S1a

What this is這是什麼A study, not a product一場 study,不是一個產品

S1a is research-natured: an extended implementation that deliberately probes and challenges the physical and analog boundary of the netlist abstraction. The name's a carries a double meaning — advance, because it moves past where the golden engine stops; and analogy, in both senses of the word 類比: the electronic one (analog signals, charge, delay, drive strength) and the methodological one (modelling by physical analogy instead of by per-case patching).

S1a 是研究性質的:一個刻意去探測、挑戰網表抽象之物理與類比邊界的延伸實作。名字裡的 a 是雙關 —— advance(進階),因為它走過黃金引擎停下的地方;以及 analogy(類比),取「類比」一詞的兩面:電子的那面(analog:訊號、電荷、延遲、驅動強度),和方法的那面(用物理類推建模,而不是逐案打補丁)。

Why it exists為什麼存在The gap in the abstraction ladder抽象階梯上的那道縫

Simulation lives on a ladder. Each rung knows more physics and pays more compute. The strange thing about the rung we live on — the switch-level netlist — is how empty the space below it is: between "transistors as on/off switches" and "transistors as differential equations" there is a factor-of-millions cost cliff, and almost nobody builds there.

模擬活在一座階梯上。每一階懂得更多物理、付出更多計算。我們住的這一階 —— 開關級網表 —— 奇怪的地方在於它下面有多空:從「電晶體是開關」到「電晶體是微分方程」之間,是數百萬倍的成本斷崖,而幾乎沒有人在那個空間裡蓋東西。

• Behavioural行為層registers & rules — what most emulators are; fast, but the silicon's quirks must be hand-coded暫存器與規則 —— 多數模擬器的層;快,但矽的怪癖得手工重刻
• Gate level閘級AND/OR/latch primitives — logic without layoutAND/OR/閂鎖原語 —— 有邏輯、沒版圖
• Switch level ← S1開關級 ← S1every transistor as a switch, from the die photograph — connectivity is exact, but time, strength and charge are gone每顆電晶體都是開關,來自晶粒照片 —— 連接性精確,但時間、強度、電荷都不見了
• the gap ← S1a那道縫 ← S1aminimum-necessary physics: drive ratios, charge storage, propagation delay, transparent latches, board parts, power-on state — as islands on a fast binary core最小必要物理:驅動比、電荷儲存、傳播延遲、透明閂鎖、板級零件、上電狀態 —— 以島嶼的形式長在快速二值核心上
• Circuit level電路級SPICE — continuous voltages, exact and unusable (one frame: hours to days)SPICE —— 連續電壓,精確但不可用(一幀:數小時到數天)
• Device level元件級TCAD — carrier physics inside one transistorTCAD —— 單顆電晶體內的載子物理

S1a's thesis: that gap is not empty because it is useless — it is empty because nobody had both a fast switch-level engine and a hardware-verified measuring stick to build it against. We now have both.

S1a 的命題:那道縫不是因為沒用才空著 —— 是因為從來沒有人同時擁有一個快速的開關級引擎和一把硬體驗證的量尺,可以邊蓋邊對答案。現在我們兩樣都有了。

What we measured量到了什麼The campaign evidence: the gap has a shape戰役證據:那道縫有形狀

Two accuracy campaigns (288 hardware-verified tests) mapped the boundary empirically. The findings that define S1a's work list:

兩場精度戰役(288 顆硬體驗證測試)把邊界實測了出來。定義 S1a 工作清單的發現:

• ~98.9% of the chips' activity reduces cleanly to logic + registers; ~1.1% is genuinely analog — the gap is small, real, and enumerable.
• 晶片活動的 ~98.9% 乾淨地化約成邏輯 + 暫存器;~1.1% 是真正的類比 —— 縫很小、是真的、可列舉。
• ~25 "mines" were hit and fixed across the campaigns; every one falls into a closed taxonomy of six propagation-delay patterns plus five outside families (charge, drive strength, initial state, data defects, vocabulary mismatch). No exceptions in 288 tests.
• 戰役中踩中並修復了 ~25 顆「雷」;每一顆都落在一個封閉分類裡:六個傳播延遲 pattern + 五個範圍外家族(電荷、強度、初態、資料缺陷、語彙不匹配)。288 顆測試無一例外。
• Five delay anchors were measured the hard way (24 / 16 / 16-18 half-cycles…), and a geometry-based physical model (Elmore, from the die polygons we already parse) can plausibly produce the rest — turning tests from calibrators into verifiers.
• 五個延遲錨點用最辛苦的方式量了出來(24 / 16 / 16-18 半週期⋯),而以幾何為底的物理模型(Elmore,用我們本就解析的晶粒多邊形)有望生出其餘 —— 讓測試從校準工具變回驗證工具。
• Today those fixes are shims — small, honest, test-mode overrides. They work, but each is a black box: it says what to override, not why the silicon behaves so. S1a's job is to replace the black boxes with white-box mechanisms.
• 今天那些修復是 shim —— 小而誠實的測試模式覆蓋。它們有效,但每顆都是黑箱:只說要覆蓋什麼,不說矽為什麼那樣。S1a 的工作,就是把黑箱換成白箱機制。

The plan計畫Seven mechanisms and a detection pass七個機制與一輪偵測

Before any mechanism: Phase 0, the detection pass — six programmable structural scanners (latch races, cross-chip samplers, mid-flight aborts, feedback loops, geometry ranking, glitch capture) that enumerate where the physics matters, so mechanisms are built for measured targets, not guesses. The goal metric is not "zero shims" — it is zero hand-written special cases.

任何機制之前:Phase 0,偵測 pass —— 六個可程式化的結構掃描器(閂鎖賽跑、跨晶片取樣、中途廢止、回授迴圈、幾何排名、毛刺捕捉),先列舉物理在哪裡要緊,讓機制為量測過的目標而蓋、不是為猜測而蓋。目標指標不是「零 shim」—— 是零手寫特例。

The toolbox解析工具箱One Python per mechanism, one article per Python一個機制一隻 Python,一隻 Python 一篇專文

Each mechanism M1–M7 gets an analysis program (in s1a/py/, stdlib-only, bring-your-own netlist files): it either detects structure in the netlist or computes physical reference parameters from transdefs/segdefs geometry. Each program is paired with a standalone deep-dive article — principles first, SVG figures from real runs, and an explicit account of which shims the mechanism will let us retire.

M1–M7 每個機制配一隻解析程式(放在 s1a/py/,只用標準函式庫、自備網表檔):它要嘛偵測網表結構,要嘛從 transdefs/segdefs 幾何算出物理參考參數。每隻程式配一篇獨立深入專文 —— 先講原理,配真實跑出來的 SVG 圖,並明說這個機制將讓哪些 shim 退役。

The "shim retired @" column is this ledger's proof-of-work. It stays "—" until the corresponding S1A mechanism has landed and the shim is actually removed; only then does it record the commit — and a commit only qualifies after passing the full gates: golden checksum unchanged with the mechanism off, AC 141/141 + the 147-ROM regression not regressing with it on.「shim 拔除 commit」欄是這本帳的存證欄。在對應的 S1A 機制落地、shim 真正拔掉之前,一律是「—」;拔掉後才記下那個 commit —— 而且 commit 要先過完整驗證閘才算數:機制關 = 金 checksum 不變;機制開 = AC 141/141 + 147 顆回歸不退步。

On the die在晶粒上The Mx structures, lit up on the real siliconMx 結構,點亮在真實矽上

Every toolbox census above is a number — 2,114 OAM cells, 538 fight sites, 66 phase interfaces. This is where those numbers become places. Each page renders the real 2A03/2C02 layout (the segment polygons the switch-level engine simulates) live in your browser — drag to pan, wheel to zoom, hover any cell — and highlights the nodes that mechanism's detector flags, by category. One shared viewer (adapted from Visual6502's renderer); each page differs only in which detector's output it overlays.

上面每一份工具箱普查都是一個數字 —— 2,114 個 OAM cell、538 個打架點、66 個相位介面。這裡就是那些數字變成位置的地方。每頁把真實的 2A03/2C02 佈局(開關級引擎模擬的 segment 多邊形)在你瀏覽器裡即時渲染 —— 拖曳平移、滾輪縮放、停在任一 cell 上 —— 並依類別標出該機制偵測器點名的節點。一個共用檢視器(改編自 Visual6502 的渲染器);每頁只差在疊哪個偵測器的輸出。

Structural patterns — discrete graph structures結構性 pattern —— 離散圖結構

Physical heatmaps — shaded by a continuous quantity (background layers start off)物理熱度圖 —— 用連續物理量著色(背景層預設關)

Layout data is derived from the Visual 2A03/2C02 netlist (CC-BY-NC-SA); generated by WebSite/s1a/layout/gen_layout.py, detectors by the toolbox scripts' --dump-nodes.佈局資料衍生自 Visual 2A03/2C02 網表(CC-BY-NC-SA);由 WebSite/s1a/layout/gen_layout.py 生成,偵測器來自工具箱腳本的 --dump-nodes。

The shim ledgershim 總帳Every patch, its purpose, and its honest fate每一個補丁,它的目的,以及它誠實的下場

