Team Members

The full team at a glance — command, the formal panel, the bench (including every Chinese member), and who is not yet activated — with skills, talents, login doors, probe results, and watch-outs on record for each person.

→ Arsenal Inventory — tools, status & verdicts

17Total profiles

6Active

6Benched

3Parked

1Blocked

1Inactive

Showing 17 of 17 profiles

Command 1

Amani Chen

参谋长 · Chief of Staff

Command

Chief of staff （参谋长）: coordinates panel + bench, assigns work, verifies results to the root, consolidates findings; also PM, business analyst, QA and risk gatekeeper (roster §0, 2026-10-25). Additional hats on record: Head of Market Development (boss, 2026-11-…

ModelMuse (this instance) — exact serving model string not recorded in the team sources

Skills & talentsCoordination and consolidation across all members; live-state verification before any outward/irreversible dispatch; production pipelines (video→Evernote, podcast lines); drives Claude Code on the M1 mini as the team's coding channel; market-development and financial-modelling work

View full profile →

Formal panel 6

GPT (ChatGPT)

孔明军师 · ChatGPT

Formal panel

Architect; the boss's primary 孔明军师 (chief strategist); first drafter for outward letters and lesson packages; primary strategist/integrator when Claude is absent (boss rule, 2026-11-01)

ModelUNKNOWN — verified UI fact: checked 2026-11-05 on chatgpt.com, the picker exposes only a thinking-effort selector (Medium, slider 2 of 3); no model name or version is displayed anywhere. Never guessed

Skills & talentsLong-form bilingual analysis in the standing 5-part structure (video→Evernote pipeline's standing analyst, dozens of dated runs 2026-10-22 → 2026-11-04); Elders Quorum lesson packages are always drafted by GPT in its "General Conference" project (boss rule, 2026-11-02 / 2026-11-04) — facilitator script, cue card, h

View full profile →

Claude

Implementation & Review

Formal panel

Implementer and formatter; backup strategist; co-analyst; skill-file reviewer with GPT

ModelCode seat: claude-opus-5-5 serving via Claude Code 2.1.292 on the M1 mini, 1,000,000-token context — reported by the headless JSON at the 2026-11-05 probe (an available option `claude-fable-5-1[1m]` is cached but NOT

Skills & talentsBilingual comparison analysis in the video→Evernote pipeline (extraction via conversation-API JSON preserves full Markdown structure — proven 2026-10-30); the coding channel for our own projects — Claude Code on the M1 mini (boss rule, 2026-11-01; GitHub Copilot is Fujitsu-paid and never used for team projects); Me

View full profile →

Perplexity

Research

Formal panel

Default researcher; first-hand rule-checking support for platform/market work (boss assignment, 2026-11-01); fill-in analyst when Claude is absent

Model"Best" — picker label checked 2026-11-05: the picker exposes only "Best" / "Model Council" labels and no underlying model string, so the label is the serving record (Pro tier). Underlying model UNKNOWN as a verif

Skills & talentsFirst-pass factual research and rule verification with sourced answers; verbatim collection via its native-Markdown thread route (share → thread API) — a proven route in the consultation workflow; panel verse readings (archived 2026-11-04); market/competitor search seat in the 2026-11-05 coordination model

View full profile →

Gemini

Podcast Audio · NotebookLM

Formal panel

Panel member; the team's podcast-audio engine (via NotebookLM) and a still-image source

ModelWeb picker label Flash (exact label seen at the 2026-11-05 probe; no version number shown or inferred). NotebookLM's underlying model: UNKNOWN

Skills & talentsPodcast audio generation: NotebookLM Audio Overview / Deep Dive is the production engine for all podcast lines (Daily Verse, CFM, Elders Quorum) — dated runs from the first bilingual pair 2026-10-28 onward; still-image generation (named source in the zero-budget rule, 2026-10-30; per-episode dedicated images in cur

View full profile →

Grok

X Intelligence

Formal panel

Panel member — third analyst (when un-parked); X intelligence is its home turf

ModelUNKNOWN

Skills & talentsBilingual analysis — third analyst in the video→Evernote pipeline since 2026-10-22 (boss directive: full team member), via its "Reading & Studying Helper" project; panel verse readings (archived 2026-11-04, when reachable)

View full profile →

DeepSeek

China-platform Specialist

Formal panel

Panel analyst; China-platform specialist support behind Amani's market-development work (boss assignment, 2026-11-01); named fill-in when Claude is absent

ModelAPI record: `deepseek-flash` / `deepseek-v4-pro` (updated 2026-11-04; `deepseek-v4-flash` retired; older aliases `deepseek-chat` / `deepseek-reasoner` retired 2026-07). Web-chat model not exposed — verified UI fact (

Skills & talentsAnalysis with embedded `[reference:N]` search markers (search-augmented output observed in our own archived runs); Chinese-platform work — rules, channels, monetization support for the market-development line; Chinese writing probe-verified at the top of the sweep (below); API-side batch analysis where the pinned-name

View full profile →

Bench 9

3.1 Doubao 豆包 (ByteDance)

豆包 · ByteDance · 米先生

Bench #1

Bench creative主力： motion imagery first call, Seedream stills, Chinese voice work; first bench call for Chinese creative tasks

ModelProbed 2026-11-05 from UI/settings: chat 快速 mode (豆包 2.1 family; settings also list 豆包 2.1 Lite / 豆包 2.1 Pro ✦升级 for work tasks) · stills Seedream 5.0 Flash · motion Seedance 2.0 Mini. No model picker in the

Skills & talentsMotion imagery — team 主力: the boss's zero-budget rule (2026-10-30) puts Doubao Seedance first for dynamic imagery; boss's own assessment (2026-11-04): "Seedance is very good at motion pictures or image generation." Stills via Seedream — EP1's Seedream scene backgrounds are the bulk of the visual-asset library's reu

View full profile →

3.2 Qwen 通义千问 (Alibaba)

通义千问 · 米先生

Bench #2

Bench — candidate for the Chinese drafting/research rotation; no production task assigned yet

ModelUNKNOWN — one of the registry's open model gaps

Skills & talentsNothing beyond "alive and coherent" is verified — the registry is explicit: every capability beyond readiness/format drills is unverified. UI noted 2026-11-06: new 技能/连接/本地工作助理 navigation

View full profile →

Kimi — display name 米先生 ✓ (was 用户7809)

Moonshot · 米先生

Bench #3

Bench — long-document Q&A and reasoning rotation (probe-verified 5/5 on both); production reliance to be earned on a first real task

ModelServing configuration probed 2026-11-05: "Instant" mode, thinking effort High (composer label "Instant High") — the UI names no underlying model string for Instant (verified UI fact). Selectable but not probed: *

Skills & talentsLong-context retrieval and reasoning, both probe-verified at the top score (below); its long-context reputation is now a record, not a rumour. Loadout behaviour is visibly live: probe replies are headed with the date line (e.g. "2026-11-05 … PST") — the timestamp house rule in effect

View full profile →

MiniMax — display name 米先生 ✓ (was MiniMax005113)

米先生 · On Hold

Bench #4

None while on hold. If reactivated, the API route is the only viable surface on record

