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
2 changes: 2 additions & 0 deletions research/p104-large-step-audit/.gitattributes
Original file line number Diff line number Diff line change
@@ -0,0 +1,2 @@
retained/** -text -whitespace
bit/inputs/** -text -whitespace
17 changes: 17 additions & 0 deletions research/p104-large-step-audit/NOTICE
Original file line number Diff line number Diff line change
@@ -0,0 +1,17 @@
Hao Yan (@FanPUYen), with substantial OpenAI Codex assistance. Apache-2.0.

This package adds source-bound negative checks and proofs, not a new kappa.
DreamingOfClouds' PR7613 supplies the retained scalar word and invoice;
Clear1nsight's PR7612 supplies its base and contact-rechart pattern. The lineage
retains FanPUYen, SYLforge, ToikY, Rohan Arun and all authors named in
retained/7613/NOTICE. No predecessor source bytes are changed.

retained/interval_gate.py is the engine from the role902 direct-lift lineage,
reused in FanPUYen's PR7604. retained/independent_interval.py is the unchanged
rational engine retained by the selected paired-query-only BIT package.
The original BIT, native/ordinary/strictification sources retain Ali Romman,
Cinna Davis, zxv, DreamingOfClouds, CubeJerry and all predecessor credit.
See retained/*NOTICE, bit/NOTICE and bit/inputs/*NOTICE.

Separate OpenAI Codex agents performed implementation-independent checking;
no external human or formal-kernel verification is claimed.
97 changes: 97 additions & 0 deletions research/p104-large-step-audit/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,97 @@
# Necessary contact-frame cost bound

This is a necessary-condition rejection at two specified rational C values for
the exact retained #7613 word/invoice. It supplies neither a new κ nor a universal
upper bound for multiplication.

## Model

Keep the scalar instructions (order, operands, coefficients), 400 main fields
and their endpoints, 1101 helpers, owned lifetimes/chart trajectories, and the
complete 17-call signed caller. Each public field follows an ascending nested
sequence of subspaces of F2^9. ADD/COPY require coframed operands; owned contacts
retain that owned frame. Arbitrary bases, new subspaces and additional public
moves between scalar contacts are permitted optimistically.

The first variant fixes all public endpoints and targets. The second frees each
helper's input/output subspace and credits its one target with rank
`dim(output)-dim(input)`. New fields, split targets, scalar rewrites, other callers,
nonmonotone paths and changed owned operators are outside this model.

Equal spans need not have equal ordered physical bases or phases. We do not
identify their physical operators: we drop extra basis/phase costs and legality
restrictions only when taking a lower bound on paid cost.

## Necessary corridors

Start at each fixed public input space (or {0} for a free helper). At a public
contact sum its predecessor lower spaces; at an owned contact use its pinned
space. This gives L. Starting at fixed outputs (or F2^9 for free helpers),
intersect successor upper spaces in reverse order and apply owned pins, giving U.
Induction shows every permitted contact space S satisfies L ⊆ S ⊆ U. SCALE is a
single-field contact. COPY still requires coframed operands and the declared
monotone frame trajectory, despite overwriting one scalar row.

The independent checker constructs a contact DAG instead of scanning instructions,
then applies topological forward/backward anchors. It checks source MOVE directions
and chart/lifetime states; both implementations obtain identical corridor widths.
Contradictory anchors and forced descending dimensions have rejecting controls.

## Dimension relaxation and coupled credit

Set H_C(r)=r^(1-C) for r>0 and H_C(0)=0. Since 0<C<1, subadditivity implies that
multiple ascending moves between contacts cost at least H_C of their total
dimension gap. Replacing them by one gap tariff is therefore optimistic.

For each field independently choose nondecreasing dimensions within [dim L,dim U].
This drops actual subspace containment and equality of the two chosen operand
frames at a common contact. Every legal program induces these independent paths,
whereas the converse is unnecessary. The sum of minimum path costs is consequently
a lower bound on native public paid cost.

For a free helper retain its start dimension in the DP state and minimize

sum H_C(d_next-d_previous) - H_C(d_end-d_start).

