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
126 changes: 126 additions & 0 deletions research/controller-stage-ceiling/PROOF.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,126 @@
# Complete stage-family ceiling and a native repair target

Prepared by Cinna Davis with substantial OpenAI Codex assistance. Apache-2.0.

## Family and scope

Fix one of the three source-bound local invoices: selected PR4470, PR4610,
or PR4626. There are v=112 ports and transverse rank b=8. For each
1 <= p <= 64, take any split a in {0,...,p-1} of the minimal-cut paired
path controller, and **any** subset S of its 2p+1 native stage positions.
Stages outside S become fully paid query shears, lose all helper credit,
and lose their stage multiplier in both main targets and idle jobs.
This is the specialization operation of PR4440, not deletion of a bill.

The family is deliberately an invoice relaxation: all these substitutions
are priced even if a particular new physical contact has not been admitted.
Rejecting this stronger relaxation is safe; positive scores alone would not
admit a word. No other idle masks, words, native suppliers, typed outputs,
or p>64 are excluded. The selected record is unchanged.

## Complete surplus

Let n=|S|, t=1-C, and let H_t and P_t be the local helper and paid moments.
Write V=112 and B=8. The complete surplus is

F(C) = n(H_t-P_t+V) - V(2p+1) + 2V(1+Bn)^t
+ sum_arm 2V[(1+B(k+e))^t - (Bk)^t - e B^t].

There are a left arm of length a and a right arm of length p-1-a. At
an arm step j, k is the number of selected bits in the first 2j positions
of that arm and e is the next, even-position bit. The right arm is read
in reverse stage order. Rank-zero idle terms mean zero cost, including
in the limit t -> 0. The central odd-position bit contributes only to n.

The first term includes all n independent helper inventories and all n
native paid inventories. The V correction replaces the otherwise paid
query at each native stage. The second term charges every query position.
The third term is the two copies of the long main target. Each arm term
includes its shorter main target, initial/terminal idle, and the separate
middle idle of rank B when e=1. Thus no middle idle is silently erased
when its stage remains native. `literal_control` independently counts every
LOCAL, QUERY_SHEAR, IDLE and diagonal target of the retained p16 word.

## Dynamic program, with outward rounding

An arm's score depends on a prefix only through its selected-bit count.
For a nonempty arm its first two bits have no separate charge and therefore
only their count 0,1,2 matters. An empty arm has one bit. At each subsequent
step choose e, add its displayed complete arm contribution, then choose the
following odd bit unless this is the last step. Keep the largest score
for each count. This is exhaustive induction on the bit prefix, not a
local hill climb. It is the recurrence of the credited historical floating
DP, now evaluated with exact, outward integer bounds.

The retained rational interval engine encloses each r^t. Scale by 2^64,
round positive occurrences up and negative occurrences down. Every DP
score is consequently an upper bound on its real surplus. Maxima preserve
upper bounds. Combine both arms, every split and the middle bit, then
maximize over every p=1,...,64. The three native profiles share that
combinatorial envelope. Their local replacement terms are added separately.

At C=0.003022, every final upper bound is strictly negative. `RESULT.json`
records the exact rational upper bounds, integer scale and witness parameters.
The covered count is sum(p*2^(2p+1), p=1,...,64) per native invoice. It
counts split/support pairs; it is not a claim that all invoices are distinct.

Acceptance also exhausts every arm bit pattern through length four and
all complete stage supports/splits through p=3; checks the actual p16
invoice for each local substitution; and recomputes a witness moment with
the independent full-invoice path of the rational engine. Omitting the
real middle-idle charge or duplicating independent helper credit changes
the rejection into an apparent positive result. These are rejected models.

## Why one test excludes every larger saving in this family

For each fixed support, divide the surplus by 8^t. Its normalized local
replacement is

L(C) = (H_t - P_t + V)/8^t.

All paid ranks are at most 8. For t in (0,1], its derivative with respect
to C has upper bound

L'(C) <= -((P_1-V)/8)*log(8)
+ sum_(r<8) H_r log(8/r) < 0.