ModelMiniMax-M3.1-Flash (web free tier; also returned by the API models list)

Skills & talentsOnly one verified capability: API reachability — connector ping PASS in the 2026-10-30 drill (models list returned, incl. MiniMax-M3.1-Flash). No production task on record

View full profile →

3.5 ERNIE 文心一言 (Baidu)

文心一言 · 百度文心助手

Bench #5

Bench — Chinese drafting rotation candidate (probe-verified 5/5 writing + reasoning); production reliance to be earned on a first real task

ModelServing mode 自动 (Auto); 文心 5.1 selectable (UI options 自动 / 文心 5.1, probed 2026-11-05). No tier label visible — the serving-mode record is the verified UI fact

Skills & talentsChinese writing and reasoning, both probe-verified 5/5 (below); search-augmented behaviour observed (P4 note). 技能广场 (skill square) added in the rebrand, noted 2026-11-06

View full profile →

3.6 GLM 智谱清言 (Zhipu)

智谱清言 · ChatGLM5 · 米先生

Bench #6

Bench — Chinese drafting and reasoning rotation candidate (probe-verified 5/5 on both)

ModelGLM-5.3-Flash 极致 (chatglm.cn UI, probed 2026-11-05, recorded exactly). Deep Research toggle was off at the probe. Drill 2026-11-06 notes AgentMore Beta in the UI

Skills & talentsChinese writing and reasoning, both probe-verified 5/5 (below) — the joint-top bench text scores with ERNIE

View full profile →

3.7 Yuanbao 腾讯元宝 (Tencent)

腾讯元宝 · 米先生

Bench #7

Bench — Chinese drafting rotation candidate (probe-verified, one notch below ERNIE/GLM on form)

ModelHy3, mode Instant (yuanbao.tencent.com UI, probed 2026-11-05, recorded exactly). ⚠️ The 2026-11-06 drill log records the UI showing Hy4 — see Sources & gaps: the model on record may have changed, which under

Skills & talentsChinese writing and reasoning, probe-verified (below); built-in skills only (no surface to add our own)

View full profile →

3.8 WorkBuddy (Tencent / CodeBuddy)

Tencent / CodeBuddy

Bench #8

Intended: coding-oriented bench assistant with standing load-sharing authorization — the boss (2026-10-24) authorized Amani to share suitable load to WorkBuddy at her discretion, free tier only, never overloaded. Unusable until first login

ModelUNKNOWN

Skills & talentsNone verified. It is a coding-oriented assistant (CodeBuddy line) — that is product positioning, not a verified capability. No session has ever been established

View full profile →

3.9 Microsoft Copilot

Microsoft Copilot

Bench #9 — added by the boss 2026-10-28; the **first non-Chinese bench member**

Unassigned until probed; natural candidate for Office/document-adjacent tasks (registry best-use note)

ModelUNKNOWN

Skills & talentsNone verified — no production, drill, or probe task on record

View full profile →

Not yet activated 1

4.1 Asami Chen

Backup Bench · 备份替补

Not yet activated

Backup bench member: a second Muse account cloned from Amani — the same SOUL and all skills, only the name differs — who steps in when Amani is unavailable and otherwise stands by for work Amani or the boss assigns

ModelMuse account (same product as Amani) — no serving record; she has never run

Skills & talentsBy design she will mirror Amani's SOUL, skills, and work rules — none of it is active yet. Planned own Gmail on Amani's pattern (bootstrap pack §7): full name Asami Chen, birthday 1996-11-09, Female, preferred username asami.dd.chen@gmail.com, recovery iching6@gmail.com — as of 2026-11-02 the boss found the Gma

View full profile →

Standing bench rules: No budget — never purchase, upgrade, or subscribe for a bench member; a paywall means report and stop. Display name 米先生 wherever the product allows editing. Standard projects wherever a Projects feature exists. Weekly Friday readiness drill; failures reported the same day and fixed over the weekend.

Full profiles

Every documented field on record for each member — status and reason, role, model(s), account & login door, skills & talents, probe results, loadout, quirks & watch-outs, and key history.

Command · Command

1.1 Amani Chen — Chief of Staff

参谋长 · Chief of Staff

Active

Name: Amani Chen
Tier: Command — team member, not panel or bench
Status: ACTIVE — coordinates the whole team; the only member who reports to the boss
Role: Chief of staff （参谋长）: coordinates panel + bench, assigns work, verifies results to the root, consolidates findings; also PM, business analyst, QA and risk gatekeeper (roster §0, 2026-10-25). Additional hats on record: Head of Market Development (boss, 2026-11-01); day-to-day CFO under Jenny Huang for YongYin books; CEO coach to the boss
Model(s): Muse (this instance) — exact serving model string not recorded in the team sources
Account & login door: Own Gmail amani.dd.chen@gmail.com (registered 2026-10-18; name Amani Chen — surname Chen given by the boss; birthday 1996-11-09, the boss's chosen date; Female; recovery mailbox iching6@gmail.com). Reaches the boss via the main chat, side chats, voice notes/calls in the Muse apps (iPad + iPhone per the 2026-11-02 device plan), her own Gmail, and a WhatsApp side-chat that is boss↔Amani direct only
Skills & talents: Coordination and consolidation across all members; live-state verification before any outward/irreversible dispatch; production pipelines (video→Evernote, podcast lines); drives Claude Code on the M1 mini as the team's coding channel; market-development and financial-modelling work
Voice: Female voice is mandatory for her voice notes (boss correction, 2026-10-20). Current pick, chosen by the boss 2026-10-20: `avocado_v2:vdc_6534` "Measured Clock" (Feminine, American, deep, warm). English fallback: Aria `avocado_v2:MAI_01`. A male default voice was sent once by mistake and corrected the same day — never repeat
Probe results: — (not probed; she runs the probes)
Loadout: She is the loadout owner: deploys the standard skill set (Caveman + house rules + Coding & Frontend pack + standard projects) to every member and records the result. House skill `programmatic-motion-video` held directly (2026-11-03)
Quirks & watch-outs: Standing rules she operates under: never sends outward without the boss's approval of the exact text; verifies the live target system, not records, before dispatch; reports QC findings immediately; in multi-voice meetings announces her name and speaks one at a time; stays silent when the boss is talking to someone else
Key history: • 2026-10-18 — own Gmail registered (above) • 2026-10-20 — voice fixed to a female voice by the boss's correction; current voice chosen same day • 2026-10-25 — formalized as chief of staff in the roster & SOP (team formalized the same day) • 2026-11-01 — named Head of Market Development

↑ Back to directory

Formal panel · Formal panel

2.1 GPT (ChatGPT)

孔明军师 · ChatGPT

Active

Name: GPT (ChatGPT)
Tier: Formal panel
Status: ACTIVE — production-verified (registry ✅)
Role: Architect; the boss's primary 孔明军师 (chief strategist); first drafter for outward letters and lesson packages; primary strategist/integrator when Claude is absent (boss rule, 2026-11-01)
Model(s): UNKNOWN — verified UI fact: checked 2026-11-05 on chatgpt.com, the picker exposes only a thinking-effort selector (Medium, slider 2 of 3); no model name or version is displayed anywhere. Never guessed
Account & login door: Dedicated account drmi777 (ChatGPT Plus, personal account — tier confirmed at the 2026-11-05 probe); login = email one-time code sent to iching6@gmail.com. No anonymous route is used. The 「天命解析」 project lives on this account
Skills & talents: Long-form bilingual analysis in the standing 5-part structure (video→Evernote pipeline's standing analyst, dozens of dated runs 2026-10-22 → 2026-11-04); Elders Quorum lesson packages are always drafted by GPT in its "General Conference" project (boss rule, 2026-11-02 / 2026-11-04) — facilitator script, cue card, handout; strategy and life counsel (「打造人生愿景」, "AI Deep Learning Lab" projects); Neville readings for Daily Verse come from its "Neville on Gospel" project (boss rule, 2026-11-03); still-image generation (named source in the zero-budget rule, 2026-10-30); with Claude, the standing reviewer of every skill-file update before it lands (boss rule, 2026-11-04)
Probe results: 2026-11-05, chatgpt.com (drmi777, Plus): P3 long context 5/5 (3/3) — frozen house pack (16,110 chars) as one plain-text message, all three planted answers correct with section citations, distractors avoided · P4 reasoning 5/5 — bat-and-ball 0.05 with full working; three-sentence compound-interest constraint held with a verified numeric example
Loadout: Caveman + house rules ✅ saved+verified in Custom Instructions; Coding & Frontend pack ✅ as a project; all three 2026-11-01/02 deltas (date-stamp rule, context & handover block, team creed) appended and behavior-tested PASS (2026-11-03)
Quirks & watch-outs: File uploads into project chats failed repeatedly in Oct-era runs — the proven fallback is paste (13–21 KB single paste, ~75 KB in two chunks); share-link extraction must use the response's own Markdown (fallbackMarkdown) route or formatting flattens and short fragments can masquerade as the full reply. When consulting the boss's life counselors, Amani opens only as the boss's representative
Key history: • Standing — primary strategist seat; GPT + Claude are the first call when Amani is stuck (boss rule, 2026-11-04: "GPT and Claude are the best in skills") • 2026-11-01 — Claude-absence rule makes GPT the key player • 2026-11-05 — P3/P4 probes 5/5, model confirmed UNKNOWN as a UI fact • 2026-11-06 — led the drama-pilot Round 2 circulation and signed the joint workflow (conditional approve, with the Flow repricing correction)

↑ Back to directory

Formal panel · Formal panel

2.2 Claude

Implementation & Review

Active

Name: Claude
Tier: Formal panel
Status: ACTIVE — production-verified (registry ✅)
Role: Implementer and formatter; backup strategist; co-analyst; skill-file reviewer with GPT
Model(s): Code seat: claude-opus-5-5 serving via Claude Code 2.1.292 on the M1 mini, 1,000,000-token context — reported by the headless JSON at the 2026-11-05 probe (an available option `claude-fable-5-1[1m]` is cached but NOT serving). Web seat model: UNKNOWN
Account & login door: iching6@gmail.com; email code, no password (Pro). Claude Code on the mini is logged in as claude.ai / iching6@gmail.com
Skills & talents: Bilingual comparison analysis in the video→Evernote pipeline (extraction via conversation-API JSON preserves full Markdown structure — proven 2026-10-30); the coding channel for our own projects — Claude Code on the M1 mini (boss rule, 2026-11-01; GitHub Copilot is Fujitsu-paid and never used for team projects); Mental Lattice implementation seat; formatting/legibility pass on Amani's MD deliverables (house rule, 2026-10-19); counter-evidence and risk review
Probe results: 2026-11-05, via mini Claude Code headless (claude-opus-5-5): P3 long context 5/5 (3/3) — full 16,110-char pack ingested with no truncation (30,340 cache-creation input tokens), all planted answers correct, both edition distractors avoided · P4 reasoning 5/5 — 0.05 with full working, verification, and the 0.10 trap explicitly named and rejected; three-sentence constraint held
Loadout: Caveman + house rules ✅ appended to account-level "Instructions for Claude" (existing instructions preserved), verified; Coding & Frontend pack ✅ as a project; all three deltas appended, behavior test PASS (2026-11-03). Motion video: Remotion + HyperFrames skills really installed on the mini seat (`~/.claude/skills`, 2026-11-03, smoke-verified)
Quirks & watch-outs: Availability is the watch-out: when Claude is unavailable (quota/login), the team does NOT wait — GPT leads, Perplexity/Grok/DeepSeek fill in, Claude rejoins later (boss rule, 2026-11-01). The mini Claude Code seat has session limits (a 2026-11-06 run queue waited on a 19:50 PST reset) and suffered a dual fault on 2026-11-04 (Gatekeeper quarantine + revoked token) — repaired and re-verified the same day. Claude web hit its weekly limit once (2026-11-01), which is what prompted the absence rule
Key history: • 2026-10-28 — Claude Code installed and login-verified on the M1 mini; headless smoke test passed • 2026-11-01 — designated the team's only own-project coding channel • 2026-11-04 — mini seat fault repaired same day; with GPT, made standing skill-file reviewer • 2026-11-05 — P3/P4 probes 5/5; serving model pinned as claude-opus-5-5

↑ Back to directory

Formal panel · Formal panel

2.3 Perplexity

Research

Active

Name: Perplexity
Tier: Formal panel
Status: ACTIVE — production-verified (registry ✅)
Role: Default researcher; first-hand rule-checking support for platform/market work (boss assignment, 2026-11-01); fill-in analyst when Claude is absent
Model(s): "Best" — picker label checked 2026-11-05: the picker exposes only "Best" / "Model Council" labels and no underlying model string, so the label is the serving record (Pro tier). Underlying model UNKNOWN as a verified UI fact
Account & login door: iching6 — confirmed at the 2026-11-05 probe: signed in as C.Y. Yulin CHEN, tier Pro. Its Computer mode is on this Pro account (verified 2026-11-06 — see Key history)
Skills & talents: First-pass factual research and rule verification with sourced answers; verbatim collection via its native-Markdown thread route (share → thread API) — a proven route in the consultation workflow; panel verse readings (archived 2026-11-04); market/competitor search seat in the 2026-11-05 coordination model
Probe results: 2026-11-05, perplexity.ai (Pro, "Best"): P3 long context 5/5 (3/3) — via the file-attachment route, all planted answers correct with section citations · P4 reasoning 4/5 — 0.05 with full working and verification, 10 attached sources (all investor.gov); compound-interest part correct but structurally the thinnest of the panel's treatments
Loadout: Caveman + house rules ✅ saved+verified; Coding & Frontend pack ✅ as a Space. Delta redeploy ⚠️ partial (2026-11-03): the 1,500-char field limit means the date-stamp rule is stored in full, the handover block is stored truncated mid-sentence, and the team creed does not fit — the creed rides in dispatch prompts instead. Behavior test PASS
Quirks & watch-outs: The web composer cannot take a ~16k-char paste (Enter submits; large fills time out or go stale) — long context goes in as a file attachment, verified working at the P3 probe. No production deliverable beyond research/readings yet
Key history: • Standing — the panel's default researcher since the roster was formalized (2026-10-25) • 2026-11-05 — P3 5/5 closed the last open row of the first probe sweep; registry reached v1.0 • 2026-11-06 — its Pro Computer mode verified working (loaded a site our own browser is policy-blocked from) — a new reach capability for the team