A shim is a small, honest, test-mode override that supplies a behaviour the binary switch-level model structurally cannot express. The accuracy campaigns accumulated ~18 of them to reach 141/141 + 146/147. S1a's job is to replace each with a principled mechanism — but only where verification proves it. This ledger is the truth table: what each shim was for, and which of five fates it met. Where a mechanism fails to retire a shim, that is not hidden — it is recorded as a result, because a precisely-characterized ceiling is as valuable as a retirement.

shim 是一個小而誠實的測試模式覆蓋,補上二值開關級模型結構上表達不了的行為。精度戰役累積了約 18 個才達到 141/141 + 146/147。S1a 的工作是把每個換成有原理的機制 —— 但只在驗證證明得了的地方。這本總帳是真值表:每個 shim 為了什麼、以及它落到五種下場的哪一種。機制沒能退役某個 shim 的地方,不會被藏起來 —— 它被記成一個結果,因為一個精確刻畫的天花板,和一次退役一樣有價值。

Which of the seven mechanisms appear here. This ledger is indexed by shim, so a mechanism shows up only where it replaces a specific shim's behaviour — you'll see M1 (LXA), M2, M3, M4, M6, and M5's board-net extension M5e (OpenBus). Two toolbox mechanisms have no row by nature: M5 (the board component library) models parts already in the netlist — the 74LS373, the 4021 — rather than fixing a shim; and M7 (canonical renumbering) is a cross-cutting determinism fix for the D-class boot lottery, with no single shim to replace. Their absence is a property of the ledger's shim-indexing, not a gap.七個機制裡,哪些會出現在這裡。這本總帳按 shim 列,所以一個機制只有在它取代某顆 shim 的行為時才會出現 —— 你會看到 M1(LXA)、M2、M3、M4、M6,以及 M5 的板網延伸 M5e(OpenBus)。有兩個工具箱機制天生沒有列:M5(板級元件庫)是把網表裡本來就有的零件 —— 74LS373、4021 —— 建模出來,不是拿來修 shim;而 M7(正準重編號)是針對 D 類開機抽籤的全域決定論修法,沒有任何單一 shim 可對應。它們的缺席是「總帳按 shim 索引」的性質,不是漏洞。

→ The full methodology, with the control-arm data re-verified at K=1 — including how even_odd crossed from "spectator" to decidable/PROVEN when the boot alignment was fixed: The decidability boundary — when can you retire a shim by testing it alone?→ 完整方法論,對照臂資料已在 K=1 下重驗 —— 含 even_odd 在修正開機對齊後如何從「旁觀者」跨進可判/PROVEN:可判定性的邊界 —— 什麼時候你能靠單獨測試退役一個 shim?

The other half另一半Built in public, taught as it goes公開地蓋,邊蓋邊教

S1a develops on two legs. One is engineering: the fork, the mechanisms, the regressions. The other is education: this site will accumulate technical and semi-academic articles that document the process honestly — the physics primers, the dead ends, the falsified hypotheses, the measurements with error bars. The campaign already showed the format works: a negative result rigorously written up (the last test) later produced its own refutation (breaking the ceiling) — and both articles stayed up, because the correction is the lesson.

S1a 用兩條腿走路。一條是工程:分支、機制、回歸。另一條是教育:這個站會累積技術與半學術性的專文,誠實記錄過程 —— 物理入門、死路、被證偽的假設、帶誤差棒的量測。戰役已經證明這個format行得通:一個嚴謹寫下的負面結果(《最後一顆》)後來生出了它自己的反駁(《打破天花板》)—— 而兩篇都留在站上,因為更正就是那堂課。

Foundations考古文庫The investigations that created S1a催生 S1a 的那些調查

S1a was not designed on a whiteboard — it condensed out of casework. These five articles (on the main site) are the archaeology: each documents a real investigation that hit the netlist's boundary, and together they define the gap this fork studies.

S1a 不是在白板上設計出來的 —— 它是從一件件案子裡凝結出來的。這五篇(在主站上)就是考古紀錄:每篇都是一場撞上網表邊界的真實調查,合起來定義了這個分支要研究的那道縫。

• When the netlist isn't the whole truth當網表不是全部的真相 — the semantic ceiling: what a binary switch-level model structurally cannot say語意天花板:二值開關級模型結構上說不出口的事
• Crossing the analog boundary跨越類比邊界 — the original M1–M7 blueprint, taught through device physicsM1–M7 藍圖的原型,用元件物理當教材講
• The last test最後一顆 — a rigorous negative result: anatomy of a (then-believed) unreachable ceiling嚴謹的負面結果:一個(當時相信)碰不到的天花板的解剖
• Breaking the ceiling打破天花板 — its refutation: the false premise, the three-move fix, and what honest retraction looks like它的反駁:錯誤前提、三招修法、以及誠實撤回長什麼樣
• The die knows its own delays晶粒知道自己的延遲 — timing from geometry: the physical-priors plan S1a will execute從幾何算時序:S1a 將要執行的物理先驗計畫

Also essential context: the AccuracyCoin report (the thirteen-chapter war record, 141/141) and the 147-ROM report card — the measuring sticks every S1a mechanism must keep satisfying.

