Skip to content

Search and Core: derive numbers from declared domains and graph slot layout (−649 register sites) - #188

Open
afflom wants to merge 40 commits into
mainfrom
claude/project-thread-kkzl28
Open

afflom wants to merge 40 commits into
mainfrom
claude/project-thread-kkzl28

Conversation

@afflom

@afflom afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member

Requested by Alex · project thread

What

Two mechanized passes over the hand-written LexLean of the client's Search and Core modules, the counterpart of the coder's domain pass (#180). Register on main: 4,735 → 4,086 (649 sites). Values are unchanged.

1. Domain pass (client/gen/srcdomains.py, new wire module Uorc.Wire.Domains).

  • Uorc.Wire.Domains states no numeral other than 0 and 1. radix is 1 + 1. An octet's bits and a word's octets are radix cubed. Every other quantity is built from radix places: the octet's extent, last value and half, the word's bits, byte places and top (its half twice, less one, because no word holds its own extent), the node names of a binary tree over the octet alphabet, and the seal digest (2^8 bits, counted in octets). Each quantity has an evaluation{} theorem that states its value.
  • The pass parses each definition body and infers the domain of the operand each literal acts with. The sources are:
    • the other side of a comparison, sum or difference;
    • declared parameters (trie walks, fitsAddition);
    • the dividend of a place-value quotient or remainder;
    • declared fields and binders (bit counts, tree nodes, seals, octet digits).
  • A literal is restated only as a quantity of that domain. Radix is used only in definitions declared binary. Every literal the pass cannot place is reported with its context. --check fails unless the sources are at the pass's fixpoint.

2. Slot pass (client/gen/srcslots.py).

  • A kernel graph block names values by slot, and an instruction names an earlier value by its distance back.
  • The pass simulates each written-out block's slots. It states each slot index once (the first is 0, each next one is the one before plus 1), and each reference as the difference of two slot indices.
  • Short blocks bind their slots in let chains.
  • The Huffman program's blocks are split into parts, so their slot indices are module definitions. A 160-deep let chain exceeds the data form's depth limit (LLT4001). The parts stay as they are.

Checks run locally (4 cores)

  • lexlean check --all and lexlean verify --all pass on 93 modules at the pin 2bf5c1e.
  • just figures passes (4,086 sites, none beyond the register). client-export --check passes.
  • xtask validate passes up to audit-bootstrap, which runs only in the SDK.
  • A full client export of this lane plus an earlier gen: derive the coder's figures from declared domains; count every coder figure #180 head succeeded at the pin (3 packages, 4 h 8 min). It was the check of these changes. The export of this exact tree is running, and generated/rust will be pushed here when it finishes.

Not closed here (unmet, with counts in Search and Core)

Class What derivation needs
Work accounting (eventWork, instructionReadWork, plannedWorkBound: 129, 192, 320, …), ~85 Each charge derived from the work it charges, not a cost table
In-program constants (n : UInt64) and block distances, ~90 A UInt64 view of the domains, and the same slot layout for blocks
Builder offsets (Overlay, Residual, LzProgram: slots + 2/4/5), ~30 Offsets taken from the emitted chunk's own layout
Sizes and budgets (ByteBuilder 4096/12/2048, Effort defaults, LzFit 24/128, Synthesis halving), ~40 Coordinates or facets decide them; they are not domains
Dictionary.indexWidth byte counts, other, ~40 Restate it as a recursion over byte places


Generated by Claude Code

The NCM decoder and encoder models read and write their globals at byte
offsets of one state table. The offsets were written out at every access
(2,086 sites of the register). They are now derived: each global's address
is the address of the global before it plus that global's width, and the
width is the one figure of its place. The register falls from 8,252 to
6,166 sites; the IR and its lowering are otherwise unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
Uorc.Performance::parametricRelease names the first release whose
implementation must be parametric (0.7.0). From that workspace version on,
the figures check rejects the tree while the register lists any site, so the
version bump that tags the release cannot land until the distance of
implementationFigures is 0.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
uorc::coordinates names the coder's figures the engine decides (the three
plasticity figures first): the coder reads each as set for a measurement or
at its standard value. frontier's CodecPoint carries the coordinates a point
sets (in its κ only when any is set, so earlier claims keep theirs), and a
run sets them for its compressions and decodes.

