UKIYO Lady Care · study · from your MANAGER tab

UKIYO Lady Care study: what went wrong, and why the fixes did not hold

MANAGER, 10 Nov 2026. This is a diagnosis and a proposal only. Nothing was built, and no rule, note or skill was created.

Your ask

"For UKIYO LADYCARE, since we have spend so much time, and making so many error and system issue on it, lets map the entire session logs from day 1 start, and all of UKIYO's information in knowledge base ... to get a full study on this. In order to diagnose the real issue of which command, ontology, terms, or what and where went wrong ..." (Full text: STUDY-BRIEF-1110.md, lines 4-10.)

Study written 10 Nov 2026, 16:30 MYT · page built 10 Nov 2026, 16:39 MYT

What to check: read the short answer below, then tell me Go ahead, Change something or Not now on the proposal.

Waiting on you

  1. 1

    The proposal: make your corrections into one table the tools must read

    One ruling table per line, kept on master, which the picture tool, the video fire and the audit read before drawing. Nothing is built yet. Go ahead, Change something, or Not now.

    Read the 7 points

    Nothing below is built. It needs your yes, and the build would be one shared place, not a new rule pile.

    1. One ruling table per line, on master.
      • Each of your corrections becomes one row with:
        • your exact words, the date and where you said it;
        • the check that enforces it;
        • the planted fault it was shown to catch;
        • its status.
      • It sits beside the brand canon on master, not in a worktree or a round folder.
      • Lady Care canon (KB.json, LINE-DNA) moves to master with it.
    2. The tools read it at draw time.
      • The picture tool, the video fire and the audit load the brand and line table before drawing, and refuse on a failed check.
      • The audit reads the design file from the current master copy, never ~/zennith-os.
    3. Proven before trusted. A check counts only after it has been seen failing on a planted fault, then passing on a clean one.
    4. Retired by you, not by drift.
      • When you rule again, the old row is marked "replaced by <your new words, date>" and its check turns off.
      • Nothing is silently deleted.
      • One word list replaces the three conflicting ones.
    5. No copying scripts from round to round.
      • Rounds hold data only: pictures, briefs, plans.
      • Checks and recorders live in shared code. The yes-recorder becomes one shared tool.
    6. What a tab loads at the start.
      • The ruling table's short form (ids plus your exact words, a few thousand tokens), so it survives every summary.
      • The current method lock for each ad type.
      • Nothing else that is not needed. That also eases the 97k fixed load.
    7. The board counts as done only when the live page shows the new file. The check compares the live page with the new file's fingerprint, not a deploy message.

    Order, by how often it came back and what it cost:

    1. Method lock read by the fire (kind 1).
    2. Ruling checks read at draw time (kinds 2 and 3).
    3. The audit reads the current design file (kind 7).
    4. The start-of-tab ruling table (kinds 5 and 6).

The answer, short

  1. The biggest fault was method. The tab used a different way of making the ad from the one you had locked: 86 of 280 corrections. Next came three others:
    • a rule was written down, but nothing read it at draw time (35);
    • a check that should have stopped the fault was missing (34);
    • the tab made content up (25).
  2. Your corrections were mostly not digested.
    • 162 of 280 (58%) left nothing written anywhere.
    • 102 (36%) became a note: memory, handover or KB.
    • Only 16 (6%) became a check in code.
  3. A note did no better than nothing. About half of the faults came back either way: 49% after a note, 54% after nothing.
    • Even the 16 code checks came back half the time (8 of 16).
    • They sat in one round's folder, a side worktree or a scratch file, so the next round never ran them.
  4. The drawing tools read none of the rule files. The audit reads an old copy of the UKIYO design file. On 7 Nov that old copy is why it marked your pink touch down.
  5. Each round copied the last round's scripts forward. That is how the SIMI mask came back. It is also why a fix made in one round stayed in that round.
  6. The tabs had little room to work.
    • About 97k of a 166k window is used up before you type anything.
    • So a tab summarises itself about every 59k tokens.
    • 85 tabs did this 4,450 times. Each time, your exact words can turn into a paraphrase.
Leave a note

The study, section by section

Tap a section to open it.

1. What went wrong, ranked

"Came back" = a later correction points to this one as the same fault again. "Re-makes" = the times an ad was drawn again on a later day just after a correction of this kind. That second figure is a proxy; see REWORK.md.

# What happened Corrections Came back Re-makes after it
1 Wrong method: your locked way treated as optional 86 50 15
2 Written down, but nothing reads it at draw time 35 16 14
3 No check that could stop it 34 22 7
4 The tab made things up (not in any source) 25 9 9
5 Lost when the tab summarised itself or handed over 15 6 -
6 One tab's ruling never reached another tab 15 6 -
7 Read an old copy (KB, design file, stale checkout) 13 11 -
8 One word, two meanings 8 1 6
9 Your ruling kept as a paraphrase, not your words 8 2 4
10 Code reads a path or field that no longer holds the data 3 1 -
- New kinds the readers found (section 5) 37 21 -

The ranking holds whether or not the second kind on each row is counted: the top four stay the same. Old copies (row 7) are rare but nearly always come back (11 of 13).