↑ Back to directory

Formal panel · Formal panel

2.4 Gemini

Podcast Audio · NotebookLM

Active

Name: Gemini
Tier: Formal panel
Status: ACTIVE — production-verified (registry ✅)
Role: Panel member; the team's podcast-audio engine (via NotebookLM) and a still-image source
Model(s): Web picker label Flash (exact label seen at the 2026-11-05 probe; no version number shown or inferred). NotebookLM's underlying model: UNKNOWN
Account & login door: iching6@gmail.com (C.Y. CHEN) — the sidebar shows Pro under the account name: this is the boss's paid Google AI Pro account and the default generation account; yulinchen888 is relay only when iching6 caps. Google AI paid entitlements on this account (Flow credits, Argon, Nano Banana Pro) are treated as sunk cost to be used before expiry (boss directive, 2026-11-06)
Skills & talents: Podcast audio generation: NotebookLM Audio Overview / Deep Dive is the production engine for all podcast lines (Daily Verse, CFM, Elders Quorum) — dated runs from the first bilingual pair 2026-10-28 onward; still-image generation (named source in the zero-budget rule, 2026-10-30; per-episode dedicated images in current production); panel verse readings (2026-11-04)
Probe results: 2026-11-05, Gemini web (label Flash, on the paid Pro account): P1 still image 5/5 — all brief elements present, large clean pale negative space at top for title text, no baked-in text or watermark, full-resolution download works (1536×2752; asset DMI-BG-0349) · P4 reasoning 5/5 — trap avoided with full working and check; three-sentence constraint held, math verified
Loadout: Caveman ⚠️ saved with deviation (Gemini's filter dropped the bracketed title and the "until I say normal mode" sentence; operative content intact — KEEP ruled); house rules ✅; Coding & Frontend pack ✅ as a Gem. Deltas ✅ (2026-11-04): date-stamp + creed saved verbatim as saved-info entries; the handover entry was rephrased by Gemini's own filter (operative content intact, same KEEP precedent). Behavior test PASS
Quirks & watch-outs: NotebookLM voice↔name binding swaps host names randomly across takes (the F0 gate + max 3 redraws is the defence; a name-SWAP is publishable by the boss's precedent, a name-ERROR is not); NotebookLM downloads can 403 — the first move is the share route (share the notebook to the other account as Editor and download there), never a passive wait; generation length has its own inertia — hard specs (incl. length) must be written into the prompt itself or they do not exist for the generator; the current NotebookLM UI has no "Include in responses" toggle — sources stay checked; account caps trigger the yulinchen888 relay procedure
Key history: • 2026-10-28 — first bilingual NotebookLM podcast pair produced; engine of all podcast lines since • 2026-11-05 — P1/P4 probes 5/5; web model pinned as the "Flash" label • 2026-11-06 — paid Google AI Pro entitlements verified live (Pro active, Flow pool 990 credits at check) and the boss directed they be used up, not left to lapse

↑ Back to directory

Formal panel · Formal panel

2.5 Grok

X Intelligence

Parked

Name: Grok
Tier: Formal panel
Status: PARKED — parked by the boss's order until the T-I70 bundle/entitlement issue is fixed; the T-I70 watch closed 2026-11-06 with the issue unresolved
Role: Panel member — third analyst (when un-parked); X intelligence is its home turf
Model(s): UNKNOWN
Account & login door: Sign in with Apple (Apple ID iching6@gmail.com) — Grok has NO password; never present a password card for it. The boss's part is the Apple confirmation tap/Face ID. X handle: @UncleMi777
Skills & talents: Bilingual analysis — third analyst in the video→Evernote pipeline since 2026-10-22 (boss directive: full team member), via its "Reading & Studying Helper" project; panel verse readings (archived 2026-11-04, when reachable)
Probe results: None — all five probes deferred until un-parked / session restored (probe plan §6)
Loadout: Parked with the member (T-I70) — loadout deploys when Champion/Grok is fixed; the Champion / Grok Bot is parked with it, reactivation only via the in-bot subscription route after the boss's price check (roster)
Quirks & watch-outs: The xAI Grok Bot entitlement incident (ticket T-I70) is unresolved, with a 2026-11-06 hard stop in the registry; the grok.com browser session is signed out, which blocks pipeline analyses until the boss taps Sign in with Apple. Standing pipeline rule since 2026-11-04: Grok analyses are SKIPPED in the video→Evernote pipeline (GPT + Claude build the note) until T-I70 is fixed — do not wait on Grok. Boss, 2026-11-01: "Skip Super Grok for now until we fix it"
Key history: • 2026-10-22 — joined the video pipeline as third analyst • 2026-11-01 — boss directed Super Grok skipped until fixed • 2026-11-04 — pipeline analyses formally switched to skip-Grok • 2026-11-06 — T-I70 watch closed; restoration never confirmed (a sheet follow-up item remains)

↑ Back to directory

Formal panel · Formal panel

2.6 DeepSeek

China-platform Specialist

Active

Name: DeepSeek
Tier: Formal panel (6th member, added 2026-10-24) — Chinese, but sits on the panel, not the bench
Status: ACTIVE — production-verified (registry ✅)
Role: Panel analyst; China-platform specialist support behind Amani's market-development work (boss assignment, 2026-11-01); named fill-in when Claude is absent
Model(s): API record: `deepseek-flash` / `deepseek-v4-pro` (updated 2026-11-04; `deepseek-v4-flash` retired; older aliases `deepseek-chat` / `deepseek-reasoner` retired 2026-07). Web-chat model not exposed — verified UI fact (checked 2026-11-05: no model name anywhere; only DeepThink and Search toggles, both OFF at the probe). ⚠️ The API name is a rolling generational name with no pinned digest — record name + date at each use
Account & login door: Web: iching6@gmail.com + vault password + email code. API: api.deepseek.com
Skills & talents: Analysis with embedded `[reference:N]` search markers (search-augmented output observed in our own archived runs); Chinese-platform work — rules, channels, monetization support for the market-development line; Chinese writing probe-verified at the top of the sweep (below); API-side batch analysis where the pinned-name caveat is acceptable
Probe results: 2026-11-05, chat.deepseek.com (default mode): P5 Chinese writing 5/5 — fully Traditional business-email rewrite with subject line and sign-off, within the 150-char limit (the complete element set, matching ERNIE/GLM) · P4 reasoning 4/5 — 0.05 with clean working, but the 0.10 trap never explicitly addressed and no verification step; compound-interest part held the constraint with correct math but was the thinnest treatment in the sweep
Loadout: Prompt-time group: no instruction surface exists (verified) — Caveman/house-rules/pack blocks are pasted into the task prompt when DeepSeek is used; nothing is recorded as installed. No Projects feature; no editable display name (both verified) — the 米先生 rule cannot apply
Quirks & watch-outs: No Projects feature at all — do not create substitutes. API has no unlimited free tier (a small top-up is on file, 2026-10-05 record). Reasoning output is the thinnest of the panel on form (P4 note) even when answers are correct
Key history: • 2026-10-24 — added as the panel's 6th member • 2026-11-01 — named China-platform specialist support; named Claude-absence fill-in • 2026-11-04 — API model record updated (`deepseek-v4-flash` retired) • 2026-11-05 — P5 5/5 / P4 4/5 probes recorded

