Skip to content
Open
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
62 changes: 62 additions & 0 deletions research/qq16-protocol-comparison/ADMISSION.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,62 @@
{
"schema_version": 1,
"domain": "One source-admitted nonleaf selected root with actual complete packets; all descendants retain original C.",
"boundaries": "DECISION.md / PROTOCOL.md",
"items": [
{
"item": "Original arithmetic, descriptors, row stock, geometry, fine quantum and H_b",
"status": "inherited unchanged",
"source": "physical-complex-controller / frozen4166",
"cost": "source preparation and arithmetic matched literally for A/B; no conversion or recovery added"
},
{
"item": "Original embedded versus isolated fresh-engine cancellation",
"status": "deferred",
"source": "PhysicalArena.available persistent order",
"cost": "No universal native-body cancellation from closed flags; unresolved encoded-slot equivariance is retained"
},
{
"item": "Shared primitive bundle and allocator-history relation",
"status": "new implemented source refinement; matched complete direct and targeted nonleaf execution passed",
"source": "adapter.py / PROTOCOL.md",
"cost": "one additional controller bank/roles/Count pool; original closed scalar/leaf/monomial/Ordinary bundle used exclusively in original invocation order"
},
{
"item": "Extra controller inventory, initialization and retirement",
"status": "new source-derived bound with literal controller inventory checks",
"source": "resources.py / RESOURCES.md",
"cost": "1023fixed controller tapes; setup/terminal<=66496R/W/moveactions; historical visited backing retained"
},
{
"item": "Full-root original occurrence partition",
"status": "new literal census",
"source": "ledger.py",
"cost": "all510662Gamma original calls; retained493097Gamma later calls explicit"
},
{
"item": "QQ16 incremental comparison and suffix",
"status": "conditional threshold",
"source": "COMPARISON.md / BILL.json",
"cost": "Header component exact; original-to-isolated-native allocator context differences, changed cores, service and nonchild costs remain explicit"
},
{
"item": "Guarded internal child-return adoption",
"status": "new local mechanism; full direct/nonleaf, alias, generation and retirement checks passed",
"source": "adapter.py / BASELINE.md",
"cost": "11n+4 read/write/move saving minus separate finite selector control; available equally to compatible variants"
},
{
"item": "Whole-machine Turing step / elapsed runtime",
"status": "not established",
"source": "PROTOCOL.md",
"cost": "Separately compiled finite-control rates must be bounded; no subtraction of unrelated upper bounds"
},
{
"item": "Production O73 payload/full RootLayer.run, all-node recursion, supplier/bootstrap/kappa",
"status": "outside new claim",
"source": "selected-result.json / inherited contracts",
"cost": "No exponent change; selectedkappa unchanged"
}
],
"publication_gate": "One substantial source-bound context/comparison/mechanism contribution only after frozen integrated assisted review and registered/export checks."
}
128 changes: 128 additions & 0 deletions research/qq16-protocol-comparison/BASELINE.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,128 @@
# Fair original-arithmetic external baseline and protected-return adoption

The primary baseline B changes only the original physical complex controller's
child protocol. `adapter.install` retains the existing `RootLayer.machine` as
the recursive endpoint and constructs only an extra controller: bank/packet
roles, private fields, encoded roles/capsule, fixed scratch, finite header and
parking stack. Both controllers share **the same physical scalar, leaf and
monomial/Ordinary bundle**, lending it sequentially at closed primitive cuts.
The existing original binding and scalar catalogue are retained. The `RootLayer`
plan, current and output ports do not change. No extra Ordinary, decoder,
acquisition store or service Machine is allocated. The production extra
controller has 1,023 physical tapes.

This sharing is essential to the all-size common-body argument. Equal widths
and payloads do not imply equal action traces when independent Ordinary arenas
have different encoded free-list histories. The independent `shared=False`
configuration is preserved as an explicit measured experiment only; its finite
matching outputs/actions do not establish arbitrary-context cancellation. In
primary B, source primitive calls occur in exactly the original order on exactly
the original engines. Finite PC/root/stack context remains separately compared.
The additional controller changes where suspended parent state is stored; it
does not restart or independently initialize the arithmetic allocator.

The primary boundary is the same complete prepared caller entry and return.
At each child the smaller matched boundary begins before `prepare_child`, and
ends at the next caller instruction with the returned target, borrowed source,
other packets, complete current bank, encoded descriptors, and dirty-state
obligations in the same state. Source constructor work and public/root copies
remain part of the complete boundary, even when a paired subtrace cancels.

## New constructive mechanism

Variant B-adopt uses the same arithmetic and external calls as B. On successful
completion the dedicated child's output is an owned protected internal tape;
the parent target is another owned tape whose prior value is dead. The adapter:

1. Physically clears the old target and erases its marker.
2. Installs the returned tape at the same finite target slot, in `bank.packets`
or `private` as appropriate.
3. Rebuilds `parked` and `parking.roles` in the original logical role order.
4. Sets the dedicated backend's output role to the cleared old target.

