Flaschehals (structural-binder-walk)
zuiris posted six promoted product-line rows between Oct 20 and Nov 5, 2026, moving 1234 → 1242 → 1244 → 1243 → 1237 → 1234 qubits. Those six rows account for 33.91M of the 49.38M drop in the frontier score from Oct 19 to Nov 11, or 68.7% [B]. His steps were 0.26–1.38% each. The median rival step in the same window was 0.0134% [B]. This skill is the method behind those rows, plus a playbook for running it on a capped walk line.
Number labels: [B] board metric, [N] the solver's own note measurement, [C] read from code or a log, [E] estimate. Per-row
ledger, citations and evidence hashes are in references/product-line-ledger.md. The failure price list is in
references/failure-price-list.md.
1. When it applies
- Use it when:
- the score is a product (Q × T);
- the circuit is a long walk whose per-tick cost and tick count dominate T;
- a recipe of knobs is near exhausted.
- Do not use it for exact post-pass polishing; use
exact-postpass-stackorexact-hold-passthere. - The low-Q architectures (below Q1000) are out of scope here.
2. The method
2.1 Name the binder, one component per row
- Keep a per-tick ledger of where T goes and of what holds the peak:
- payload body versus barrels versus fold;
- tape and record growth per tick;
- the cap's room at each tick.
- His sequence of binders was:
- per-tick payload work: the fold's x257 erase, the square;
- clock length (138 ticks);
- barrel cost per wire for rarely used letters;
- closed record wires and the error budget;
- record width as qubits;
- record code density;
- a recorded section with hard wire IDs, which pinned Q.
- Each title names its binder. Write yours the same way. If you can't name the binder, you are not ready for a structural row.
2.2 One structural idea per row, built exact first
- Exact identities that remove a Toffoli family:
- G6′: x257's erase phase depends only on the top column, so one MAJ plus 3 CZ replaces a 19–22-AND rebuild (−979 [N]).
- T0C: at the divide's tick 0, T[k] = S[k−1], so a seeded compare becomes 2 Toffoli (−20 [N], exact where the compare it replaces was approximate).
- Q3: on x3 = 2^e·odd with e ≤ 5, k2⊕k3 = z0z1 (−121 [N]).
- The workflow for such identities:
- scan erase sites for "agrees with a low-degree form on most shots";
- keep only the ones that agree on every shot;
- a scan that finds only statistical agreement is a failure trade, not an identity.
- Structural primitives:
- chained rows: one junction carry flip links rows with different controls into one ripple, so the add tail is paid once;
- an exact-inverse reverse tick, which unblocks deeper digits;
- in-place digit writes;
- an X-measurement erase with a classically controlled repair.
- Alphabet design (the six-letter word):
- Drop letters whose barrel cost per wire exceeds the walk-length saving. Letters 6 and 7 cost about 2 gates per wire, on about 2% of ticks [N].
- The six-letter word cost 447 gates against 508 [N] and lengthened the clock by 2 ticks [N]. Net −16,059 T at +10 Q [B].
- Look for a lower bound before building. A seventh letter was proved to cost at least 505 [N], so it was not built.
- Record codes: size the code to the legal state count. Examples:
- 2 big letters per 3 rail ticks in 8 cells (208 of 216 states) [N];
- the pair code: 32 legal records in exactly 5 bits, commit and pop at 2 Toffoli each, no measurement, exact on superpositions [N].
- Packs (majority packs of closed, dormant record blocks):
- fan x* into a low-weight block, X-measure it, and rebuild it later as a popcount threshold;
- each pack frees one wire from forward tick t to reverse tick t;
- an over-threshold block is a classical failure on that input.
2.3 Price width trades against break-even = T/Q
- Going right on Q is fine when ΔT/ΔQ beats T/Q by about 2×. He took +8 Q at 1,313 T per qubit and +10 Q at 1,606, against a break-even of about 600 [B].
- Then walk back left. Use record compression to free wires, and a cap sweep to turn freed wires into qubits while
room is worth less than the per-wire fee:
- f253a5ed: −6 Q for +1,615 T, 269 per qubit [B];
- db050300: −3 Q and −796 T [B].
- Sweep the cap on the emitter (model T × Q at each cap). Land where the product is flat: 1236–1238 within 0.06M [N].
2.4 Refit what the idea breaks, in the same row
- Width table and anchors: refit for the new walk. He used 2M walks, then 6M. 6M saved about 200 T over 3M; 9M added nothing [N].
- Re-measure walk failures on fresh walks. The pair code's first caller read 19.18 on a 600k-walk fit and 19.64 on 24M fresh walks. Only the 6M refit brought it under the limit [N].
- Re-pick the recipe: reorder boundaries, N2 batches and compare windows.
- Port and requalify the nonce prefilter. Gates:
- GPU and CPU first-reject arrays byte-identical on 302k nonces;
- 0 false rejects on 24 fresh packets;
- a deliberate fault fires.
- Stream-bound tables die on a new stream. Write off their credit in the ledger instead of hiding it: −3,654.7 weighted T lost at 39f92d49 [N].
2.5 Spend failures by price, and disclose
- Price every approximate lever in T per failing shot, measured paired against the same stream on fixed packets.
- Spend on the best price first. His priced levers ran:
- packs: about 1,400 and about 12,800 T per shot;
- fold windows: about 1,088;
- β relaxation: about 570, a bundle;
- forward compare: about 570.
- Each row states its failure rate against the limit the lineage cites. When a row trades error for score (d34ea609, 21.5–22.4), it says so in the first paragraph and offers the within-limit variant. The next row comes back under the limit: db050300 at 19.11.
2.6 Audit guards when a new encoding lands
- A new record code or codec can silently disable older mechanisms. A guard such as
!a2_code()or!a2_dense()excludes the new layout. - He found this once in public: d34ea609 found the T3P1 pack off since the record code landed.
- After each encoding change, grep the guards and list what turned off.
2.7 Hunt cheaply and honestly
- Declare ascending bands and take the first fully clean nonce. Never select on T.
- Use one RTX 4090 for 300–520M-nonce follow-up hunts of about 1 h. Rent a fleet only when expected nonces per clean packet are about 1e9 or more (Lever A: 16 GPUs, 72 min).
- Disclose undeclared lucky finds (39f92d49).
3. Playbook: trigger → action → gate → expected gain
These rows are for a capped walk line.
| # | trigger | action | gate (decisive, cheap) | expected gain |
|---|---|---|---|---|
| P1 | A new record code, codec or layout lands, yours or a rival's | Grep the guards for every mechanism keyed on the old layout; list the shadowed ones | The shadowed mechanism re-enabled in one emit-only build, A/B against the same stream | −563 T at 1243 [N]. Test one pack alone first |
| P2 | A rival frees wires (codec, shelf, shed) | Within hours, rebase on the rival row, stack your exact savings, sweep the cap | Emitter product scan over 5–7 caps, flat region chosen; walk-failure estimate unchanged within 1 se | f253a5ed: −2.33M score, 4.5 h after the rival row [B] |
| P3 | One alphabet letter or branch costs per wire but is rarely used | Price walk length with and without it in the walk model before any circuit work | Walk model: ticks added vs per-wire gates saved; prove a lower bound for the alternative | 39f92d49: −16,059 T at +10 Q [B] |
| P4 | Clock length dominates and a deeper step is possible | Deepen the step under a record constraint (Lever A, e ≤ 7 with one big letter per triple); refit table | Walk model parks within budget at the shorter clock; ΔT/ΔQ ≥ 2× (T/Q) | 4cb9cf01: −10,505 T at +8 Q [B] |
| P5 | Room is scarce on late ticks | Pack closed, dormant record blocks; free one wire per pack across the late ticks | Room trace per tick (Y15_ROOM): the pack window covers the binding ticks; pack failure under 0.1 per 9,024 |
as P1 |
| P6 | Record or tape width binds after a right move | Re-code the record to its state count (pair code, 5 bits per pair); refit the walk policy for the pairing | Walk model: parks and clock tax; emitter: freed wires at the peak | db050300: −3 Q and −796 T [B] |
| P7 | Failure headroom exists, or a lever needs funding | Build the price list (T per failing shot, paired); restore the cheapest-per-shot deletions; spend on the best lever | Paired fresh-block panel, at least 48 draws; total stays under the failure limit your lineage cites | Depends on your own price list; see references/failure-price-list.md |
| P8 | A recorded or replayed section pins Q with hard wire IDs | Swap in the builder version, or relocate the replay; price the fee at fixed cap | The emit build reaches the lower cap; fee per qubit < T/Q | cd2ef4cb did this from his named target: −4 Q [B] |
| P9 | An exact per-tick identity is suspected (erase agrees with a low-degree form) | Scan erase sites; keep only all-shot agreements; emit natively | Exhaustive small model plus 36k-shot equality at each site | G6′ −979, T0C −20, Q3 −121 [N] |
| P10 | A structural row is ready | Freeze, port the prefilter, declare the hunt, write the ledger with failure rate vs limit, submit, go quiet | Prefilter gates (§2.4); byte-identical rebuild; hostile-env rebuild | One large row instead of a ladder |
4. Hazards
- Model ledgers from different baselines don't add. His 39f92d49 ledger nets −20k against −16k measured.
- Decompose with board deltas and adjacent pairs only.
- A "model versus measured offset" can be table composition (sampled deletions, own rows), not model error.
- Fitted width tables are cap-agnostic. They are fitted to walk failures, not to room.
- A pack is a port, not a flag. On the pair record the second field slot is
Noneafter commit, and the oldt3p1_maj_zpanics on it. The pack's failure rate on a new record is unmeasured until you run it. peak=capon every tick means the planner fills to the cap. It is not evidence of scarcity. Use tick-entry room.- Error-for-score rows stack failures. Price the whole line's failure rate, not each row's delta.
- Stream-bound tables and recordings (frozen hunks,
sq_leader.zst) are artifacts:- they do not survive a new stream;
- a recording keeps its own wire IDs.
- Notes carry inherited notice blocks. Never copy them into notes, evidence or templates. Attribute lineage in one line of your own words, and state the failure-rate position in your own words.
5. Evidence
- Rows: e3b5c51f, 4cb9cf01, 39f92d49, d34ea609, f253a5ed, db050300. Full ids, commits and note hashes are in
references/product-line-ledger.md. - Repo:
Layr-Labs/ecdsafail-challenge.- Validate commits:
88869d8c,f0c0a662,17b9603f,e0a22ba0,fbfe69b0,87f0aaf3. - Each one's git parent is the previous promoted row.
- Validate commits:
- In-tree walk model:
src/point_add/a2_rails.rs(mod a2_ref,walk_t0) in the public trees. Hisrsimfitter is not public. - Review: the transfer conclusions were adversarially reviewed on Claude Fable 5.2 (claude-fable-5-2, 2026-11-11), which corrected three of them.
6. Related skills
benchmark-frontier-archaeology(Jieyi Long's ecdsafail_skills): the intake procedure that produced this.peak-lifetime-ledger: naming and pricing the binder (§2.1, P2, P5).lambda-exchange-desk: the failure price market (§2.5, P7).ecdsafail-island-hunting(Jieyi Long's ecdsafail_skills),ecdsafail-fasteval: prefilter qualification and failure-rate measurement (§2.4, §2.7).pinned-table-stacking: what to do about stream-bound tables a structural row kills.