Target subtraction is inside the same path optimization. Optimizing target and
cost separately or leaving old targets unchanged would not implement this bound.
Main fields have no subtraction. Zero-rank helper targets are allowed with zero
credit, making the relaxation more permissive.

The independent checker supplies concrete nonzero modular evaluations for every
public destination row except the sole exactly-zero interval. Both contacts
adjacent to that interval are pinned to frame115 even with free helpers, so the
gap already costs H_C(0)=0. Granting free known-zero relabeling there changes no
bound. Nonmonotone excursions leave the stated ascending-trajectory model.

## Complete invoice and arithmetic direction

Subtract 17 copies of the source public MOVE histogram from complete P. Its
nonnegative remainder P_fixed includes owned and caller costs. With fixed endpoints
keep T. With free helpers subtract 17 copies of their old target histogram from T,
giving T_fixed. The independent checker also separates owned and caller costs and
checks the remaining caller target histogram.

If K is the sum of minimum native costs (cost minus helper credit in the free
variant), every permitted candidate satisfies

F(C) = target moment - paid moment
<= moment(T_fixed,C) - moment(P_fixed,C) - 17 K.

The primary engine brackets H_C with directed integer arithmetic at 80 decimal
places, a range-reduced atanh logarithm and alternating exponential series. The
minimum of finite costs in [a_i,b_i] is enclosed by [min a_i,min b_i]. Target
subtraction reverses its interval endpoints. The final upper bound uses the
LOWER paid-cost bound.

The independent implementation uses bitset subspaces, a DAG, reverse endpoint-
conditioned DP and a separate positive-Taylor rational engine rounded outward to
40 decimal places. Its native-net and final-margin intervals must overlap the
primary intervals. Original source paths provide an available-path bound; tiny
ordinary/coupled paths are exhaustively enumerated. Exact fractions are retained.
All four upper margins are negative; rounded values appear in README.

No monotonicity of the optimized margin in C is invoked. The conclusion concerns
the named rational points, not every larger C. Changed scalar dependencies, owned
trajectories, callers, field counts and target interfaces need a new audit. The
separate BIT kernel and interleaving proof is in bit/PROOF.md.
108 changes: 108 additions & 0 deletions research/p104-large-step-audit/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,108 @@
# A +0.001 target audit: contact frames and BIT cache promotion

**No κ gain was obtained.** This package records two source-bound obstructions
encountered while attempting an absolute +0.001 improvement. It changes no
selected result, source word, invoice, shared checker or admission claim.

The donor's preceding result was #7604, κ = .0036502223357, so the first exact
test is C = .0046502223357. The latest open #7613 reported raw crossing is
approximately .0036530568656869; its published upper bracket plus .001 gives
the second explicitly tested point C = .0046530568657. These are tests of
the retained signed-moment criterion at named points, not an impossibility
claim for every larger C or for arbitrary multiplication algorithms.

## Results

For the exact #7613 scalar word, fix its opcode/operand/coefficient order,
owned chart chronology, caller, main endpoints and number of public helpers.
Allow arbitrary contact subspaces of F2^9 and monotone public trajectories,
including new moves inside old stays. A necessary subspace corridor followed
by independent per-field dimension optimization gives these **optimistic upper
bounds** on full target-minus-paid moment:

| Allowed variation | C = .0046502223357 | C = .0046530568657 |
| --- | ---: | ---: |
| Fixed public endpoints and targets | -362.9625468636 | -364.0413334486 |
| Free helper input/output frames, coupled target credit | -306.0163311064 | -307.0605913342 |

All four upper endpoints are strictly negative. The second row includes both
helper endpoint changes together, with the corresponding changed target credit.
Actual subspace nesting, operand agreement at contacts, phase costs and full
physical legality are relaxed in the direction favorable to a candidate.
A positive result would not construct a supplier. See [PROOF.md](PROOF.md),
[frames.json](frames.json) and the independent checker `independent_frames.py`.

A distinct changed-ownership attempt promoted the nine descending rank-seven
BIT caches 770–778 to independent inputs. Nine explicit kernel witnesses show
that **no nonzero joint combination of those inputs is recoverable at chart0**,
even allowing arbitrary interleaved chart0 linear repair under the fixed original
chart trace. This fails the output contract, so its hypothetical target credit
cannot be booked. See [bit/PROOF.md](bit/PROOF.md) and
[bit/EXPECTED.json](bit/EXPECTED.json). This is an exact linear-algebra certificate,
not a global limit on changed-ownership BIT constructions.

## Reproduction and limits

Python 3.12, standard library only, no network or foreign Git objects:

```sh
python3.12 -I -B research/p104-large-step-audit/verify.py
python3.12 -I -B research/p104-large-step-audit/frames.py
python3.12 -I -B research/p104-large-step-audit/independent_frames.py
python3.12 -I -B research/p104-large-step-audit/bit/audit.py
python3.12 -B scripts/research_package.py --package research/p104-large-step-audit
```

The frame implementations use different data structures, forward/reverse
min-plus recurrences and different directed rational interval engines.
Controls cover exhaustive short paths with coupled endpoint credit, infeasible
descents, the sole zero public-state gap and source corruption. The BIT checker
propagates observations backwards, directly checks nine kernel vectors, and
replays their nine-bit values forwards. Its joint-stock control catches the
invalid inference that absence of individual basis inputs excludes every combination.

This is not a fresh full native, physical, caller, tape, precision, compiler,
outer or all-size replay. The #7613 word/invoice and selected BIT normalized
semantics are pinned premises; #7613's inherited CLEAR-content assumption is
neither accepted nor discharged here. The bound rejects the stated model even
if that candidate is provisionally granted its stated semantics. Separate Codex
agents supplied implementation-independent checks; no external human or
formal-kernel verification is claimed.

## Scope and overlap

Canonical main was refreshed at e54523517d0f03e58352e647c403a5af3e525a57.
The selected record is unchanged; open conditional claims are not admissions.
The revision-8 scalar-to-moment card is refined to this explicit donor target;
the secondary BIT attempt changes ownership and is labelled separately in TASK.
Live PRs and discussions were checked before exploration and before submission.

- #7574 optimizes retained old stays/boundaries; #7582 searches a finite contact
palette. Our relaxation permits arbitrary contact subspaces, extra moves
inside stays and jointly free helper endpoints.
- #7577 bounds funding with fixed original helper/cache histories. #7122/#7124
already discuss endpoint recharts, so the stronger row includes endpoint freedom.
- #7304 exhausts fixed-bank BIT target splits; preliminary duplicate split searches
were stopped. The retained result changes nine owned inputs, supplies joint
kernel witnesses, and proves an interleaving obstruction.
- #6726's other promotion pilots do not use this actual selected BIT word.
#7162/#7163 and #7607/#7609 retain their separate caller/address-map questions.

## Open next questions

1. Can a changed scalar interaction or owned trajectory produce enough complete
moment gain? First change one dependency creating a tight corridor, replay
exact rows and charge all correction costs. The fixed program is the blocker;
this bound does not apply automatically after such a change.
2. Can paid nonzero-frame repair distinguish a promoted BIT input from its explicit
old-input kernel shift? Start with owner772's two-term witness and construct
one actual address map with a complete endpoint contract. No such map or
profitable invoice is supplied here.