In your words: one or two per kind

  • Method.
    • "Why not propose first u proceed to generation?" (14 Oct)
    • "So the reels are not made in 1 fire? using either Wan 3.0 or Seedance 2.5 like other approved reels?" (20 Oct)
    • "so u do so long only reskin 3?" (1 Nov)
    • "Did u use jev skill to map the entire prd, skill, prompt, pipelline, workflow, and the reviw and audit before presenting to me?" (16 Oct). You asked "use Jev / as before" about 20 times between 7 and 19 Oct.
  • Written but not read.
    • "are the method used for reels and reskin static correct for social and simi?" (4 Nov). This was the 6th time in two days: 3 Nov ×4, 4 Nov ×2.
    • "And did u also remark the negative prompts that made those that failed previously?" (15 Oct). The tab answered "No, I didn't".
  • No check.
    • "it is still morphed. when the box opens." (7 Nov, the third time)
    • "i think the ref static from behyskin is it taken from reels?" (3 Nov)
  • Made up. "Why no hook is show the bumps first, all alos treatment first." (10 Oct)
  • Lost at handover. "looks ok for me, but lack of pink colour touch. We do have pink touch for ladycare series." (7 Nov). You had ruled the pink on 3 Nov.
  • Between tabs. "ok but social also need to move in bulk, not one by one." (4 Nov)
  • Old copy.
    • "no, we already generated so many in the ads, so this rule is not valid." (3 Nov, the design file's hero rule)
    • "Yes CellLight 3.0 trial is Rm198, it should be in KB." (5 Nov). A memory already said RM198.
  • Two meanings.
    • "4 reels in total is what i meant." (8 Nov)
    • "no simi reels amendment as my notes." (4 Nov). "Dropped" was read two ways.
  • Paraphrase, then a reversal. "Open call, you may use all terms used by client." (10 Nov). This came after three days of the tabs banning words you had never banned.

By period (TIMELINE.md has each day)

  • 7-19 Oct:
    • You restated "use Jev / as before" about 20 times.
    • Handheld camera faults came back 7 times in 9 days.
    • The review page was unusable 7 times.
    • 17 of the 19 notes written were memory notes; there were no code checks.
  • 19-28 Oct:
    • "One fire on Wan" was restated 6 times in 8 days.
    • Two rounds over-corrected to your latest note and broke earlier ones.
    • "No people on fal" was bypassed once.
  • 28 Oct-2 Nov (Lady Care starts):
    • Your first command: "we need to build the DNA for this line, and KB first. brand design all same as UKIYO current one." (28 Oct)
    • The KB and line DNA were written in a side worktree.
    • A tab-written "no anatomy" line hid the references. You lifted it on 29 Oct ("remove that never. for intimate care series we need to show."). It came back twice.
  • 3-4 Nov:
    • The one-fire reel method was restated 7 times across 3 tabs. The method lock lived in handover text only.
    • About 22,000 steps were planned for what you saw as a 4-step job.
  • 4-5 Nov:
    • The board did not show the work, 6 times.
    • The RM99/RM198 price was wrong although a memory held the right one.
  • 6-10 Nov:
    • "Steps, not end results" was corrected 6+ times. Each time only a note was written.
    • The client's own results were never pulled into the copy bank, then reversed.
    • The R14 morph took 4 rounds.
    • That night, 0 of 42 statics made it.
Leave a note

2. Is the learning digested?
What the tab did with your correction Corrections Of those, came back later
Wrote nothing 162 88 (54%)
Wrote a note (memory, handover, KB) 102 50 (49%)
Wrote a check in code 16 8 (50%)

Why a note did not stop it. A note works only if the next tab reads it, believes it over newer text, and remembers it after summarising. All three failed:

  • 328 memory notes match UKIYO.
  • The memory index is cut at load (5 lines lost today).
  • Handovers number 112 files.
  • No drawing tool opens any of them.

Why the code checks did not stop it either. Where the 16 live:

Check Where it lives Does the next round run it?
review_receipt.py and gate scripts (3 rows, cdrama) a round folder no, it came back 3 times
LINE-DNA.json (3 rows) side worktree only; not on master no drawing tool reads it
Board approve page (1) the board it came back once
brands/ukiyo/DESIGN.md hero rule (1) master yes, but the audit reads an OLD copy
fit.py word-fit (1) UKIYO round folder came back 3 times
check_intent.py, mk_intent.py (2) remake-1107 only came back
verify_b2.py (1) a scratch folder lost when the scratch goes
Plan edit, bank, board check script (3) data or round files per round only
Reel-still reference check (1) ad-reference-research exists; a still-only check was not found unverified

So a check that lives in a round folder is the same as a note: the next round starts from a copy, or from nothing.

Leave a note

3. What is wired wrong (file and line, master d8691f27b0)
  • The drawing tool reads no rule file. codex-image-gen gen.py checks only that the brief is approved (its run-gate). It reads no UKIYO or Lady Care rule.
    • reskin-autopilot reads a compiled pack that holds 4 Lady Care rules out of 45 UKIYO rules.
    • The Lady Care canon (KB.json, LINE-DNA) is read by no engine script (RULE-STORES.md).
  • The audit reads an old design file.
    • god-audit.sh lines 826 and 1315 open ~/zennith-os/brands/ukiyo/DESIGN.md first. That is a checkout 4,371 commits behind, with a different version of the file.
    • 292 audit runs loaded it.
    • Neither copy has your pink touch, so on 7 Nov pink was marked as a second accent.
    • Filed as #5736 (the same fault hit BASIL).
  • Three word lists disagree.
    • fit.py:11 and scorer.py:54 ban one set of words; mgr_check_s4_copy.py:20-21 bans another.
    • The brand pack (pack.yaml:73) bans 紧致, 修复, 精雕, 激活 and 异味, which are the client's own words.
    • The newest checker (fit_1110.py) reads no word list at all.
    • So your 10 Nov ruling ("all terms used by client") is enforced nowhere, and the old bans still sit in live scripts.
  • The copy bank files the client's end results under the wrong label.
    • build_banks_1110.py:116 types them as "mechanism" (90 of 103).
    • The objective is fixed as "awareness" for every line (:76-78).
    • The "steps vs end results" rule has no check in code; scorer.py:12 mentions it only in a comment.
  • A missing file. pack.yaml:29 points at a facts.json that does not exist.
  • