool/tuple/list/Option total order,
guard-else EARLY RETURN with Result if-let, named-arg record constructors,
the Repr derive, C-064 assign auto-unwrap, C-077 dependency-ordered globals,
the C-062 RawPtr bridge, C-213 endian-argument writers (plus a measured
copy_within signature fix), C-041 arena checkpoints, C-031 json path writes
with the full autovivification matrix, C-132/C-226 mut-param write-back, and
the rope/attention kernels — all bit-exact.

Arc 4 — the ALS-D4 regex engine (3 → 1). Transcribed verbatim from the
normative oracle (runtime/rs/src/regex.rs): backtracking matcher, empty
alternation arms, brace quantifiers with literal fallback, captures with
whole-match index 0, zero-width advance rules. The 90-case fuzz batch is
byte-identical.

Arc 5 — the registerized integer-loop tier (1 → 0). The last abstention
was range_bind_huge.almd's 2^32-iteration cell — 4 minutes of general
tree-walking, ruled unaffordable by the fixture itself. The tier compiles a
for-range body of wrapping-int assignments to RPN over i64 registers and
EXECUTES EVERY ITERATION (no closed-form shortcuts) in ~11s, inside the
existing fuel envelope at 1/64 charge; the deterministic METER is never
bypassed (metered regions keep the general per-head charges). The one cell
the interpreter "could not vote on" now votes: huge=9223372030412324865,
byte-identical to 0.59.1, and the whole-corpus snapshot shows exactly one
changed program. The fixture's abstention prose is updated to the new
reality.

Ratchets (own commits): 52 → 34 → 22 → 1 → 0. 547 stdlib functions.
R1's「参照評価器の棄権 → 0」line item is DONE.

Author: Claude agent (session als/ref-codec). Verifier: gates only — no
independent human review (QUALIFICATION.md limitation 1).

🤖 Generated with Claude Code

@O6lvl4 O6lvl4 changed the title The ALS-DT fan family lands in the reference evaluator (abstains 52 to 34) The fan family and the sized-numeric family land in the reference evaluator (abstains 52 to 22) Aug 27, 2026
@O6lvl4 O6lvl4 changed the title The fan family and the sized-numeric family land in the reference evaluator (abstains 52 to 22) The reference evaluator reaches totality: 610/611 programs, abstains 52 to 1 (the ledgered floor) Aug 27, 2026
@O6lvl4 O6lvl4 changed the title The reference evaluator reaches totality: 610/611 programs, abstains 52 to 1 (the ledgered floor) The reference evaluator reaches totality: 611/611 programs, abstains 52 to ZERO Aug 27, 2026
@O6lvl4
O6lvl4 changed the base branch from ref-matrix to main August 27, 2026 03:58
O6lvl4 added 21 commits August 27, 2026 12:58
…ize and 2^28 ceiling, C-282 index abort vs reduction identity, C-228/C-229 f32/Q1_0/Q8_0 selectors and loaders with OOB-to-zeros, C-270 fp16 scale domain and the accumulate-into-zeros sign ruling, C-223 pow through vendored libm; gelu abstains until the deg-6 fast-exp is pinned
…x stdlib classes burn to the kernel frontier (gelu, softmax_rows, rope_rotate, masked_multi_head_attention, rms_norm_rows)
…o-even range reduction and the bit-shift 2^k, softmax row walk (left-to-right sum, reciprocal-multiply), gelu identity form and associations; C-223 re-points to it
…exp (ties-to-even, unfused, wasm NaN-propagating clamp), softmax_rows and gelu — both fixtures byte-identical to 0.59.1
…outside it, non-finite display forms, signed zero preserved
…96 digit counts through the exact big-integer expansion, the T6 abort outside, inf/-inf/NaN display forms
…to the piped result, postfix ?? suppressed only inside a |> RHS (pipe_probe against 0.59.1)
…es, non-empty errs), fs.is_symlink lands, bytes.to_string answers the ok/err carrier with the validator message, immutable list.pop abstains
…n position answers Result[B, String] (EffectWrap), and in-place list.pop answers some(last)/none
…94 inliner exemption, loop-head checks 1 each, bulk concat 1+len/16, check-then-charge, the ledger-constant Err message
….3 meter (fn entry with C-294 inliner exemption, loop-head checks, bulk-concat dynamic charge, check-then-charge, EIP-150 min-cap with C-320 exit bookkeeping), time values and DT1 algebra, fan.bounded, fan.race block and mapper forms with the DT3 winner rule and C-205 trap window, and the DT5 fan.timeout ends
…ped-position narrowing through retag (lists, tuples, record fields), C-179 UInt64 upper half (unsigned div/rem/ordering/printing, BigInt literals), C-180 width-wrapping arithmetic and the per-width MIN/-1 abort, C-182 Float32 at-birth narrowing with the widened-f64 spelling, C-190 sized captures, the C-195 checked-conversion matrix and ALS-T24 exactness
…22, the sized stdlib and uint64-upper-half classes burn
…/list/Option total order (C-053/C-099), guard-else early return (C-035/C-224), Result if-let, named-arg record constructors (C-175), Repr derive (C-060), C-064 assign auto-unwrap, C-077 dependency-ordered globals, the C-062 RawPtr bridge and C-213 endian-argument writers, C-041 arena checkpoints, C-031 json path writes with autovivification, C-132/C-226 mut-param write-back, the rope and attention kernels, and the ALS-D4 regex engine transcribed from its normative oracle
…e remaining abstention is range_bind_huge's by-construction fuel exhaustion, ledgered in the fixture itself
…2-iteration cell: a for-range body of wrapping-int assignments compiles to RPN over i64 registers and executes every iteration (11s, was a 4-minute fuel exhaustion); the meter never bypasses it, fuel charges at 1/64; the fixture's abstention prose updated to the new reality
…the reference evaluator abstains on nothing in the corpus
@O6lvl4
O6lvl4 merged commit 6d1d8d5 into main Aug 27, 2026
1 check passed
@O6lvl4
O6lvl4 deleted the ref-fan branch August 27, 2026 04:00
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