The engine's coordinate stage gallops each coordinate from its identity by
doubling while the corpus gets strictly smaller, then bisects the bracket
around the best value. Every value measured is a closed claim under its κ,
so later ticks walk the path free and measure at most hypothesisBudget new
values. A finished search is decided when every family codes smallest at the
corpus's best value, and per-input (a frame-word facet is needed) with its
distance otherwise; model/coordinates.toml records it. `uorc-engine
coordinates` runs the stage alone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
… the budget did not reach has no row yet

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
…allop past the standard value, key stored warmed models by the coordinates

A warmed-model image was keyed by build, kernel, profile, mask and state but
not by the coordinates the model learned at, so a measurement at a
non-standard value could read a model warmed at another value: the encoder
coded a state frame from it and the native decoder refused the payload,
sending the frame to the reference decoder. The image key now carries every
coordinate's value.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
…ched them)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
The decoder's in-process cache of warmed models matched state, profile and
dimensions but not the coordinates, so after a measurement at one value a
frame coded at another decoded from the wrong model: the native decoder
refused it and the frame fell to the reference decoder. The test fails
without the key.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
…best; first stage over uorc-v0.3.0

plasticityGain is decided at its standard 104 (families that prefer another
value tie it). plasticityFloor is per-input with a distance of 1 byte (tiny
codes 1 byte smaller at 22). plasticityHalfLife is per-input: the corpus codes
453 B smaller at 47888 than at 4096, and the families lose 171 B against
their own bests. 71 closed claims in frontier.toml; a rerun reads them all
and measures nothing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
…rged tree (6,151 sites)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
Setting a coordinate changes what every coder in the process reads, so the
warmed-model test ran under the library's parallel tests and moved
stored_warm_keeps_bytes off its bytes (rest, ubuntu-24.04-arm). It now lives
in tests/coordinates.rs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
…one LexLean pin

The three byte-index trees (Types.singletonSequence, ByteBuilder.byteSlice,
Enumerate.natByteSlice: 763 register sites) become one definition,
Types.byteCell = prim update of a one-byte seed at 0, reached through
helpers that match only on parameters. Register 6,151 -> 5,388.

The trees had forced every byte to a literal, and seven reflexivity theorems
over long runs leaned on that: without it Lean's unifier passed 13 GB. Every
parameterless theorem proved by reflexivity (399) is now an evaluation
(LexLean#553), checked by the kernel alone; axiom policies restate what Lean
observes. The authority verifies in 9 minutes at width 3.

The gate, the formal legs, the offline construction and the client export all
run one LexLean, pinned in lexlean.pin.json (2bf5c1e, branch
claude/evaluation-proof: #441, #443, #537, #553); the SDK's 0.4.0 has no
prim update.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
The recipe containers mount a 2 GiB /tmp; cargo install's default build
directory filled it on the arm rest leg. The build now runs under
.lexlean/build/<rev> and is removed once the binary is installed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
…ions

The review-only projection job and the offline construction's own lexlean
calls ran the SDK's LexLean, which rejects the pinned lock (LLC0102).
offline-validation names the pinned binary (LEXLEAN, PATH) before any
lexlean call, so copied roots reuse it; the review job puts the pin on PATH.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
The Justfile's PATH reads lexlean.pin.json through scripts/lexlean-pin.mjs,
so the figures leg, which keeps only its declared inputs, could not
evaluate it. The three files are one factory set, `just`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
… .lexlean

The offline setup installs into its sealed cache and then requires a clean
checkout; a build directory under the checkout's .lexlean failed that
(URP002).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
The SDK refuses a PATH that shadows a tool it declares (PP5401 in
template-check and every prismpm lock check). Recipes now run LexLean
through `node scripts/lexlean-pin.mjs exec`, and the validation scripts
spawn lexleanBinary(); nothing changes PATH. The offline setup builds the
pin under a cargo home of its own, so the sealed cache holds its binary
and no git checkout (whose symlinks it refuses, URP001). The claims
harness accepts the kernel's refusal of an evaluation proof, which names
the owning theorem, beside the elaborator's rfl failure.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
Both spawned `lexlean` from PATH, which is the SDK's compiler since the
pin is named rather than put on PATH; its semantics refuse the lock the
pin wrote (LLC0102).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
The model and xtask linked lexlean 0.4.0 from crates.io while the
compiler ran the pin, so the executed controls staged a runtime without
`LexLeanRuntime.update`, which the pinned compiler now emits for
Uorc.Core.Types. The workspace `lexlean` is now the pin's git revision;
scripts/lexlean-pin.mjs refuses a Cargo.toml whose revision differs from
lexlean.pin.json. deny.toml admits that one git source. The offline
acquisition seals cargo's bare repository of it and drops the derived
checkout, whose symlinks the sealed inventory refuses; cargo makes it
again offline.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
…ree is main's

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
client/gen/domains.py: a model declares each domain of its quantities once,
by its extent (stretch, slot, weight with origin and unit; every stored cell
by its width), and the tables and globals that hold it. The pass infers
every operand's domains to a fixpoint and states each literal that is a
characteristic quantity of the domains it acts with, or of where its value
goes (extent, last, half, origin, unit, byte lanes, a rescale between two
domains), as that quantity derived from the extent. The offset-form mixer
dot product's constant term is derived from the two origins
(domains.origins). Values are unchanged; the kernel graph and basis bytes
are identical; ir_check 25/25 identical.

python3 client/gen/domains.py reports the literals left, by value, with
the domains in scope: the next declarations.

