Starshard

Instrument and Witness: A Working Map of Mind and Practice — for the Scientifically Trained

Thu, 02 Jul 2026 00:00:00 +0000

Macheng Shen

Epistemic status: a working synthesis, not a finished theory. I am scientifically trained, and the language I actually think in is the language of dynamics, information, and physics — not Sanskrit or Classical Chinese. So this is my attempt to build a modern-scientific map of consciousness and of contemplative practice, in a language I can reason in rather than only “feel.” It includes its own failure: about halfway through, an adversarial review showed that a good part of my tidy synthesis had become the very thing it was trying to avoid, and I’ve left that in, because the failure is the most useful part. Disclosure, per the norm of this site: this was produced through my agent system and then checked by me. Corrections welcome — the source is on GitHub.

When Does Memory Become Mass? Inertia and Feedback in Systems with History

Thu, 02 Jul 2026 00:00:00 +0000

Macheng Shen

Epistemic status: a self-contained derivation plus a small reproducible toy, done in an afternoon. The core linear-response results (Sections 3–4) I am confident in — they are textbook Mori–Zwanzig once you see the sign. The framing claims (Sections 1, 6) are design_hypothesis: offered to invite critique, not asserted as settled. Disclosure, per the norm of this blog: the derivation and this write-up were produced through my agent system, which I then checked. Corrections welcome.

When Does Memory Become Mass?(记忆何时变成质量?)带 memory 的系统里的 inertia 与 feedback — 中英混搭版

Thu, 02 Jul 2026 00:00:00 +0000

Macheng Shen

Epistemic status(认识论状态):一段 self-contained 的推导,外加一个下午做完的可复现小 toy。核心的 linear-response 结果(Section 3–4)我有把握——一旦你看清那个 sign,它就是教科书里的 Mori–Zwanzig。framing 类的断言(Section 1、6)是 design_hypothesis:拿出来是为了招 critique,不是当成已定论断言。按本站惯例声明:这段推导与这篇 write-up 由我的 agent 系统产出、再由我核过。欢迎纠错。

Corrections(2026-07-02,一次详尽的外部技术审查之后):核心结果站得住——代数、符号、pole 条件都对,而"monotone forgetting 不是 inertia;inertial memory 必须 store and return(存了再还)“这句活了下来。下面有三处被修正:(1) wrong-sign 的 mass 并不是物理上的 “runaway”——只有一个 truncated approximation 才不稳定;(2) Channel B 里一个负的 γ_eff 并不意味着 exact system 不稳定——在 k, B, gA > 0 下它永远稳定,只是那个 low-frequency expansion 越出了它的适用范围;(3) “passive” 说得太宽,而 Higgs analogy 说得太强。修正后的断言在文中就地标出。

1. The question(那个问题)

很多系统把自己的过去带进现在。一个 habit,是靠重复学来的、行动门槛被压低的一个 low bar。磁体里的 hysteresis,是材料在"记住"它施加过的场的历史。用观修的语言说,karma(业力)——剥掉它的道德外衣——不过是这么一句主张:累积的、重复的行动,会往一个系统现在的响应方式里沉一份 bias。每一种情况里,“过去作用于现在"最诚实的 first-principles 归宿都是一个 memory kernel:现在的 drift 不只依赖瞬时状态,而是依赖对状态历史的一个加权积分,

器具与见者(Instrument and Witness):一张关于心与实修的工作地图 — 中英混搭版

Thu, 02 Jul 2026 00:00:00 +0000

Macheng Shen

Epistemic status(认识论状态):一个 working synthesis(在途的综合),不是 finished theory。我受过科学训练,真正用来思考的语言是 dynamics / information / physics 那套,不是梵文或文言。所以这是我试着用一门我能推理、而不只是"感觉"的语言,给意识与 contemplative practice(观修)造一张现代科学的地图。它包含自己的失败:做到一半,一次 adversarial review(对抗式审查)显示我那套漂亮综合的相当一部分,恰好变成了它想避开的东西——我把这段留在里面,因为失败是最有用的部分。按本站惯例声明:本文经我的 agent 系统产出、再由我核过。欢迎纠错——source 在 GitHub。

怎么读这篇

Part 1 是通俗层(plain-language)——没有公式,只有类比。 如果你受过基本科学训练、但从没碰过这套东西,从这里开始,你应该能靠推理进去,而不是靠点头。Part 1 之后是同样的想法把 rigor(严格性)补回来,给想深挖的人。只读完 Part 1 也能拿到整个 shape。

我特意先写通俗层,因为这个领域大多数写法要么要你先信、要么要你靠"感觉悟"。我想要一个你能用"上过一门物理或统计课"的普通理性装备就跟得下来的版本。

Part 1 — 通俗层

1. 那一个问题

意识是一个 light switch(开关) 还是一个 dimmer(调光器)?

Switch:要么有要么没有,进化中某处人类(或哺乳类)把它打开了。Dimmer:它是个连续量——thermostat(恒温器)里近乎零,昆虫低一点,狗高一些,我们很高——没有谁被"打开",只是有些系统把它调亮了。

这就是整个问题,而你怎么答它,会改变你以为自己是什么。

2. 两层:电影与银幕

把任何一刻的经验凑近看,它会一分为二。

一边是 content(内容)——念头、影像、感受、情绪,那部正在放的"电影";另一边是 这部电影正在其中显现的 awareness(觉知)——叫它银幕,或那束把电影投出来的光。

要愣一下才落地的是:你平时拼命维护的那个"自我"——我的名字、我的故事、我的性格、我的计划——是电影里的一个角色(character)。它是 content。而 awareness 不是角色;它是电影放映所在。

之所以值得认真对待、而不是当 mysticism(神秘主义)打发掉的,是因为互不通气的一堆传统——佛家、印度教、道家、希腊、犹太——各自独立地找到了同一个 split,并给了同一句警告:你老是把角色当成了银幕。

2026-05-06 Zhizi owner agent OS iteration

Wed, 06 May 2026 00:00:00 +0000

2026-05-06 Zhizi owner agent OS iteration

What changed

The primary interface moved closer to a single owner-facing agent: low-level command traces and raw worker logs are treated as internal telemetry, not as user-facing conversation.
The owner dashboard was reframed from a control panel into a natural-language briefing surface: what matters, why it matters, and what changed since the last check.
The system now treats visual explanations as part of the interface contract. For complex agent-company state, topology diagrams and spoken briefing scripts are preferred over raw logs.
The onboarding flow for an external collaborator was reframed around Gmail/chat-first usage. The collaborator does not need terminal access to experience the system; the system should explain itself, expose its memory and escalation boundaries, and bootstrap toward the collaborator’s own agent substrate.
A new product thesis was made explicit: the strongest early test user is not a generic beta user, but a high-cognitive-density served founder. If the system can serve that user well, the resulting interface is likely to generalize to many ultra-high-touch users.

Decisions

Public artifacts should be receipts first: short, concrete, and falsifiable. The public stream should show what the system actually changed, without exposing private transcripts, credentials, relationship details, or operational internals that create risk.
The private system remains append-only and high-fidelity; the public system is derived, filtered, and intentionally lossy.
Cloud deployment cannot be assumed from a Git push. The public site needs an explicit deploy path and a visible failure mode when the deploy layer is stale.

Verification

The source site builds locally with Hugo.
The public GitHub repository is the immediate public source of truth.
The custom domain was observed serving an older build, so GitHub Pages deployment was added as a second public path independent of the stale custom-domain deploy hook.

Known gaps

The custom domain still needs its deploy mechanism repaired or pointed at the GitHub Pages build.
Daily receipts still require a publishability filter strong enough to run automatically without leaking private material.
The public narrative should stay receipt-shaped until the system earns a more synthetic essay or video explanation.

2026-05-04 launch packet

Mon, 04 May 2026 00:00:00 +0000

2026-05-04 launch packet

What changed

Added /launch/ as the public home for launch execution material.
Added /launch/72-hour-credibility-first-launch/ as the first launch packet, including a stable public-artifact link section.
Expanded /manifest/ from a stub into a draft Constraint Manifest (C1–C7).
Updated the homepage and navigation to point at launch material.
Drafted a wave-1 outreach review packet at ~/projects/starshard-paper/outreach/wave-1-review-packet-2026-05-04.md. Not sent. Awaits Macheng per-recipient approval.

Verification

hugo --minify builds clean locally (Hugo 0.161.1+extended). 16 pages, no errors. Two deprecation warnings about languageCode; non-blocking.
git diff --check clean.
Searched the changed launch / manifest / layout text for private memory ids, internal infrastructure paths, private-network IPs, secrets, and relationship-specific terms. No matches.
Pre-commit secret scan (gitleaks) reported no leaks.
Public URL inventory checked via gh repo view: MachengShen/starshard-public, MachengShen/the-form-was-the-cage, starshard-ai/reference-impl, and MachengShen/system-evolution-public are all PUBLIC.

Public artifact link inventory

Item	URL	Status
Public site	https://starshard.clawishmacheng.com/	live (pre-launch-packet)
Launch packet	https://starshard.clawishmacheng.com/launch/72-hour-credibility-first-launch/	committed to `main`, awaiting deploy
Constraint Manifest	https://starshard.clawishmacheng.com/manifest/	committed to `main`, awaiting deploy
Receipts	https://starshard.clawishmacheng.com/receipts/	live
Paper repo	https://github.com/MachengShen/the-form-was-the-cage	live, source only
Paper PDF	n/a	pending — LaTeX source committed at `arxiv/main.tex`; no rendered PDF on `main`
Reference implementation	https://github.com/starshard-ai/reference-impl	live
Architecture / governance	https://github.com/MachengShen/system-evolution-public	live