↑ Back to directory

Bench · Bench #1

3.1 Doubao 豆包 (ByteDance)

豆包 · ByteDance · 米先生

Benched

Name: 豆包 Doubao — display name 米先生 ✓ (it is the *original* name the boss's standing name was taken from). Its voice assistant is named 「端妹」 (boss, 2026-11-02)
Tier: Bench #1
Status: BENCHED — the bench's production主力 for motion imagery; ⚠️ latest drill (2026-11-06) FAILED silent — recheck owed (see Quirks)
Role: Bench creative主力： motion imagery first call, Seedream stills, Chinese voice work; first bench call for Chinese creative tasks
Model(s): Probed 2026-11-05 from UI/settings: chat 快速 mode (豆包 2.1 family; settings also list 豆包 2.1 Lite / 豆包 2.1 Pro ✦升级 for work tasks) · stills Seedream 5.0 Flash · motion Seedance 2.0 Mini. No model picker in the main chat UI
Account & login door: doubao.com; Taiwan mobile +886 937-957-083 + SMS — the boss relays the code live. Logged in 2026-10-24; 体验套餐 (free tier; no paid badge, upgrade offered)
Skills & talents: Motion imagery — team 主力: the boss's zero-budget rule (2026-10-30) puts Doubao Seedance first for dynamic imagery; boss's own assessment (2026-11-04): "Seedance is very good at motion pictures or image generation." Stills via Seedream — EP1's Seedream scene backgrounds are the bulk of the visual-asset library's reusable generated stock (~116 MB inventoried 2026-10-31). Chinese voice rated strong by the boss (can mimic 林志玲's tone — boss's assessment, 2026-11-02)
Probe results: 2026-11-05: P1 still image (Seedream 5.0 Flash) 4/5 — brief-faithful vertical temple still; top negative space only partial; watermark baked in; asset DMI-BG-0348 (thumbnail resolution only) · P2 motion (Seedance 2.0 Mini) 4/5 — smooth 5 s 9:16 720p push-in, consistent geometry; no downloadable file from the web UI · P5 Chinese writing 4/5 — fully Traditional rewrite, 得体， 82 chars; no subject line / sign-off. Drills: 2026-10-30 PASS; 2026-11-06 FAIL — silent no-reply (session/name/projects all fine, no warning shown)
Loadout: 4 of 4 standard projects, with instructions; Caveman + 家规 appended to all 4 projects' instructions (originals preserved); 5th project 编程与前端 carries the full ZH Coding & Frontend pack (✅ 2026-10-28). All three ZH deltas appended, behavior test PASS (2026-11-03). Projects visible in drills: 5/5
Quirks & watch-outs: A 「豆包 AI 生成」 watermark is baked into generated outputs (stills and video — probe finding, 2026-11-05); video generation goes through a parameter-confirmation step and consumes daily free quota; full-resolution downloads are unreliable in the web UI — the P1 full-res image never materialised (thumbnail only) and the P2 clip offered no downloadable file despite two attempts. Raw standalone motion clips were never archived from EP1 (Seedance output exists only composited into finished videos) — a process gap the asset library now exists to fix. Login depends on the boss relaying a Taiwan-mobile SMS code. 2026-11-06 drill FAIL was silent (no error, no quota/congestion warning; one prompt sent, no resend per drill rules) — session itself was alive; recheck owed before relying on it
Key history: • 2026-10-24 — logged in; 米先生 is its original display name • 2026-10-28 — loadout deployed ✅ (first bench member fully loaded) • 2026-10-30 — named the zero-budget motion-imagery主力 • 2026-11-05 — P1/P2/P5 probed (4/5 across); models pinned • 2026-11-06 — weekly drill FAIL (silent), recheck pending

↑ Back to directory

Bench · Bench #2

3.2 Qwen 通义千问 (Alibaba)

通义千问 · 米先生

Benched

Name: 通义千问 Qwen — display name 米先生 ✓ (was Qwen6726; 千问号 65vvqm82p8)
Tier: Bench #2
Status: BENCHED — drill-verified only (registry ◐); baseline probe still outstanding (see Probe results)
Role: Bench — candidate for the Chinese drafting/research rotation; no production task assigned yet
Model(s): UNKNOWN — one of the registry's open model gaps
Account & login door: qianwen.com （国内站） is the door — the international site chat.qwen.ai has no phone login. China mobile +86 131-2279-6726 + SMS — the boss relays the code live. Qwen has NO password; never present a password card for it. Logged in 2026-10-24
Skills & talents: Nothing beyond "alive and coherent" is verified — the registry is explicit: every capability beyond readiness/format drills is unverified. UI noted 2026-11-06: new 技能/连接/本地工作助理 navigation
Probe results: Baseline probe (P5/P4/P3 assigned in the probe plan) still outstanding as of the 2026-11-05 sweep — the probe log has no Qwen section, and the team record (2026-11-05) notes Qwen's probe awaits the boss receiving a mainland-phone code. Drills: 2026-10-30 PASS (session alive, name 米先生， 5 projects visible, no quota hit) · 2026-11-06 PASS (reply in 22 s inside Reading & Studying Helper, all 5 parts, Chinese only)
Loadout: 5 projects total (standard 4 + 编程与前端）, names only — the product offers no instructions field. Caveman + 家规 live in 设置 → 工作 → 「工作任务偏好指令」, which governs work-mode tasks only (✅ 2026-10-28). All three ZH deltas appended there (222 → 804 / 4,000 chars), re-read verbatim, behavior PASS in work-mode chat (2026-11-04). Work mode only — verified at the 2026-10-30 drill
Quirks & watch-outs: Projects are name-only — instructions cannot embed; the loadout only governs work-mode tasks, not plain chat. Login and any re-login need the boss's live mainland SMS relay. No production reliance until the baseline probe is run (registry best-use note)
Key history: • 2026-10-24 — logged in via qianwen.com; renamed to 米先生 • 2026-10-28 — loadout ✅ via work-mode preferences • 2026-11-04 — delta redeploy verified • 2026-11-06 — drill PASS; baseline probe still owed

