Skip to content

Prove two-helper shears gain nothing from target feedback at any target count (no new κ) - #6034

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
petalas:research/w13-1022ai-two-helper-feedback
Oct 28, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
petalas:research/w13-1022ai-two-helper-feedback

Conversation

@petalas

@petalas petalas commented Oct 28, 2026 •

Copy link
Copy Markdown
Contributor

Result

Claim type: scoped proof for every target count, with finite corroboration. No new supplier, κ or selected-frontier change.

In the zero-frame dirty-shear model of PR #1306 and PR #2567, take two helpers and T targets with unit-cost single-source XORs and no extra registers. Let a, b, c count the rows (1,0), (0,1), (1,1) of K. For every T ≥ 1 and every K, the restricted and unrestricted minima are equal:

d_P(K) = d_full(K) = a + b + c + min(c, 2).

So target-to-helper feedback never helps for a two-helper shear. This answers the 2×7 part of Open next question 1 in rtyu193-rgb's PR #2567. Through that package's duality it also covers T×2, including 7×2. The 3×6 and 4×5 shapes stay open.

The upper bound is an explicit restricted word: direct reads, plus one shared h0 ^= h1 … h0 ^= h1 block when c ≥ 2. The lower bound counts destinations and allows feedback:

  1. every nonzero row needs its target as a destination;
  2. if no helper is ever a destination, a target modified once can only receive a pure row;
  3. a helper modified once cannot end unchanged.

The full proof is in research/two-helper-zero-frame-feedback/README.md.

Evidence and reproduction

verify.py uses only the standard library and imports no predecessor code. It checks:

  • explicit restricted witnesses for every K with T = 1..8 (87,380 matrices), replayed exactly; the gates are linear, so this covers all dirty inputs;
  • exact unrestricted and restricted BFS for T = 1..3 (84 matrices) against the formula;
  • the formula against PR Finite certificate: no target-feedback saving for any shape with at most eight registers (no new kappa) #2567's pinned 2×4, 2×5 and 2×6 certified minima (5,376 matrices). This calibrates the formula; it is not a replay of that search;
  • four rejection controls, including the wrong formula wt(K).
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -B research/two-helper-zero-frame-feedback/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -O -B research/two-helper-zero-frame-feedback/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/two-helper-zero-frame-feedback
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/check_research_commit.py --package research/two-helper-zero-frame-feedback --check-source-history

All four commands passed locally on CPython 3.12. Each direct run took about 0.45 s and printed PASS_TWO_HELPER_ZERO_FRAME_FEEDBACK witnesses=87380 bfs=84 pr2567=5376 controls=4 feedback_gains=0 no_new_kappa.

Limitations

Open next questions

  1. Does feedback help for 3×6, the other nine-register shape besides 4×5? Step 2 of the proof assumes two helpers.
  2. Is there a row-type closed form for d_P at every H?
  3. Does destination counting extend to twisted two-helper exits, beyond PR Negative result: twisted helper exits have no target-to-helper feedback gain at any fixed 3+3 or five-register exit, plus an every-word contamination-charge lemma (no κ claim) #4071's fixed small shapes?

Related open work

CI notes

Head d27351424186a70b06cf3dae0eb87e425acd5f2d: both hosted checks passed, plan and replay (research/two-helper-zero-frame-feedback), in run 40608381317.

Both minima equal a + b + c + min(c, 2) for every target count,
answering the 2x7 part of the PR2567 nine-register question.
@Mister-Meeseeks
Mister-Meeseeks merged commit 2b35e6c into CrocSwap:main Oct 28, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants