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
31 changes: 31 additions & 0 deletions research/qq16-output-stock/PREDECESSORS.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
{
"canonical_base": "20ecf2891d80b5acf18d980b2164fb44ea7faa90",
"canonical_sources": "All five SOURCES input paths authenticated before parsing; source executables read only as provenance.",
"pending_data": {
"pr": 4015,
"head": "0fff81d8997cef646538bcd4f60d7d455cf93b8e",
"source_package": "research/qq16-streaming-recovery",
"local_path": "vendor/pr4015",
"files": {
"WORD.json": {
"sha256": "272d8c1bd172c29044eb227ec6d67a882501250978234685e035158f4479aa58",
"git_blob": "9a57a6f9c2fd8cdc4e5934408d78ef54e5f266b6"
},
"RESULT.json": {
"sha256": "b4af55d3db3212e79f8f31108e048c5c2d4c32d520fc980ebaa66f47933ad622",
"git_blob": "e9b6b8ea6f5a7dc47d8547cd7ed255e3d5864f7d"
}
},
"reuse": "Exact word and invoice data; no pending executable; final F17 algebra inherited, not independently replayed."
},
"lineage": [
"PR4007 output pebbling",
"PR3985 fusion",
"ywz65 PR2707 / Arham Tahir PR2681 / Ali Romman QQ16 lineage",
"commdata2338 PR2651",
"Nick Petalas PR3339",
"Jeremy Kothe, Rohan Arun, Hao Yan and selected hosting/invoice contributors"
],
"license": "Apache-2.0",
"ai_assistance": "Substantial OpenAI Codex assistance in source tracing, model construction, code, proof and validation. Automated peer review is not human or formal-kernel review."
}
165 changes: 165 additions & 0 deletions research/qq16-output-stock/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,165 @@
# Actual stock and the missing exclusive lease

This audit closes a bounded allocation question, not the QQ16 producer or
physical supplier theorem. The source is canonical main
`20ecf2891d80b5acf18d980b2164fb44ea7faa90`; the demanded word is PR4015 at
`0fff81d8997cef646538bcd4f60d7d455cf93b8e`. Only its exact WORD and RESULT data
are retained. No pending executable is imported, and the decoder algebra and
native kernels are inherited rather than replayed here.

## 1. Actual caller and stock

The selected replay archive's `t16/global/hosted/GLOBAL-TEMPLATE.json.gz`
contains16 replicas and144 complete local invocations. Replica0 begins
F0 (source event0), B1 (event1), pair1 idle preparation (event2), and bridge1→0
(event3). Its B1 invocation uses provider `gauge44`, orientation backward,
pair0 and stage1, at asynchronous time4. This is an actual selected boundary;
it is not silently replaced by the QQ16 producer.

The template and its independently retained finite invoice agree on:

| Stock | Half-open physical family interval | Count |
|---|---|---:|
| Productive data/owner families | [0,14336) | 14336 |
| Private centres | [14336,14344) | 8 |
| Reserved tableau | [14344,14824) | 480 |
| Coefficient work | [14824,14826) | 2 |

The invoice's physical stock identity explicitly includes the480 local-helper
reserve. The hosted helper roles themselves occupy productive families below
14336. The final template has no dedicated helper banks or bank-padding
families. Historical `replica_lift.bank_stock` is inherited construction
metadata, not additional free stock in the final hosted template.

Source anchors: `research/gauge44-slots-t16/code/emit_allcold.py` lines80–104
route every helper to a distinct productive owner and declare this private
namespace; lines107–132 serialize the complete asynchronous continuation.
`code/invoice_allcold.py` lines101–117 charge a cleared coefficient work stream
for each scalar macro; lines270–300 charge the physical stock, including the
helper reserve. The local emitter
`research/complex-sw4-frame8/local/hpe_emit_local.py` lines53–120 emits scalar,
adapter, climb and centre events in namespaces L and C. Its audit checks
private centre and coefficient return. These exact source files are pinned,
read as evidence, and never executed by this checker.

## 2. A concrete helper-bank candidate fails

Consider assigning the224 output roles Z0…Z111,W0…W111 to the first224
helper owners of the actual B1 call. The receipt supplies all those physical
family IDs, and all480 helpers' ordered routes, entry and core-exit frame
bases, last local read/write indices and next explicit global use.

