Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions research/r437-oq4-caller-binding/.gitattributes
Original file line number Diff line number Diff line change
@@ -0,0 +1,2 @@
inputs/** -text -diff
R437-* -text -diff
228 changes: 228 additions & 0 deletions research/r437-oq4-caller-binding/BINDING.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,228 @@
# The current source-to-physical boundary

Chris R / Clear1nsight, with substantial OpenAI Codex assistance. Apache-2.0.

## 1. Exact source delta

Let S_old be PR3690's R437 source at36ff441 and S_new be the a24f9ff/OQ4
source. Both catalogues are compiled with the same canonical selected compiler,
with its exact661-to659 role adaptation. The adapter retains the source checks;
it does not accept a new catalogue merely because its matrices have equal rank.

The complete ordered records differ exactly at

18189,18464,18920,19854,20212,27776,27777.

The five MOVE records are stage6, helper role644, family209. Their host chart
moves from axes22..29 to30..37. Their residual ranks and source event ordinals
are unchanged. The IDLE records retain their source ordinals and native family
charts, but change their unused-axis projectors from ranks34/34 to26/42.
Thus they move32 paid children from34 to26 and32 from34 to42. Counts and rank
mass are unchanged. This calculation includes owner routes: accepting only the
new histogram would miss five actual source changes.

All COPY records, their identity common charts, ERASE guards and source ordinals
are unchanged. This statement concerns the complete ordered request records,
not equality of incoming dirty payloads. The latter retain the all-function
COPY contract proved in PR3936/3950. The selected source's original661-role
Catalogue still rejects the new source; our adapted Catalogue authenticates
the complete current word/template/profile/digest before use.

The old-data positive control enforces its own exact old word/template/profile
and ordered digest using the same retained compiler. It is not an independent
implementation of that compiler. A discovery-only exploratory run determined
the new digest; the committed adapter has no discovery or disabled-guard path.
Replays recompute the catalogue against the frozen digest and reject the old
digest on new data. SOURCES.json records complete source identities.

## 2. The real two-level generated program

Choose current native IDLE27776 and common COPY1603. The first is the actual
new rank26 idle in window6/family209; the second is stage0/position22014,
frame515/centre475. Each source code comes from the full current Specifications
table, its actual chart and the canonical LowCodes scheme. No two-label
replacement vocabulary or synthetic ancestor is used.

For each request choose low matrix equal to the inverse of its actual
preparation modulo3. Native preparation includes the family chart. Hence the
compiled unweighted projector is the leading-coordinate projector of its
stated rank. With w=1, zero high fields and root unit2, all incoming diagonal
weights remain2. Diagonal scalar conjugation leaves the projector unchanged;
every good pivot equals1. The required good child coefficients therefore equal
2. Coordinates0 and90 in the second child test that independently stated
identity through the actual generated callback.

The exact generated steps are

ROOT width2048, unit2, depth0
-> (good, actual IDLE27776 code, rank26): width832, depth1
-> (good, actual COPY1603 code, rank7): width91, depth2.

Indeed832=26 floor(2048/64), and91=7 floor(832/64). Both widths exceed the
64 cutoff; neither step is merely a terminal descriptor example. The encoded
program contains the two actual source codes, ranks, flags and parent/child
widths. The current COPY prefix is appended after the frozen ancestor field.
The old rank34 label paired with the current ancestor code rejects at
`actual selected rank`; the other actual COPY label in the current prefix
rejects at `original global raw-field code`.

This is an admissible generated source *grammar* instance, not a claim that
the root's native family209 is literally followed by a stage0 COPY in one
flat global schedule. The outer IDLE central recursive child runs another
selected node, which has its own local source vocabulary. We test this nested
descriptor/caller boundary, while the complete recursive node body is not run.

## 3. What the physical protocol establishes

For a complete source call of parent width n and child width c, the good
chunks are the first c head atoms and their reflected tail partners. The
existing Layout.child_extents emits

B'=B, G'=G*A^(2(n-c)), S'=S,

with U in its separate row port. Thus

Z*U*B'*G'*S'*A^(2c) = Z*U*B*G*S*A^(2n).

This is the whole-packet volume identity; it does not discard spectator
addresses or replace U by U*B. The actual fixture checks the encoded grants

(prefixes,U,B,G,S) = (2,11,2,5*3^2432,3)
and then (2,11,2,5*3^3914,3).

Both equal prefixes are generated by the canonical directory producer with
their separate ordinals. It is not a deduplicated header set.

Machine.prepare_child checks that the event names the same parent Program
object, current source code and active geometry code. It copies current,
directory, extended program, prefix count, U and all three extents into fixed
handoff ports. CALL writes the exact successor PC and numerical-role labels,
parks the parent streams and numerical state, then installs the child grants.
The tested current source codes use this unchanged path.

On protected complete_return, the child frame is closed before inverse
parking. The result stays in its distinct output port while numerical and
stream parents restore; it is then transferred to current. Two nested returns
recover the child-parent successor4 and root successor1, without replaying a
prefix. Snapshot checks cover exact numerical objects/bytes/heads and every
canonical role. Additional audits include erased-but-allocated numerical slots
and template heads; clearing live content alone is insufficient.

Within the child parent, the actual COPY dispatch preserves stage475, clones
into copy_source, transfers the returned residual output to private stage659,
and feeds all 63 actual literal XOR readers before clearing it. Each reader
uses the unchanged physical XOR dispatcher on two complete reduced 23-bit
prefix fibres. Source and private output remain immutable through all 65
source events22014..22078. This tests ownership, actual private-result
consumption and the physical disposal scan.

