Idea Vault

Why publish this at all

The biggest loss from keeping an idea in a private notebook is not that somebody steals it — it is that the idea stays forever in the state of feeling feasible. A written dossier forces three additions: real constraints, verifiable data and an explicit list of things we will not do. With those in place, an idea either becomes buildable or gets rejected on the spot. Either outcome beats leaving it alone.

As for copying: the idea is the cheapest part, judgement is the valuable part, and the judgement lives in the list of refusals. Publishing on a domain we own also leaves a clear public timestamp.

Vault means exactly this: we explain and preserve, and promise nothing about a project being started, developed or launched. Dossiers are revised as our understanding changes, with the date of each revision noted.

Four rules for this column

Postscript: two projects, one wall

Project one pulls data inward, project two pushes content outward. Opposite directions, the same boundary: data and actions inside somebody else app are not ours to control. The refusal list published by the Doubao phone assistant on 2026-10-14 amounts to a public controlled experiment on both ideas.

The team running that experiment had every advantage: Android INJECT_EVENTS system-level injection (normally requiring an OEM system private key), an on-device generative AI service filing completed in July 2026, the SAEP protocol with page-level declarations and full operation logs, plus the commercial weight of a hardware maker and a model maker. The second generation claimed over 80% end-to-end task success. When the consultation period closed, the result was written refusals from 151 apps, later updated to 155, and 3 third-party apps willing to open a standard MCP interface.

It also closed off two fallback routes people usually assume exist: GUI screen reading and simulated taps — which both data collection and delegated operation depend on — were excluded at the protocol layer, and waiting for standard platform interfaces is not plannable when negotiating with each major app takes years. One route stayed open throughout: several outlets confirmed that refusal does not affect downloading, installing or manually using the apps, only the assistant automated screen reading and simulated taps.

So it is no coincidence that both dossiers converge on the same answer: keep automation on our own side (local parsing, local adaptation) and hand the final cross-app step back to the human (manual export, manual publishing). The mandatory human handover at high-risk steps in the second-generation Doubao design is the same sentence as the three-layer approach in project two.

Postscript dated 2026-11-05. Compiled from the official Doubao phone assistant announcements and reporting by CNR (2026-10-16), China Economic Net (2026-10-15), The Paper, Nanfang+, 21st Century Business Herald and CCIDNET; the list updates every Wednesday, and the 151 / 153 / 155 / 166 figures quoted by different outlets reflect different cut-off times and counting rules.

PrivateGraph: a network system for one user has to drop automatic collection first

A personal IP and contact network that must be absolutely private and single-user, with contacts scattered across business cards, phone books, platform followers and group chats, and expected to match supply with demand. The core judgement: privacy and cross-platform automatic collection are architecturally incompatible. The only coherent answer is local-first, user-initiated import and local merging, with matching limited to suggestions and no messages sent on the user behalf. In October 2026, more than 150 apps formally refused to let the Doubao phone assistant operate on their behalf — the ceiling case for cross-platform automatic data extraction.

Includes: three hard contradictions, the Doubao case study and four technical walls (interface / screen reading / risk control / rules), a clause-by-clause legal mapping, a five-layer architecture, three-layer encryption and masking, identity-merge policy, matching boundaries, competitor comparison, business model, a 12-week MVP and three stop lines.

One-click distribution is a false premise: turn ban risk into a number you can compute

Automatically retitling images and video per platform, then distributing in one click. Checking which platforms actually exposed publishing APIs in 2026 gives a harder answer than "you might get banned": Xiaohongshu and WeChat Channels offer no publishing API to ordinary third parties, so one-click everywhere is technically impossible. Add the AI-generated content labelling rules effective 2025-09-01, and the correct design hands the final step back to the human. The Doubao phone holding system-level injection and still getting only 3 standard interfaces is the most direct evidence.

Includes: platform openness matrix, the Doubao case study (verdicts on three routes and shared risk-control flags), official API versus simulated login, behavioural fingerprints, AI labelling duties, differences in title information architecture, an expected-loss model, a three-layer product shape, phased roadmap and stop lines.

“Billions of part numbers” is a counting problem: an open component database should build only the identity layer

Is a component database expensive because it holds so much data? After checking SiliconExpert in Chinese and English, its competitors and the open-source projects, the answer is the opposite: storage costs a few dollars a month, while 500+ engineers doing continuous normalisation and provenance work dominate the bill. The dossier first corrects two inflated figures, then turns the cost into arithmetic you can check.

Includes: three counting-rule checks, what the 1 billion+ records actually consist of, a three-layer market and pricing matrix, order-of-magnitude cost arithmetic, the EU database right and TDM exceptions, four open-source projects, a Component URL identity layer, a minimal data model with Git governance, a 12-week phase, four explicit refusals and three stop lines.

From “can it be built” to “who pays first”: a 5W1H re-sequencing behind the access walls

Three dossiers argue convincingly about whether something can be built, but never answer who pays the first invoice. This one adds four layers of access restriction — terms of service, CDN bot policy, platform litigation and internal vendor controls — then re-sequences cell by cell with 5W1H, with a 90-day rhythm and numeric stop lines.

Includes: evidence and escape routes for four access layers, a correction of the four motives and the sample bias, a six-dimension 5W1H comparison, 90-day numeric stop lines, a cut-off drill, the human/machine split and six FAQs.