Indeed, retain only the negative rank-one paid contribution; its magnitude
is at least (P_1-V)log(8)/8. Bound each possible positive helper contribution
by H_r log(8/r), and drop the other negative contributions. The final strict
inequality is enclosed rationally for each frozen invoice.

For an arm step with k>=1, the derivative of

(k+e+1/8)^t - k^t - e

is negative, since x^t log(x) is increasing for x>=1. If k=0,e=1,
its derivative is again negative. The only positive case k=e=0 contributes
at most (1/8)^t log(8) before the multiplicity 2V. There are at most p-1
such arm steps. The normalized negative query term contributes
-V(2p+1)(1/8)^t log(8), more than cancelling all of them. If n>=1,
the normalized long main target has negative derivative. If n=0, its
positive derivative leaves the total derivative exactly
-V(1/8)^t log(8). Thus F(C)/8^(1-C) is strictly decreasing throughout
0<=C<1 for every support in the family.

The negative bound at .003022 therefore excludes **all C>=.003022,
C<1**, not merely that one numerical point. This is a conditional invoice
ceiling, not a universal limit on integer multiplication.

## Changed model: the H repair budget

The old H112 reconstruction changes a local rank-eight correction bill
8 H8 to 8 H1, saving 56 units of rank per full call. Applying that
56-unit change to PR4610's stronger native invoice is an **optimistic
unrealized combination**. It is not a source-compatible H word: changed
warp contacts, cross-port response, head/hole normalizers, absolute phases
and arbitrary-input helper restoration all still need proof.

For that explicitly hypothetical invoice, the same complete DP gives
exact upper repair budgets at .004, .006 and .01. A full-invoice lower
bound at each saved witness confirms the scale of the positive budgets.
A rank-one repair request costs exactly one unit at every C; a rank-r
request costs r^(1-C). Storage, scalar work and all other requests also
need their actual ledger. These numbers are spending limits, not free
allowances or proposed K values. In particular the .01 screen fails even
before a repair in this bounded family. No conclusion for p>64 follows.
63 changes: 63 additions & 0 deletions research/controller-stage-ceiling/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,63 @@
# Exact ceiling for native/thin path-controller changes

**No higher K or new supplier is claimed.** For the source-pinned PR4470,
PR4610 and PR4626 invoices, every minimal-cut path controller with 1..64
pairs, every split, and every native/thin stage subset fails at
C=0.003022. A normalized-moment monotonicity proof excludes all larger
C<1 in this family. This turns the earlier floating stage search into an
exact certificate and identifies a model change required for a substantial
increase. The selected conditional K remains 0.00302015 at the base commit.

[PROOF.md](PROOF.md) derives the complete invoice, the exhaustive max-plus
recurrence, outward integer bounds and the monotone rank-eight separator.
[RESULT.json](RESULT.json) contains the exact receipt. This does not exclude
new words, native interfaces, idle masks, coded corrections, or p>64.

The changed-model calculation also prices an optimistic combination of
PR4610 with H112's 56-rank-unit local saving. Its repair budgets at .004,
.006 and .01 are included, with the unproved native/caller compatibility
explicitly retained. A budget or positive invoice is not an implementation.

| Candidate C | Maximum optimistic rank-one repair budget |
| --- | ---: |
| .004 | 2641.1932795534 (p=64) |
| .006 | 342.1085144324 (p=37) |
| .01 | Already negative: -93.1789150428 |

These are bounded-family upper limits; source compatibility is unproved.

Reproduce from the repository root:

```sh
python3.12 -B scripts/research_package.py --package research/controller-stage-ceiling
```

The verifier uses the standard library and the pinned rational interval
engine. It does not require a solver. Controls compare small complete
censuses and literal p16 event bills, and show why dropping middle idles
or duplicating independent helper credit would give false apparent gains.
The external PR inputs are exact unmodified Git blobs, not newly admitted
native words; their provenance and heads are in SOURCES.json.