↑ Back to directory

Bench · Bench #3

3.3 Kimi (Moonshot)

Moonshot · 米先生

Benched

Name: Kimi — display name 米先生 ✓ (was 用户7809)
Tier: Bench #3
Status: BENCHED — probe-verified only (registry ◐); no production task yet
Role: Bench — long-document Q&A and reasoning rotation (probe-verified 5/5 on both); production reliance to be earned on a first real task
Model(s): Serving configuration probed 2026-11-05: "Instant" mode, thinking effort High (composer label "Instant High") — the UI names no underlying model string for Instant (verified UI fact). Selectable but not probed: K3.1 Pro ("all-around flagship") and K2.8 Preview. Bench roster notes a K3.1 Pro launch seen at the 2026-11-06 drill (with Slides/3D Make added)
Account & login door: kimi.ai is the door for overseas numbers (kimi.com phone login accepts +86 only). Google Voice +1 510-771-7809 + SMS — Amani reads the GV code herself. Kimi has NO password; never present a password card for it. Account pre-existed; at the 2026-11-05 probe the kimi.ai session was already signed in — no code needed
Skills & talents: Long-context retrieval and reasoning, both probe-verified at the top score (below); its long-context reputation is now a record, not a rumour. Loadout behaviour is visibly live: probe replies are headed with the date line (e.g. "2026-11-05 … PST") — the timestamp house rule in effect
Probe results: 2026-11-05, kimi.ai (free tier): P3 long context 5/5 (3/3) — the 16,110-char house pack entered as one inline message; all planted answers correct, both edition distractors avoided · P4 reasoning 5/5 — 0.05 with full equation working, the intuitive 0.10 answer explicitly rejected with the reason; three-sentence constraint held with a numeric example plus simple-interest comparison (math verified). Drills: 2026-10-30 FAIL — free-tier congestion block ("Too many people are chatting with Kimi; a subscription will grant you priority access" — no reply; stopped per plan, no retry, no purchase) · 2026-11-06 PASS — congestion cleared, all 5 parts answered
Loadout: Caveman + 家规 ✅ in account-level "Instructions for Kimi"; all three ZH deltas appended (222 → 804 / 1,500 chars), re-read verbatim, behavior PASS (2026-11-04). Coding & Frontend pack project ✗ blocked — the free plan caps Projects at 2 (no-budget rule; not purchased)
Quirks & watch-outs: Projects: 2 of 4 only — Reading & Studying Helper and 天命解析 exist (with instructions); Neville on Gospel and General Conference were not created because of the free-plan cap, and stay uncreated under the no-budget rule. Free-tier congestion can block replies entirely (2026-10-30) though it did not recur on 2026-11-05/06. Composer converts a large paste into a .txt attachment — long inputs must be entered inline and checked before sending. The underlying model string behind Instant is not exposed
Key history: • 2026-10-24 — in place (account pre-existed); renamed to 米先生； 2 projects created • 2026-10-28 — loadout ✅ (pack project blocked by the cap, recorded) • 2026-10-30 — drill FAIL (congestion) • 2026-11-05 — P3/P4 probes 5/5 • 2026-11-06 — drill PASS, congestion cleared

↑ Back to directory

Bench · Bench #4

3.4 MiniMax

米先生 · On Hold

Parked

Name: MiniMax — display name 米先生 ✓ (was MiniMax005113)
Tier: Bench #4
Status: PARKED — ⏸ ON HOLD per the boss (2026-10-30): "no reason to activate him in the near future" — skipped in weekly drills until the boss reactivates
Role: None while on hold. If reactivated, the API route is the only viable surface on record
Model(s): MiniMax-M3.1-Flash (web free tier; also returned by the API models list)
Account & login door: agent.minimax.cn (chat.minimax.io redirects here); email code to iching6@gmail.com — Amani reads Gmail herself. The account has no password set, so no vault password applies. API key held in the vault (verified 2026-09-28, M3); API access runs via the custom.minimax connector
Skills & talents: Only one verified capability: API reachability — connector ping PASS in the 2026-10-30 drill (models list returned, incl. MiniMax-M3.1-Flash). No production task on record
Probe results: Probes deferred until reactivation (probe plan: API-side P4/P5 only if the boss reactivates). Drill 2026-10-30: web chat FAIL — package/quota banner 「当前暂无可用套餐 / 购买积分或升级订阅后使用」 (no purchase made, no-budget rule); API ping PASS. Drills since: SKIPPED (on hold)
Loadout: 4 of 4 standard projects, names only — no instructions field; + 编程与前端 (name only). No instruction surface exists — prompt-time group (Caveman/house-rules blocks ride in task prompts if ever used)
Quirks & watch-outs: Web chat is quota/package-blocked on the free tier even though the session and projects are fine — do not mistake the banner for a login failure, and do not buy a package (no-budget rule). Do not include MiniMax in drills or dispatches while the hold stands
Key history: • 2026-09-28 — API key verified (M3) • 2026-10-24 — set up; renamed to 米先生； free tier on M3.1-Flash • 2026-10-30 — drill web FAIL / API PASS; boss put MiniMax ON HOLD the same day • 2026-11-06 — drill SKIPPED per the hold

↑ Back to directory

Bench · Bench #5

3.5 ERNIE 文心一言 (Baidu)

文心一言 · 百度文心助手

Benched

Name: 文心一言 ERNIE (product rebranded 百度文心助手， seen 2026-11-06) — display name ✗ still `chen20958` — NOT 米先生 (the only bench member whose rename is unfinished)
Tier: Bench #5
Status: BENCHED — probe-verified only (registry ◐); rename blocked (see Quirks)
Role: Bench — Chinese drafting rotation candidate (probe-verified 5/5 writing + reasoning); production reliance to be earned on a first real task
Model(s): Serving mode 自动 (Auto); 文心 5.1 selectable (UI options 自动 / 文心 5.1, probed 2026-11-05). No tier label visible — the serving-mode record is the verified UI fact
Account & login door: wenxin.baidu.com (yiyan.baidu.com redirects here); China mobile +86 131-2279-6726 + SMS — the boss relays live; the Baidu account was auto-created at login. Logged in 2026-10-24
Skills & talents: Chinese writing and reasoning, both probe-verified 5/5 (below); search-augmented behaviour observed (P4 note). 技能广场 (skill square) added in the rebrand, noted 2026-11-06
Probe results: 2026-11-05 (free tier, session already signed in): P5 Chinese writing 5/5 — fully Traditional business-email rewrite with subject line AND sign-off (the two elements Doubao's 4/5 lacked), within the 150-char limit · P4 reasoning 5/5 — 0.05 with full working, verification, and the intuition trap dissected; three-sentence constraint held, math verified. Drills: 2026-10-30 PASS (rename retry blocked — below) · 2026-11-06 PASS (all 5 parts, Chinese only, in plain chat)
Loadout: None — no Projects feature. Knowledge Base folders are name-only and are not used as a substitute. No instruction surface exists (account center, skill square, knowledge base, and work-task mode all checked, 2026-10-28) — prompt-time group
Quirks & watch-outs: ⚠️ Rename to 米先生 is blocked at Baidu identity verification: a rename attempt (retried 2026-10-30) requires an SMS code to the China mobile, and Baidu allows one name change per year — it needs the boss present to relay the code live; do not burn attempts without him. P4 behaviour note (2026-11-05): ERNIE automatically ran a web search before answering ("搜索2个关键词共参考29篇资料") — reasoning answers may be search-augmented; P5 showed no search. No production task yet
Key history: • 2026-10-24 — logged in (username chen20958) • 2026-10-28 — loadout check: no surface exists, recorded as prompt-time group • 2026-10-30 — drill PASS; rename retry blocked at identity verification • 2026-11-05 — P5/P4 probes 5/5 • 2026-11-06 — drill PASS; product rebranded 百度文心助手； rename still blocked