The allocation model is deliberately bounded: fixed source owners; arbitrary
independent productive contents; no evacuation, recomputation, destructive
retirement of the original output family, or changed representation. An
exclusive output may overwrite arbitrary old *owned* dirt. It may not erase
a still-productive family merely because its helper role has ended.

Every candidate is below14336, and every one has a later explicit use in the
actual continuation. Local completion preserves its productive content under
the paid invertible residual transformation; it does not turn that content
into disposable zero. Thus every candidate conflicts with an outstanding
productive obligation. In the bipartite allocation graph from224 output roles
to eligible candidate banks, the eligible neighborhood is empty: matching
size0, deficit224. Equivalently, contracting the protected productive elements
of the free-bank partition matroid leaves rank0 on this candidate set.

This is a capability/interference obstruction, not a universal information
lower bound. Changing a source/consumer contract, evacuating data, or proving
a different representation may change eligibility. The audit does not count
an already eliminated centre child as new stock or new savings. It is distinct
from PR3976's appended-output-dirt counterexample: the candidates here are
actual source-owned physical families with exact future uses.

## 3. The charged reserve supplies an address-level positive fixture

There is a concrete alternative inside the existing physical stock:

- Zp ↦14344+p and Wp ↦14456+p for0≤p<112;
- Fk ↦14568+k and Gk ↦14576+k for0≤k<8;
- current X/Y map to the first224 actual B1 owners (families0…223).

The output and scratch leases use240 of the480 reserved tableau fields,
leaving240 reserve fields, all8 centres and both coefficient fields separate.
The checker scans the entire frozen3813-event PR4015 word. Each target belongs
to an active owned role; every read follows initialization; borrowed inputs
are never written; physical output/scratch/input roles are disjoint; private
windows return zero before release; no private role spans a native event.
It records each output's first initialization and final word access. The
unchanged PR4015 theorem supplies the final F17 frame and exact output values;
this audit does not infer phase correctness from address disjointness.

Independently, the checker enumerates all482 explicit selected global events.
Local owners, idle child owners, bridges, terminal translations/exchanges and
final live-family correction do not address any of these480 reserve IDs.
For the two local provider words, every explicit register reference is L or C;
L is mapped through the selected owners and C through the centre interval.
Coefficient macro work uses the separately declared final two fields.
Therefore the proposed map has zero *explicit top-level* conflicts, including
through the actual next bridge. The bridge reads current pair0/pair1 X/Y; it
has no read of the newly decoded Z/W outputs.

This is a positive literal-address allocation fixture. It refutes the claim
that224 *additional addresses* are forced by the decoder's output count.
It does **not** establish that those240 reserve fields are exclusively
available inside every opaque native child, wrapper, map implementation or
six-tape compiler. The inherited general reserve has not been proved unused
by those interiors. Absence from a macro trace is not an exclusive lease.

## 4. A discriminating native-interior test

A complete substitution must provide an effect contract excluding the leased
240 families from every child/map/wrapper's internal work, or supply an
explicit disjoint scoped allocation. The positive fixture assumes that effect
contract. The checker also evaluates the conservative alternative in which a
native child may use the shared tableau reserve. The first native line occurs
after live output definitions; a legal-to-the-unrefined-arena clobber of a
retained output violates its lifetime. The negative control rejects it.

This is a precise obstruction to inferring a composed allocation from the
current macro contracts: two address-level refinements differ on protected
output effects while the displayed native target is unchanged. We do not
claim an existing native implementation actually performs that clobber, or
that it fits the selected finite allowance. The countermodel concerns missing
effect restrictions only. Proving a disjoint native workspace would discharge
this obstruction; assigning unknown workspace behavior zero would not.

The selected `gauge44` producer and the changed QQ16 producer also remain
separate premises. Carving addresses from the selected reserve does not prove
that selected F0/B1 produces PR4015's borrowed input state. That algebraic
binding is a different investigation.

## 5. Complete local stock/return bill and remaining costs

The unchanged PR4015 word costs224 H1 children,448 wrappers,352 maps,
416 CLEARs and26944 scalar primitives. The proposed address renaming adds no
scalar/native/map event, no productive family, and no physical family address
in the literal model. It uses224 exclusively owned output fields and16 private
scratch fields within an already charged480-field reserve. This is not a
physical compiler-space bound or a new complete invoice.