NcmEncode/NcmDecode: 1,827 -> 922 fixed-figure sites; register 5,498 -> 4,593.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
cargo deny (wildcards = "deny") counts a git dependency without a
version as a wildcard. The workspace `lexlean` now names =0.4.0, the
version the pinned revision carries; cargo deny check passes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
Corpus 574,441 B at the standard coordinates. Half-life: best 19,912 saves
145 B over the corpus, family bests differ (per-input, distance 94 B).
Floor 24 and gain 104 stay the corpus best; each is per-input (distance 4 B
and 36 B), so they wait on a frame-word facet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MrzwoXqXQR1FdKaod
Conflicts: scripts/client-export.mjs keeps the pinned LexLean of
lexlean.pin.json (2bf5c1e, which contains #174's 560b649);
model/figures.toml is rewritten by `figures write` over the merged
tree: 4,735 sites (main 5,498, minus the 763 byte-tree sites).
Under the pin: lock --check, fmt --check and check --all pass; the two
src modules #174 changed (Page, Performance) verify.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
client/gen/srcdomains.py types the hand-written LexLean of the client's
Search and Core modules as client/gen/domains.py types the coder's IR. The
domains are declared once in the wire module Uorc.Wire.Domains by three
format identities (radix 2, an octet of 8 bits, a word of 8 octets); every
other quantity is derived there from them: the octet's bit places and
extent, its last and half, the word's bits, byte places and top (its half
twice less one, since no word holds its extent), and the node naming of a
binary tree over the octet alphabet.

The pass reads each definition body as a tree, infers the domain of the
operand a literal acts with (the other side of a comparison, sum or
difference; the parameter it is passed to, declared for the trie walks and
fitsAddition; the dividend of a place-value quotient or remainder; the
field it is added to), and states the literal as the quantity of that
domain it equals. Radix places are taken only in the definitions declared
binary. What it cannot place keeps its numeral and is reported with its
context. --check fails unless the sources are at the pass's fixpoint.

Values are unchanged; lexlean check --all passes on 93 modules.
Register 3,830 -> 3,546 (284 sites) on main + #179 + #180.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D
…ctets

Uorc.Wire.Domains derives nodeBytes (the octets a node name of nodeBits is
written in) and nodeRecordBytes (radix children of it). The pass types a
count of inner nodes (a node less the first inner one) and the binders
HuffmanFit.perform names nodes, so the Huffman tree's size and its
leaf/inner tests are stated from the node domain.

Register 3,546 -> 3,495.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D
Uorc.Wire.Domains declares the seal's digest (SHA-256, 32 octets: the
archive's UORC-SHA256-CHAIN-v1 format identity) and derives the last node
name and the last bit place of an octet. The pass types the binders that
hold seals, the digits of big-endian octet strings and state names, let
bindings of naturals, and the bits past whole octets; var{} binders are
read as the names they bind.

Register 3,495 -> 3,464.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D
A kernel graph block names its values by slot (parameters first, then one
per instruction), and an instruction names an earlier value by its distance
back from the slot it defines. client/gen/srcslots.py simulates the slots
of every written-out block of Search and Core and states each slot index
once (the first zero, each next the one before plus one) and every slot
reference as the difference of the referring and referred slot indices.
A block written in short lets binds its slots in a chain of lets; a block
whose list is split into parts (the Huffman program) states its slots as
module definitions, which keeps each module's data form within its depth,
and its parts stay parameterless. Block references (distances among prior
blocks) and the programs' natural constants are left for the next pass.

Graph bytes are unchanged by construction (each reference evaluates to the
distance it replaces); lexlean check --all passes.
Register 3,464 -> 3,225.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D
…e block's values

Register 3,225 -> 3,181.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D
radix is one plus one; the octet's bits and the word's octets are radix
cubed; every extent is a product of radix places; the seal's digest is
radix to the octet's bits many bits, in whole octets. Each quantity
carries an evaluation theorem of the value the pass places it for, so
Uorc.Wire.Domains states no numeral besides 0 and 1 and stays at zero
under a gate that admits no named numerals.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D
…-kkzl28

# Conflicts:
#	model/figures.toml
#	src/Uorc/Search/Enumerate.lex.tex
…by domains and slots

One export of the merged closure (#179 + #180 + this lane) at pin 2bf5c1e:
3 packages from 79 modules, 4 h 8 min on 4 cores.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D
… main

#180 (the coder domains pass) is still open and has moved on, so this
branch no longer carries its earlier head: client/src, the coder
generators, frontier/coordinates and generated/rust are main's, and the
register is rewritten on main plus this lane: 4,735 -> 4,086 (649 sites).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D
…by domains and slots

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M7pu1nwsA7VLR4o6D

This branch has not been deployed

No deployments
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.

2 participants