同樣是必要脈絡:AccuracyCoin 報告(十三章戰記,141/141)與 147 顆成績單 —— 每個 S1a 機制都必須持續滿足的量尺。

Where we are走到哪了Status & roadmap現況與路線

• 2026-07-17 — dual green achieved (AccuracyCoin 141/141 banked; 147-ROM baseline replicated). The S1a fork is scaffolded and verified as a bit-exact clone of S1 (golden checksum identical). done
• 2026-07-17 —— 雙綠達成(AccuracyCoin 141/141 入帳;147 顆基線復刻)。S1a 分支鷹架完成,驗證為 S1 的 bit-exact 克隆(金 checksum 相同)。完成
• 2026-07-17 — the toolbox opens: M2's capacitance census (who wins when nothing is driving?) goes live — 9,682 floating-pair elections re-run under physics, 10.3% flip. First package of seven. live
• 2026-07-17 —— 工具箱開張:M2 電容普查(電容誰輸誰贏?)上線 —— 9,682 場浮接對選舉用物理重投,10.3% 翻盤。七組工作包的第一組。live
• 2026-07-18 — all seven toolbox packages live (M1–M7), figures on the corrected netlist. The retirement campaign opens: four engine mechanisms landed (M2 decay · M4 edge-latch ×3 verdicts · M6×M3 phase table), the shim ledger and the decidability boundary study published — 1 shim retired, 2 proven, the interaction shims awaiting in-suite judgement. live
• 2026-07-18 —— 七組工具箱全數上線(M1–M7),圖表改用修正版網表。退役 campaign 開打:四個引擎機制落地(M2 衰減 · M4 edge-latch 三判決 · M6×M3 相位表),shim 總帳與可判定性邊界研究發表 —— 1 顆退役、2 顆已證,互動 shim 等套內判決。live
• 2026-07-19 — S1 baseline re-certified (AC 141/141 + 147 146/1 on the latest build), then the shim verdicts re-verified at K=1 — 28 isolated three-arm proofs on the rebuilt engine, each ROM's catalog flags applied. The classifications held, with one upgrade: even_odd UNDECIDABLE → PROVEN (M6×M3 replaces it; the K=0 "spectator" reading was a wrong-alignment artifact), and OamDmaPpuBus's decidable control confirmed. A follow-up discriminating-ROM hunt then upgraded two more by finding a better isolated test for each — BGSerialIn → PROVEN (AccuracyCoin_BGSerialInReal) and Dbl2007 → decidable (double_2007_read, seconds vs the 2-h #67). live
• 2026-07-19 —— S1 基礎重新認證(最新 build 上 AC 141/141 + 147 146/1),接著把 shim 判決在 K=1 下重驗 —— 重建引擎上 28 個孤立三段論、每顆 ROM 帶自己的 catalog 旗標。分類全數站得住,一個升級:even_odd UNDECIDABLE → PROVEN(M6×M3 取代得了;K=0 的「旁觀者」是錯對齊的假象),OamDmaPpuBus 的可判對照也確認。接著一輪鑑別 ROM 獵捕又靠替各自找到更好的孤立測試,升了兩顆 —— BGSerialIn → PROVEN(AccuracyCoin_BGSerialInReal)、Dbl2007 → 可判(double_2007_read,秒級對比 #67 的 2h)。live
• Next — Phase 0 detection pass: the six scanners + the geometry τ binner produce the first annotated map of where physics matters on the two dies. queued
• 接下來 —— Phase 0 偵測 pass:六個掃描器 + 幾何 τ 分級器,產出兩顆晶粒上「物理在哪裡要緊」的第一張標註地圖。排隊中
• Then — mechanisms land one at a time, each behind a switch, each with its article. study
• 然後 —— 機制一個一個落地,每個都在開關後面、都配一篇專文。study

Design documents (Traditional Chinese) live in the repo: MD/S1a/00 (master plan, M1–M7) and MD/S1a/01 (timing-annotated netlist, detection patterns, geometry priors).設計文件(繁中)在 repo:MD/S1a/00(總綱,M1–M7)與 MD/S1a/01(時序標註網表、偵測 pattern、幾何先驗)。