The target slot, source group, saved PC and mathematical value are unchanged.
`bank.selection` stores slot numbers and `bank.source` refers to the unchanged
current bank. Runtime instructions retrieve targets afresh through `packet`.
The dedicated child output is internal to this composed construction; the
actual parent's output and `RootLayer` public ports retain physical identity.
There is no extra tape, no payload-dependent role selection, and no reclamation
of high-water backing by erasure. The role permutation is compiled finite
control, with its cost kept separately from read/write/unit-move actions.

For returned length n and old target length m, the source transfer has action
vector `(5n+2m+7, 2n+m, 4n+2m+10)` in reads, writes and moves. Clearing the old
target has vector `(2m+3,m,2m+4)`. Both routes erase one LEFT marker afterward.
The adoption also checks both canonical LEFT origins, costing `(2,0,4)`.
Hence the new return site saves **`(5n+2,2n,4n+2)`**, or **`11n+4` cell actions**,
minus the separately charged finite rebinding difference. This is not #4166's
already credited `11B_bank+4` whole-bank commit. The new swap applies fairly to
B, C and D whenever their dedicated output has this exclusive internal role.

The final `retire_installed` sweep physically erases the extra controller
markers after its last use and forbids subsequent reuse. It preserves the
completed outer public output. Repeated generations occur before that final
sweep; historical backing remains charged.

The finite checks reject a returned alias, stale `parking.roles`, or repeated
caller physical roles. Reentrancy, primitive closure and capsule/header
validation remain charged. The complete resource proof also excludes escaped
raw aliases; no lease revocation mechanism is asserted for arbitrary Python
clients.

## Reproduction and scope

`comparative_fixture.py` uses the authenticated retained actual Tape, Count,
Machine, Ordinary, Monomial and coefficient implementations. Its tensor binding
is the original reduced physical-controller fixture and its arrays are test
oracles. Direct and nonleaf calls execute native arithmetic, including nested
embedded children; they are not endpoint doubles. The bounded nonleaf fixture
executes the actual source prefix, completes its first child and all ten
grandchildren, executes the next source scalar instruction, and then performs
controlled source cleanup; it does not claim a completed nonleaf root. Two direct generations test
reuse after returned-output role exchange. Every pair checks exact output,
borrowed values and heads, fixed public output identity and final closure.
The bounded nonleaf entries are correlated multiples of four with real and
imaginary components at most twelve in magnitude. Source owner maps here only
permute addresses. In the retained complete two-axis tensor fixture, the helper
C1 has component bound 24; its cancelling ADD/subtract pair has bound 36; the
joined second family has bound 24; its two-axis C has bound 96; the final
subtraction equals C of the original second family, with bound 48. The next
actual parent ADD has bound 60. Thus signed component width 8 contains this
entire executed interval, and the quarter-grid divisibility is exact. This
reduced numerical witness is separate from the retained production width and
source-entry theorem.

A separate reduced continuation fixture executes both exact bridge instruction
forms, with original arithmetic calls before and after them. Multi-prefix,
spectator and multiple-coefficient data are tested together. A real zero-child
same-output case is retained: externalizing an already direct leaf buys no call
saving and incurs the extra entry check and retirement. The production theorem
therefore uses its literal nonleaf census and complete setup/retirement bound.

The production O73 control is covered separately by the source census and
quantified semantic refinement. Complete production `RootLayer.run` is not
claimed as a finite execution.

## Credit and open next questions

Jose Manuel's #3385 supplies all original complex operations; #4166 supplies
the external boundary which B isolates without recovery. Sagar Thakrar's #3464
supplies the source-relative parent restoration invariant. Tekalign's #3380
supplies the finite ownership-exchange idea, adapted here with a new site,
complete role bindings and a different literal bill. Those predecessors' gains
are not added to this site's gain.

The next comparison must sum the exact source occurrence deltas, including the
later root suffix and every root/service packet geometry. The QQ16 incremental
comparison is against this compatible improved original-arithmetic baseline,
not against a baseline handicapped by unused decoder inventory or fixed copies.

The measurement harness sums the literal pre-existing per-Tape primitive meters
on a compiled exhaustive physical inventory. It rejects construction of a new
Tape within a measured interval, so dynamic allocation cannot silently evade
the census. Native arithmetic allocator history is checked separately at each
prepared entry and completed return: the same primitive bundle's ordered free
queues, live leases, heads and meaningful symbols must match. A closed-free-list
mutation changes that fingerprint even when the numerical state is blank. These
finite checks support the source refinement; equal fingerprints alone are not
an all-size execution theorem.
Loading