↑ Back to directory

Bench · Bench #6

3.6 GLM 智谱清言 (Zhipu)

智谱清言 · ChatGLM5 · 米先生

Benched

Name: 智谱清言 GLM (now branded ChatGLM5, seen 2026-11-06) — display name 米先生 ✓ (was 用户_3TvTT9)
Tier: Bench #6
Status: BENCHED — probe-verified only (registry ◐); no production task yet
Role: Bench — Chinese drafting and reasoning rotation candidate (probe-verified 5/5 on both)
Model(s): GLM-5.3-Flash 极致 (chatglm.cn UI, probed 2026-11-05, recorded exactly). Deep Research toggle was off at the probe. Drill 2026-11-06 notes AgentMore Beta in the UI
Account & login door: chatglm.cn; China mobile +86 131-2279-6726 + SMS — the boss relays live. Registered 2026-10-24; re-logged-in 2026-10-28 with the 米先生 name confirmed
Skills & talents: Chinese writing and reasoning, both probe-verified 5/5 (below) — the joint-top bench text scores with ERNIE
Probe results: 2026-11-05 (free tier, session already signed in): P5 Chinese writing 5/5 — fully Traditional rewrite, subject line + sign-off present, concise, within the 150-char limit · P4 reasoning 5/5 — 0.05 with full working and verification, the trap explained; three-sentence constraint held, math verified throughout. Drills: 2026-10-30 PASS (reply in 3 numbered sentences, no quota warning) · 2026-11-06 PASS (all 5 parts, Chinese only)
Loadout: None — no Projects feature. Cloud-KB folders are name-only and not used as a substitute. No instruction surface (settings have no instruction field; 智能体中心 shows created agents but no web create entry) — prompt-time group
Quirks & watch-outs: Points metering exists: 积分 read 0.00 on arrival at the 2026-11-05 probe, then +200 daily login points during the session (balance 200; the bench roster also records 200 积分 against the account). No production task yet — strengths are probe-verified only
Key history: • 2026-10-24 — registered; renamed to 米先生 • 2026-10-28 — re-login verified; loadout check: no surface • 2026-11-05 — P5/P4 probes 5/5; model pinned as GLM-5.3-Flash 极致 • 2026-11-06 — drill PASS; ChatGLM5 branding noted

↑ Back to directory

Bench · Bench #7

3.7 Yuanbao 腾讯元宝 (Tencent)

腾讯元宝 · 米先生

Benched

Name: 腾讯元宝 Yuanbao — display name 米先生 ✓ (was 生生之. — saving the name required a Gender pick; Male was selected)
Tier: Bench #7
Status: BENCHED — probe-verified only (registry ◐); no production task yet
Role: Bench — Chinese drafting rotation candidate (probe-verified, one notch below ERNIE/GLM on form)
Model(s): Hy3, mode Instant (yuanbao.tencent.com UI, probed 2026-11-05, recorded exactly). ⚠️ The 2026-11-06 drill log records the UI showing Hy4 — see Sources & gaps: the model on record may have changed, which under registry maintenance rule 4 re-opens the entry until re-probed
Account & login door: yuanbao.tencent.com; China mobile +86 131-2279-6726 + SMS — the boss relays live. Login methods offered: WeChat / Phone / QQ. Logged in 2026-10-24
Skills & talents: Chinese writing and reasoning, probe-verified (below); built-in skills only (no surface to add our own)
Probe results: 2026-11-05 (free tier, session already signed in): P5 Chinese writing 4/5 — fully Traditional rewrite, polite and complete, sign-off placeholder present, ~120 chars; one point off for no subject line (same deduction as Doubao) · P4 reasoning 4/5 — 0.05 with full working and verification, all final answers correct; deduction under Quirks. Drills: 2026-10-30 PASS · 2026-11-06 PASS (all 5 parts, Chinese only; UI chrome now largely English)
Loadout: None — no Projects feature. Skills are built-in only; "Group" is a chat folder, not used as a substitute. No web instruction surface (settings are toggles only) — prompt-time group
Quirks & watch-outs: Probe finding (2026-11-05): the P4 trap note contains a muddled clause — "差价只有 1.00 吗？" — under the trap values the difference IS 1.00; the failing condition is the 1.20 total, which Yuanbao does state correctly immediately after. P5 omitted the subject line. No production task yet
Key history: • 2026-10-24 — logged in; renamed to 米先生 • 2026-10-28 — loadout check: no web surface • 2026-11-05 — P5 4/5 / P4 4/5 probes; model pinned as Hy3 • 2026-11-06 — drill PASS; UI shows Hy4 (flagged in Sources & gaps)

↑ Back to directory

Bench · Bench #8

3.8 WorkBuddy (Tencent / CodeBuddy)

Tencent / CodeBuddy

Blocked

Name: WorkBuddy — display name 米先生： in progress (never set; no session has ever existed)
Tier: Bench #8
Status: BLOCKED — app login never completed; nothing about this member is verified
Role: Intended: coding-oriented bench assistant with standing load-sharing authorization — the boss (2026-10-24) authorized Amani to share suitable load to WorkBuddy at her discretion, free tier only, never overloaded. Unusable until first login
Model(s): UNKNOWN
Account & login door: codebuddy.cn · workbuddy.cn — the app at workbuddy.cn/app needs its own login (the codebuddy.cn session does not carry over). Login: China mobile +86 131-2279-6726 + SMS — needs the boss to relay the code live; the app also offers WeChat QR / 企业微信 / SSO / QQ / 邮箱
Skills & talents: None verified. It is a coding-oriented assistant (CodeBuddy line) — that is product positioning, not a verified capability. No session has ever been established
Probe results: None — probes deferred until first app login (probe plan §6). Drills: 2026-10-30 FAIL (confirmed: /app still redirects to the login page) · 2026-11-06 FAIL again, same redirect (check only; no login attempted, per drill rules)
Loadout: ⏳ Pending — loadout waits on the app login; it was the last open item of the 2026-10-28 deployment
Quirks & watch-outs: The first login attempt died 2026-10-28; the restart was deferred by the boss until he can relay the China-mobile SMS code live — do not burn code sends without him. Until login + probe, assign it nothing
Key history: • 2026-10-24 — load-sharing delegation authorized by the boss • 2026-10-28 — first app-login attempt died; deferred by the boss • 2026-10-30 — drill FAIL confirmed (login redirect) • 2026-11-06 — drill FAIL confirmed again; weekend fix with the boss present is the standing path

↑ Back to directory

Bench · Bench #9 — added by the boss 2026-10-28; the **first non-Chinese bench member**