There is also a uniform consumer implication, independent of the test bits.
Let x be the borrowed source function and z the correctly completed residual
result on the same complete packet. For arbitrary, possibly correlated target
functions y_t, the63 literal readers produce y_t XOR z, preserve x and z,
and leave every other staged role unchanged. This follows by induction over
the actual65-event interval: the source/private roles never appear as XOR
destinations, every target occurs once, and the last event is the matching
ERASE. Thus a correct completed residual is consumed without another residual
call or reversed uncompute. If X(V) is the completed physical XOR tariff,
retain the full bill `C_residual +63*X(V)+C_copy/transfer(V)+C_disposal(V)+gamma_window`.
Only under the retained linear XOR/copy/transfer contracts and admitted
metadata bounds does the extra consumer work simplify to O_q(V). The probe
tests this actual consumer using controlled z; it does not supply z's complete
residual contract or erase its cost.
The transported equality guard itself is checked over the complete literal
controller census; the controlled bit transformation is not a claim that it
equals the semantic COPY map. Three malformed physical gates reject before
handoff mutation, followed by exact parent/scratch checks.

These checks establish actual bound protocol execution at the finite stated
context. **They do not establish a full residual body or global node.** The
input is46 bits with an explicitly controlled fibre23, outputs are declared test stimuli, stock11 is a controlled
fixture, and the receipt's global macro count is0. The complete volume at
the encoded root would be660*3^4096 bits. Native gather/route, the adapter
sweeps, complete primitive calls and the ordinary padded wrapper are not
executed on that volume by this probe.

## 4. Uniform caller substitution: the remaining hypotheses

A complete production source binding would use the actual current topology,
Specifications and SelectedBinding in macro_code.compile_node. Its loop order
is schedule, replica, low class, literal event. That finite compiler accepts
only q-dependent labels and emits the same fixed block vocabulary for every
input width/precision; all 64-dimensional low classes are a fixed finite
set for a fixed q. The full expansion is effective but enormous. This package
does not pretend that a sampled class is that complete emitted tuple.

The source-relative structural implication is as follows. Suppose each emitted
NODE, GATHER_ROUTE, RESIDUAL_BEGIN, ordered sweep/leaf, CHILD, route/scatter,
XOR, bridge and exchange has its stated complete Boolean/ownership contract
on admitted encoded inputs; suppose the numerical source callback supplies
the actual source-bound coefficients at every frozen history; and suppose
the existing parking codecs preserve the encoded parent at each matched cut.
Then induction over the finite emitted word and decreasing-width recursive
tree gives the current caller contract. COPY introduces an owned clone and
the transported guard clears its owned result; it never projects arbitrary
dirt away. At a recursive cut the capsule volume identity above supplies
exactly the complete child input, and protected return supplies its one
substitution. The five chart changes and two idle changes are consumed by
the current authenticated bindings, rather than inferred from rank alone.

This implication is not an independent proof of its primitive, callback,
stock or source-level premises. The main physical-contract map still marks
the inherited universal refinement partially reviewed. Its archival presence
does not certify the all-width current program. A uniform source/native/router
cohort must discharge or explicitly retain these premises and connect all
constructors at one source; the bounded probe does neither by itself.

## 5. Complete cost boundary

Let T(j,w) be the inherited uniform completed-child cost per *complete payload
bit*, including coefficients, owned output, restoration and head returns.
PR3950 supplies the rank7 low-class bound

delta7 <= min(1,448*(q-1)/(q*(q-2))).

Retaining its completed good/fallback contracts, the residual cover bill is

T(7f,w)+delta7*49911*max_(1<=j<=f)T(j,w)
+ A_q*(f+leaf_toll(w)+1) + gamma_orig/(N*v).

Here gamma_orig includes encoded source/history/Count preparation, transfer
and cleanup not already charged by completed primitives. The initial clone,
output commits and final disposal remain paid. Gather/route/scatter and the
containing node's producer/wrapper work remain additional. Neither N low
classes nor repeated high prefixes create an extra coarse-child factor;
normalization is by the full complete cover volume N*v.

At source-admitted q>2^80 the bad fraction fits10^-16. The familiar cheap
local toll additionally needs the canonical root/history, numerical preparation
and backend reserve arguments. This package does not supply them merely by
changing the request digest. The fixture's q=3 is not an admitted exponent
parameter and its finite parking instruction counts are not an asymptotic
constant. No ordinary/native/router tariff is inferred from them.

For storage retain the full logical frame envelope M_j, the fixed owned
role/backend inventories, current and protected output, and the complete active
parking stack. Only if the inherited admitted bounds prove

sum_j M_j <= C_q*V_root + F_q

does the same fixed-role source lowering have O_q(V_root)+F_q visited cells.
All physical inventories must include cleared-but-visited backing, markers and
templates. A live-state snapshot or peak10 scratch leases does not prove
this uniform bound, the sharp logical coefficient11, or a reserve of nine
family packets. Interruption recovery is also outside this successful-return
contract. These are precise remaining obligations, not costs set to zero.

## 6. Attribution and next useful result

The adapter comparison extends zxv/PR3690's exact source adaptation and
controlled return. The physical protocol, source compiler and proofs remain
David Arias/Jose Manuel/ywz65 and their retained lineage; R437/OQ4 remains
ai-nanachi and credited predecessor work. PR3936/3950 supply the immediate
all-function and conditional-cost boundary. Full notices are retained.

The next substantive result is the complete production constructor/refinement
join at this authenticated current catalogue, including actual native/router
consumers, original-width stock and root/precision constraints. Another finite
controlled return would add coverage, but would not discharge that join.
Loading