## Open next questions

1. Construct a physical H caller with a complete correction invoice below
the recorded repair budget. Start with one actual head/hole boundary and
check all columns, source/complement ownership and absolute phases. The
old local H invoice and its scalar involution do not close that caller.
2. Change the native word, interface or controller topology. A further
native/thin subset scan inside the certified family cannot reach .004.
A new candidate must provide an exact endpoint and a complete paid bill.

## Attribution and status

Prepared by Cinna Davis with substantial OpenAI Codex assistance, Apache-2.0.
The branch recurrence is credited to the pre-existing
`research/kappa-hybrid-controller/search/direct_stage_dp.py` authors and
Cinna Davis's PR4440/PR4470 controller lineage. Cao Yifan's PR4610 supplies
the changed warp invoice; hexapode's PR4626 supplies the helper621 invoice.
Retain zxv's exact interval/native/strictification contributions and Jacob
Alan Sussman's mixed-job/compiler work. The new contribution is the exact
integer bound, the monotonicity proof and the current source-bound repair
budget. It does not assert independent human review, a formal proof-kernel
check, or a new multiplication theorem.
82 changes: 82 additions & 0 deletions research/controller-stage-ceiling/RESULT.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,82 @@
{
"all_native_stage_subsets": true,
"all_splits": true,
"ceiling": "1511/500000",
"controls": [
"all_5_small_arm_censuses",
"all_complete_patterns_p1_to_p3",
"all_3_literal_p16_invoices",
"missing_middle_idle_changes_sign",
"duplicate_helper_credit_changes_sign"
],
"fixed_point_scale": "18446744073709551616",
"new_kappa": null,
"optimistic_H_repair_budgets": {
"1/100": {
"best_surplus_upper": "-53713987464422013341/576460752303423488",
"status": "OPTIMISTIC_INVOICE_ONLY_NO_H_NATIVE_ADMISSION",
"witness_parameters": [
1,
0,
6
],
"witness_surplus_lower": "-292446764658482249839017434957133446527108325308117437477329/3138550867693340381917894711603833208051177722232017256448"
},
"1/250": {
"best_surplus_upper": "24360708238561784637317/9223372036854775808",
"status": "OPTIMISTIC_INVOICE_ONLY_NO_H_NATIVE_ADMISSION",
"witness_parameters": [
64,
31,
680564733840226340207702041255296892927
],
"witness_surplus_lower": "2072379864822052426775990683944584074709758170880344027228125/784637716923335095479473677900958302012794430558004314112"
},
"3/500": {
"best_surplus_upper": "3155394105585959091153/9223372036854775808",
"status": "OPTIMISTIC_INVOICE_ONLY_NO_H_NATIVE_ADMISSION",
"witness_parameters": [
37,
18,
31482443219130968091307
],
"witness_surplus_lower": "2147449949634335534650179046572602361862313365888596319687777/6277101735386680763835789423207666416102355444464034512896"
}
},
"pair_counts": [
1,
64
],
"patterns_per_profile": "57772384072799330241337377795082425233864",
"profiles": {
"PR4470": {
"separator_derivative_lower": "834868369783004008952768229585313202122157833070803325402459/1569275433846670190958947355801916604025588861116008628224",
"surplus_upper": "-4184234018831163103/9223372036854775808",
"witness_parameters": [
16,
7,
7158278827
]
},
"PR4610": {
"separator_derivative_lower": "410091947704675574684453227077902260907026584522260683631425/784637716923335095479473677900958302012794430558004314112",
"surplus_upper": "-1014518977939581775/9223372036854775808",
