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
101 changes: 101 additions & 0 deletions research/final233-portable-conditional/CONDITIONS.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,101 @@
# Conditional claim and unresolved applicability

The finite signed application supports κ ≥ 37/10000 **only within the retained
framework and a supplier satisfying the complete contract below for this exact
candidate**. We do not establish that such a supplier exists. In particular,
this is not a drop-in improvement under an already admitted unchanged entrance
or erasure premise. No selected record, common-U implementation, formal-kernel
acceptance, independent human proof or top-level compiled multiplier is supplied.

## Five inherited framework assumptions

- `native_all_columns`: retained native operators, phase conventions, spectators
and whole-job adapters work at every positive column multiplicity.
- `owned_erasure`: owned streams and suspended centers obey allocation, erasure
and ownership contracts.
- `compiler`: mixed-target compilation, chronological product, completion and
splitting satisfy uniform closure.
- `precision_tape`: fixed-tape movement and precision/storage prescriptions
apply to this new finite template.
- `analytic_recovery`: analytic decomposition, fine-prime avoidance and exact
integer recovery follow the retained outer framework.

These are assumptions, not five new proofs. Their presence does not prohibit a
clearly scoped research archive contribution; it does limit the conclusion.

## Candidate-specific common-U and entrance contract

`initial[457]` changes from chart 3 to chart 281. H5B, #7633 bookends, joint E4
resettlement, permutation and exact-zero join reconstruct the declared word.
An actual caller must furnish this changed entrance and preserve the role242
cancellation/return interfaces at every admitted recursive width/column count.
Algebraic reconstruction is not physical admission of that changed interface.

The source #7633 common-U declaration is preserved verbatim in
`custody/7633-COMMON-U.json` (205 source-candidate lifetimes there); its applicability
does not transfer automatically to our 233 lifetimes. For final233, one common
supplier must cover every tagged allocation, complete private coded extent
including represented blanks, arbitrary old private junk at INIT, and every
reachable admitted rounded residual state. INIT resets its entire assigned
extent; CLEAR resets that extent and returns the child atomically, preserving
public fields, spectators, promises and every other live bank, with no aliasing.

The combined lifetime charge is 4,000 ordinary units per lifetime pair; retained
non-MOVE allowances include 32 scalar services per INIT or CLEAR and coefficient
2^20 per service. These are declared normalized tariffs, not measured physical
steps at arbitrary widths. All width, size, storage, addressing, precision,
reset, movement, return and recursion factors must be paid. No hidden recursive
children, productive credits, free scans or extra uncharged MOVEs are admitted.
Keep 68 labels, 1,156 signed banks and the fresh precision context 2,458 (within
the retained 4,096 cap). The full ordinary bill must remain below 10^9.

## Concrete failures a provider must avoid

**Zero scalar rows do not reset physical noise.** The exact-zero lifetime join
removes an ideal CLEAR/INIT boundary. The retained residual fixture injects ε
at field1568 before the removed boundary. Final events 28870 and 29353 carry it
to field440 as -ε before CLEAR29354. Clearing the private field does not erase
the already propagated public error. This is an abstract perturbation response,
not proof that an actual provider reaches that state. The final program's
precision envelope must propagate the error; the removed boundary does not
justify resetting its majorant to zero.

**Exact scalar service domains differ from rounded MOVE contracts.** Final
events45330/45331 are `MOVE 457,966->218` followed by
`ADD 210,457,-1/2,chart218`. Under the inspected one-unit MOVE defect contract,
a one-tick perturbation may reach the half-tick division. The retained #6968
exact ADD table fixture accepts 2 ticks and rejects 1 tick at widths3/17/65.
Therefore that exact service is not proved to implement the broader rounded
contract. A provider must guarantee the componentwise divisibility/signed-range
return domain, or supply a total rounded service with suitable error and cost
bounds. Refining a grid alone does not resolve a unit defect on its last tick.

**The existing generic compact ADD is not a proved same-tariff replacement.**
For equal numerical data volume B per ADD and unchanged other bill items,
the inspected #6998 forced-action calculation requires at least
ρ = 141657821/61035029 ≈ 2.320926578 primitive read/write/moves per tariff unit
per B; ρ=1 yields a replacement ledger of at least1,322,491,168, exceeding10^9.
With the same flat-row-fetch unit the necessary ratio is about2.703351124.
This limited comparison uses upper/model charges for the other items, so it
is not a lower bound on every possible multiplier or all physical runtimes.
Other implementations, widths, volumes, clocks or a finite-cutoff route need
their own complete finite upper bounds. See the retained residual/tariff reports
in `context/`. We do not claim those particular suppliers meet our premise.

## Verified scope and exclusions

The optional finite233 replay checks exact conjugated scalar coordinates on
grid1125*2^25; MOVE is scalar identity, INIT creates zero, CLEAR discards an owned
row. All233 recoverability expressions use distinct other currently live public
fields. This is not free physical uncomputation/reset. The pinned output-row
hash matches prior evidence; that audit does not independently derive intended
multiplication from scratch. Inherited `other_rank` metadata is not a new rank
proof. The original CLEAR builder's insufficient local self/live/division
guards are addressed for the actual finite word by the separate scalar replay,
not by rewriting the original builder or imposing our helper on everyone.

Exact phase/column checks and the finite modular physical replay have different
quantifiers. Rounded/noisy registers, MOVE transport, complete non-J caller IR,
full private-extent services and recursive top-level lowering require the
supplier/framework work above. Archive intake preserves this evidence; it
cannot discharge those obligations or select a new frontier.
Loading