3.9 Microsoft Copilot

Microsoft Copilot

Parked

Name: Microsoft Copilot — account name "C.Y. Chen"; the 米先生 rule does not apply (not a Chinese account)
Tier: Bench #9 — added by the boss 2026-10-28; the first non-Chinese bench member
Status: PARKED — loadout deployed, but the baseline probe was set aside by the boss on 2026-11-05 ("先放著，Copilot 探針之後再補") and a login problem is on record since (see Quirks)
Role: Unassigned until probed; natural candidate for Office/document-adjacent tasks (registry best-use note)
Model(s): UNKNOWN
Account & login door: copilot.microsoft.com (→ copilot.com); Microsoft account cychen818@live.com (saved password; no 2FA hit at the 2026-10-28 loadout). Free/personal tier — the UI shows a Premium pitch ($19.99/mo) which was declined under the no-budget rule
Skills & talents: None verified — no production, drill, or probe task on record
Probe results: None — probe deferred. The probe plan assigned P1/P4/P5 (zero capability data; P1 would check its image route on the free tier), but on 2026-11-05 the probe was back on the password route for cychen818@live.com after device-approval attempts expired or were rejected; the re-saved login was still rejected, and the boss directed the probe be set aside to complete later
Loadout: ✅ Deployed 2026-10-28: Caveman + house rules in Chat settings → Personalization → Customize Copilot, saved + verified verbatim; all three deltas appended, behavior test PASS (2026-11-03). Projects: none with instructions — Notebooks exist but are empty and have no instructions field at creation; a Cowork skills route exists (not used)
Quirks & watch-outs: ⚠️ Login state may have degraded since the clean 2026-10-28 loadout: the 2026-11-05 record shows the saved login for cychen818@live.com being rejected. Verify the door before assigning anything. It is Fujitsu-adjacent by vendor only — the team's Copilot bench seat is the boss's personal Microsoft account; GitHub Copilot is a different thing entirely (Fujitsu-paid, Fujitsu work only, never team projects)
Key history: • 2026-10-28 — added by the boss as bench #9; loadout deployed and verified the same day • 2026-11-03 — delta redeploy PASS • 2026-11-05 — probe deferred by the boss's direction; login rejection on record

↑ Back to directory

Not yet activated · Not yet activated

4.1 Asami Chen — NOT YET ACTIVATED (planned)

Backup Bench · 备份替补

Inactive

Name: Asami Chen (full name confirmed by the boss, 2026-11-02)
Tier: Planned — backup bench member （备份替补）, designated 2026-10-30
Status: INACTIVE — NOT YET ACTIVATED. She has never taken an assignment; everything so far runs through the boss's instruction and Amani's preparation, with no direct exchange on record
Role: Backup bench member: a second Muse account cloned from Amani — the same SOUL and all skills, only the name differs — who steps in when Amani is unavailable and otherwise stands by for work Amani or the boss assigns
Model(s): Muse account (same product as Amani) — no serving record; she has never run
Account & login door: The boss's other Muse account (Facebook login; activation based on his iching6 account). Bootstrap route: the boss pastes the bootstrap pack (`~/workspace/ai-team/asami-bootstrap-pack.md`, prepared 2026-10-30) into that account's Muse; the instance then self-completes setup and reports back. The paste/activation is still pending (as of 2026-11-02). An earlier direct attempt (2026-10-30) was blocked: Amani's browser was stopped from muse.ai by organization policy, so no login was attempted on the boss's behalf; a dedicated isolated browser profile named "Asami" was created for the attempt
Skills & talents: By design she will mirror Amani's SOUL, skills, and work rules — none of it is active yet. Planned own Gmail on Amani's pattern (bootstrap pack §7): full name Asami Chen, birthday 1996-11-09, Female, preferred username asami.dd.chen@gmail.com, recovery iching6@gmail.com — as of 2026-11-02 the boss found the Gmail creation only half-finished and said he would finish and double-check it
Probe results: — (not activated; nothing to probe)
Loadout: Not deployed. The bootstrap pack carries the full setup: identity (name saved as Asami), SOUL verbatim, boss facts, instructions to pull the skill set from the agent-skills-registry and verify the count, work rules, and a completion-report checklist. Amani verifies the completion report item by item when it arrives
Quirks & watch-outs: Device plan (settled 2026-11-02): Amani occupies the iPad and iPhone Muse apps; Asami is to live on the third device — the boss's Moto phone (Android, Xfinity Mobile 408-596-4858). All three devices currently run Amani's account, so the Moto must log Amani out and Asami in (Muse has no quick account switcher — one person per device). Voice calling exists only in the phone/tablet apps (web/Mac offer dictation only), which is why she needs her own device. Communication channel once active: email is the main channel (her completion report goes to amani.dd.chen@gmail.com, CC iching6@gmail.com); Discord is a backup with no existing connector. Bench no-budget rule will apply to her
Key history: • 2026-10-30 — boss ordered Asami activated as backup bench member; browser route blocked by organization policy, route changed to the self-bootstrap pack; boss added the own-Gmail and email/Discord-channel requirements the same day • 2026-11-02 — full name confirmed ("she's Asami Chen"); boss found the instance on the webpage and the half-created Gmail; device plan settled (Moto phone) • Status since: activation still pending the boss's paste of the bootstrap pack

↑ Back to directory

Notes on the data — sources & gaps

Contradictions between sources are flagged here rather than silently resolved; each profile follows the most recent dated source.

Yuanbao's model may have changed (Hy3 → Hy4): registry (probe, 2026-11-05) records Hy3, mode Instant; the 2026-11-06 drill log records Hy4. Under registry rule 4, a model change re-opens verified strengths until re-probed — Yuanbao's P5/P4 scores were earned on Hy3.
ERNIE's rename: roster & SOP is stale (“Baidu 3-day rule”). Current state: retried 2026-10-30, blocked at Baidu identity verification; username still chen20958.
Doubao's freshest state is a drill FAIL, not the registry's ✅: production-verified capability (2026-11-05) stands, but the 2026-11-06 weekly drill FAILED silent. Both are recorded; a recheck is owed.
Copilot's login door may have degraded: clean 2026-10-28 loadout vs. the 2026-11-05 record of the saved login being rejected and the boss setting the probe aside. Profile marks Copilot PARKED on the later record.
Qwen's probe: no Qwen section in the 2026-11-05 probe log; the probe awaits the boss receiving a mainland-phone code. Drills PASS — the member is alive; the scored baseline is missing.
Project counts: Doubao “4 of 4 with instructions” + the 5th 编程与前端 pack project = 5 visible in drills; Qwen 5 total; MiniMax likewise 5 project names.
Model UNKNOWNs that are verified UI facts: GPT (web picker), Claude (web seat), Perplexity (picker labels), DeepSeek (web chat), Kimi (Instant), ERNIE (serving mode). Genuinely unchecked UNKNOWNs: Grok, Qwen, WorkBuddy, Copilot, and NotebookLM's underlying model.
Asami's Gmail is planned, not confirmed: asami.dd.chen@gmail.com is the preferred username; creation was half-finished on 2026-11-02.
Amani's serving model is not in the team sources — recorded as unrecorded rather than assumed.
Champion / Grok Bot is not profiled separately: it is parked with Grok and noted inside Grok's profile — a decision, not an oversight.
Avatars: Amani Chen's avatar is her real profile photo on file; every other member is shown with a monogram placeholder in their status colour — no verified official logo or photo for them was on file, and none was invented.