"witness_parameters": [
16,
7,
7158278827
]
},
"PR4626": {
"separator_derivative_lower": "1640764499882385906456930065487073635005120770603694469307877/3138550867693340381917894711603833208051177722232017256448",
"surplus_upper": "-977466852836146591/9223372036854775808",
"witness_parameters": [
16,
7,
7158278827
]
}
},
"scope": "Frozen native invoices and minimal-cut path family, p<=64. No global operator/controller ceiling or new supplier.",
"status": "PASS_EXACT_STAGE_FAMILY_CEILING"
}
70 changes: 70 additions & 0 deletions research/controller-stage-ceiling/SOURCES.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,70 @@
{
"schema_version": 1,
"base_commit": "b58ad9ef4d4e861b0eaa5dba41742bdc072dca51",
"files": [
{
"path": "certificates/selected-result.json",
"usage": "provenance",
"git_blob": "c35482cd9035843b15a634cc8276a7e1e40e8395",
"sha256": "6fe0a099b89aa2aed796a13c3a21dafc63f3388ebf397464e44e0451893b9375"
},
{
"path": "docs/research/agent-task-catalog.json",
"usage": "provenance",
"git_blob": "f126dd2e2dad2b669e75d1d0d0bd05c600725602",
"sha256": "f7a48bfa2659b88378282a8eee7d9fe52cdf505594124c4d1686b58510fffd72"
},
{
"path": "research/higher-moment-strictification/verify.py",
"usage": "input",
"git_blob": "4b24a5096154736ce060474eb8c72dc64c303699",
"sha256": "0ce7e7137422d5ffeda22085c81fa88471bbe963100fdad62e659f1e3bf89326"
},
{
"path": "research/joint-helper-lattice/NATIVE.json.gz",
"usage": "input",
"git_blob": "0910f74138decf498e7fc445cc2daff4390169d6",
"sha256": "74582cce642c80f0235be7fcbce3716174666326ac6a1d46d02a62a75d7be4c6"
},
{
"path": "research/kappa-hybrid-controller/search/direct_stage_dp.py",
"usage": "input",
"git_blob": "0717a3a577fd7a4e3eea9901bb266955bdb908de",
"sha256": "4a457525a8c041d3e0fd2671634627d831e3c0dc2970219b11b3b2692ba32125"
},
{
"path": "research/kappa-rebuild/bit/complex_controller_factor.py",
"usage": "input",
"git_blob": "cb0cff7c8be5f38f2ee52201d9d0907f9dfd7947",
"sha256": "1de9aaf918e6a4f2f5448754c1f3c3da3b85e6257849ebd097af09859c4e5497"
},
{
"path": "research/query-thinned-controllers/controller.py",
"usage": "input",
"git_blob": "d4557e08cfb9fe57cc1ae4024705a72dc95f3583",
"sha256": "8b462d660487b1be4940357c4c7522967b1bfaede3c02b1c66109b85acae0fd2"
},
{
"path": "research/controller-stage-ceiling/inputs/pr4610-native.json",
"usage": "input",
"git_blob": "88fa05cb3d0b7227f85ba0c42eb427184949447d",
"sha256": "c6f16840b3f7bed8686ad494b479df599b18b420c239dc6167c214130b78eb64",
"pin": "head:6ba2f0ad1809172be7ebf04ad750592f14475dab:research/controller-stage-ceiling/inputs/pr4610-native.json",
"original_source_pin": "head:1eb8da6cfb35512c0cc2533413e5ca42f48e13c6:research/warp-interface-20261022/NATIVE-VALIDATION.json"
},
{
"path": "research/controller-stage-ceiling/inputs/pr4626-arithmetic.json",
"usage": "input",
"git_blob": "bec84f7dc990a0a892fd77fd74665d49df66d4de",
"sha256": "5787bae88594f8d05e1a7bbecdc98c8635523c817093b717af621ac893277d28",
"pin": "head:6ba2f0ad1809172be7ebf04ad750592f14475dab:research/controller-stage-ceiling/inputs/pr4626-arithmetic.json",
"original_source_pin": "head:e4413c43eeaacdd20be123fdd7d7add53427f028:research/helper621-warp-combination/ARITHMETIC.json"
},
{
"path": "research/joint-helper-lattice/CONTROLLER.json",
"usage": "input",
"git_blob": "49a2f0675819f9deeea27ed77be2c2b6ff2e87e5",
"sha256": "1f5d6aa267e4e66048818714b5d3b5eae17a58abe91d15deeaa34574f7391a5f"
}
]
}
Loading