Outputs return in F17. Private scratch returns zero/identity by the fixed
source contract. If the outer private-stock contract requires eventual zero
return after the outputs have been legitimately consumed or discarded, add
224 full-cover CLEARs:448 real scalar primitives, zero native children and
zero maps. The module-plus-output-erasure scalar allowance is27392. No inverse
frame transform is needed to erase an owned output; zero may be initialized in
the identity frame. This erasure destroys the decoded demand, so it cannot be
moved before an actual last consumer. The next selected bridge is not that
consumer. A terminal consumer/retirement schedule is still required.

| Obligation | Status |
|---|---|
| Actual field IDs, helper future use, stock identity | Proved by authenticated data census |
| Helper-only fixed-owner allocation | Precisely obstructed (0 eligible of224) |
| Reserve address injection and explicit continuation survival | Constructive, checked on the complete literal words |
| Exclusive native/map/wrapper interiors | Unknown; explicit clobber test rejects its omission |
| Selected gauge44 → QQ16 input binding | Unknown here; independent obligation |
| Output last consumer and legal erasure | Unknown; conditional erasure bill448 is exact |
| Six-tape movement/control, finite precision, productive coverage, both supplier caps | Unknown; no κ claim |

The next task is a narrowly specified240-field exclusive lease or a smaller
proved native workspace footprint for these exact addresses, together with a
real output consumer. Stop before another decoder optimization or complete
supplier claim. PR3339's distinction between core, occurrence and consumer
costs applies: an unused local address is not productive-family coverage.
59 changes: 59 additions & 0 deletions research/qq16-output-stock/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,59 @@
# Audit actual QQ16 output stock

The first selected B1 caller does not release its productive helper fields as
scratch. A concrete224-bank candidate drawn from its helper owners has **zero
eligible banks** under a fixed-owner, no-evacuation allocation model. All480
helper lifetimes, frames and next explicit uses are recorded.

There is nevertheless an existing charged **480-field private tableau
reserve**. This package constructs a precise224-output plus16-scratch address
map within it, verifies the complete PR4015 word's lifetimes and aliases, and
checks survival through all482 explicit selected continuation events. It adds
no physical family address in that literal model.

The missing implication is now explicit: **exclusive native/map/wrapper
interiors must preserve the240 leased fields**. Macro-trace absence alone does
not establish that lease; a clobber control rejects the inference. Selected
gauge44-to-QQ16 producer binding and an actual output consumer are separate
obligations. Conditional eventual output erasure costs224 CLEARs /448 scalar
primitives. No complete physical supplier or κ improvement is claimed.

See [PROOF.md](PROOF.md) for the models, source anchors, positive map and
obstructions. RESULT.json contains the exact field/route/lifetime ledger and
all controls. PR4015 is pinned at0fff81d8997cef646538bcd4f60d7d455cf93b8e and
retained only as data. No pending or canonical predecessor executable runs.

## Reproduce

```sh
python3 -I -B research/qq16-output-stock/verify.py
python3 -I -O -B research/qq16-output-stock/verify.py
python3 -B scripts/research_package.py --package research/qq16-output-stock
```

Standard-library Python3.12. Normal replay is read-only; explicit regeneration
refuses overwrite. The verifier authenticates SHA256 and Git blobs before
reading inputs and validates the complete typed receipt. Native/phase algebra
is inherited from PR4015 and is not independently reimplemented here.

## Scope and credit

User-assigned actual-output-stock follow-up, priority6, canonical main20ecf289.
This closes an actual caller allocation test rather than repeating the local
inverse, PR2651's early-record reuse, or PR3976's appended-dirt witness. Overlap
checks are partial; no novelty guarantee. The selected frontier is unchanged.

Credit PR4015 and PR4007 for the zero-extra-temporary words; PR3985 and the
retained ywz65/Arham Tahir/Ali Romman QQ16 lineage; commdata2338/PR2651;
Nick Petalas/PR3339; Jeremy Kothe, Rohan Arun, Hao Yan and the retained selected
hosting/invoice contributors. Chris R / Clear1nsight with substantial OpenAI
Codex assistance, Apache-2.0. No human or formal-kernel review is claimed.

## Open next questions

- Bind a disjoint native/map/wrapper workspace contract to the named240-field
lease; the explicit clobber control is the first acceptance test.
- Supply QQ16 input semantics at the actual source boundary independently of
storage availability; selected gauge44 does not become QQ16 by renaming.
- Pin an actual last consumer, then charge output return, movement/control,
precision and productive coverage before evaluating either supplier cap.
Loading