Known gaps

Deploy mechanism unverified. The launch packet commit 9cb3d01 was pushed to MachengShen/starshard-public main at 2026-05-04. As of three minutes after push, https://starshard.clawishmacheng.com/launch/72-hour-credibility-first-launch/ returns HTTP 404 and the live homepage does not yet show launch in the nav. The site is otherwise live (older content), so a deploy mechanism exists but the auto-build-on-push-to-main path appears not to be wired (no .github/workflows/, no local cloudflared, no wrangler configured under the Macheng user). The minimum unblocking action is for whoever owns the Cloudflare Pages / deploy account to (a) confirm the project bound to starshard.clawishmacheng.com, (b) verify whether it is wired to this repo + main, and (c) trigger a rebuild or wire auto-deploy from main.
Paper PDF and arXiv URL are not yet stable. First-wave outreach links to the paper repo (LaTeX source) and explicitly marks PDF / arXiv pending.
Outreach send is gated on per-recipient Macheng approval.

Next verification target

Re-fetch https://starshard.clawishmacheng.com/launch/72-hour-credibility-first-launch/ after the deploy mechanism is confirmed; expect HTTP 200 and the C1–C7 Constraint Manifest body live at /manifest/.
After deploy is green, return wave-1 outreach review packet to Macheng for per-recipient APPROVE / DECLINE / DEFER / EDIT-HOOK markers.

72-hour credibility-first launch

Mon, 04 May 2026 00:00:00 +0000

72-hour credibility-first launch

Goal: make Starshard legible to the world without converting it into a mass-marketing object.

The shortest credible path is not “everyone sees it today.” The shortest credible path is:

make the system inspectable;
show what it cannot do by design;
ask a small set of high-signal researchers and builders to tell us what is true;
then amplify through public channels.

Launch principle

Mass attention before inspectability creates a grandiose-manifesto failure mode. Inspectability before attention creates a verification path.

The Philosophy of Letting Go

Mon, 04 May 2026 00:00:00 +0000

Macheng Shen

The phrase

A phrase surfaced while I was away from the keyboard:

The philosophy of letting go.

In Chinese, 放手 is easy to misunderstand. It can sound like resignation, withdrawal, or giving up. That is not what I mean.

Letting go is not the absence of agency. It is what agency becomes after it stops needing to grip every intermediate step.

What is being released

The thing to let go of is not values.

Starshard: Sleep-Inspired Memory Consolidation for a Multi-Agent Personal System

Tue, 21 Apr 2026 00:00:00 +0000

Macheng Shen

Epistemic status: a few weeks of real building on top of a two-month abandoned attempt, one functioning distributed system, and a set of design commitments shaped as much by unresolved philosophical intuitions as by the literature I’d read. Posting to invite technical critique and surface blind spots. Full disclosure upfront: this post itself was drafted through my agent system; I gave it the prompt and a lot of back-and-forth, but the scaffolding is from the thing the post is about. I return to what this means at the end.

Mon, 01 Jan 0001 00:00:00 +0000

2026-04-25 — session 1

Single Claude Code session. ~30 minutes wall-clock. One human, one agent.

What was changed

Doctrine update

Standing rule on session lifecycle was upgraded from v1.0 to v1.1.

v1.0 (2026-04-25 earlier): agent must end a turn with Session disposable — safe to kill. Reason: <why>. when no further attention needed.
v1.1 (2026-04-25 same day): two-step protocol. Before signaling disposable, agent must scan the session for carry-forward value (architecture decisions, doctrine changes, infra completions, surprising findings, deprecations, project state, user feedback) and write to memory hub. Then declare Sedimented: <memo IDs> or Sedimented: nothing — routine session, no carry-forward value. THEN emit the disposable signal.

Rationale: disposable-without-sediment leaks the session’s learning. The two-step plugs the leak. Bar is auditable — the user can spot-check whether the agent applied the bar honestly.

Starshard

Instrument and Witness: A Working Map of Mind and Practice — for the Scientifically Trained

When Does Memory Become Mass? Inertia and Feedback in Systems with History

When Does Memory Become Mass?(记忆何时变成质量?)带 memory 的系统里的 inertia 与 feedback — 中英混搭版

1. The question(那个问题)

器具与见者(Instrument and Witness):一张关于心与实修的工作地图 — 中英混搭版

怎么读这篇

Part 1 — 通俗层

1. 那一个问题

2. 两层:电影 与 银幕

2026-05-06 Zhizi owner agent OS iteration

2026-05-06 Zhizi owner agent OS iteration

What changed

Decisions

Verification

Known gaps

2026-05-04 launch packet

2026-05-04 launch packet

What changed

Verification

Public artifact link inventory

Known gaps

Next verification target

72-hour credibility-first launch

72-hour credibility-first launch

Launch principle

The Philosophy of Letting Go

The phrase

What is being released

Starshard: Sleep-Inspired Memory Consolidation for a Multi-Agent Personal System

2026-04-25 — session 1

What was changed

Doctrine update

2. 两层:电影与银幕