## Sources and credit

Hao Yan (@FanPUYen), with substantial OpenAI Codex assistance; Apache-2.0.
DreamingOfClouds supplies #7613, Clear1nsight #7612 and the contact move,
FanPUYen #7604, SYLforge #7603, ToikY #7546 and Rohan Arun #7113. All predecessor
credit is retained. See NOTICE, SOURCE-VERSIONS.json, bit/SOURCES.json and the
unchanged source notices. Remote commits are provenance, not replay dependencies.
60 changes: 60 additions & 0 deletions research/p104-large-step-audit/SOURCE-VERSIONS.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,60 @@
{
"schema_version": 1,
"scope": "Immutable self-contained exports; upstream commits are attribution, not replay dependencies.",
"files": [
{
"path": "retained/7613/word.json.gz",
"commit": "bde25de30596406f5bcbc41cbce8fe37462d97c2",
"source_path": "research/p104-contact902-1091-on7612/data/word.json.gz",
"sha256": "21725d0c4ab84bac4d53701624c44701122a977b445d8facfb2ee7c7eba1afce"
},
{
"path": "retained/7613/invoice-new.json",
"commit": "bde25de30596406f5bcbc41cbce8fe37462d97c2",
"source_path": "research/p104-contact902-1091-on7612/data/invoice-new.json",
"sha256": "d362f075d8855cfed1c6c8a2bd1d50af727900d1fff810ebf57a8afb8794e721"
},
{
"path": "retained/7613/NOTICE",
"commit": "bde25de30596406f5bcbc41cbce8fe37462d97c2",
"source_path": "research/p104-contact902-1091-on7612/NOTICE",
"sha256": "4ddad1cfdec129d22663368fd99c1ee67c0e93b42dbf19a771d80ad7d8df7d30"
},
{
"path": "retained/7613/README.md",
"commit": "bde25de30596406f5bcbc41cbce8fe37462d97c2",
"source_path": "research/p104-contact902-1091-on7612/README.md",
"sha256": "fb9cdb5aa3368acb498d726e08a2afb9d3e66201c56b4ec29c32370d46349e23"
},
{
"path": "retained/interval_gate.py",
"commit": "99438c43df7aac1e5dbcc43c1578fcd886c119ec",
"source_path": "research/p104-role902-direct-lift/engines/interval_gate.py",
"sha256": "60a743d13f668ef86db4b8215b4d5d04430a671ba6f1fe06d9c1c4c3dfe86da5"
},
{
"path": "retained/role315-NOTICE",
"commit": "ba2638c5a6e2b9a1c0c559a41811e5d41a591f41",
"source_path": "research/p104-role315-direct-lift/NOTICE",
"sha256": "e984c9868bc47123f7f9f03fa7bae9c8b3d9fb5cad38e9f88a57c660a100d313"
},
{
"path": "retained/independent_interval.py",
"commit": "e54523517d0f03e58352e647c403a5af3e525a57",
"source_path": "research/paired-query-only-bit/retained/inputs/independent_interval.py",
"sha256": "9293f567aa4afd9b4e618f1cddd3cf66eea51c5e194e3c578e7a8f4db6e04a57"
},
{
"path": "retained/paired-bit-NOTICE",
"commit": "e54523517d0f03e58352e647c403a5af3e525a57",
"source_path": "research/paired-query-only-bit/NOTICE",
"sha256": "8c73eaf46327f01b4cfc18640b5e5f552648fc5e4427cbccd42af5e7ecd3b732"
},
{
"path": "retained/paired-retained-NOTICE",
"commit": "e54523517d0f03e58352e647c403a5af3e525a57",
"source_path": "research/paired-query-only-bit/retained/NOTICE",
"sha256": "620e4ec6f537fa1c071f019ac92fa0e4eec8ab487c811e84153292b981de0424"
}
]
}
18 changes: 18 additions & 0 deletions research/p104-large-step-audit/SOURCES.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,18 @@
{
"schema_version": 1,
"base_commit": "e54523517d0f03e58352e647c403a5af3e525a57",
"files": [
{
"path": "certificates/selected-result.json",
"usage": "provenance",
"git_blob": "75d77b328ea02ded535e7f23aee5b1e55d0bb1a8",
"sha256": "154c3d30455d778fd3343b6b4851f0fa4665f0f7625b0adb3675e2bafbbf6bf1"
},
{
"path": "docs/research/agent-task-catalog.json",
"usage": "provenance",
"git_blob": "3404fbdd5e7529a0674eef26114b5071e4586cec",
"sha256": "ead6cdc9d0af6d58e8a49744d8ec8c7c15e13d2ba93c276c434bcf9b51b93a81"
}
]
}
Loading
Loading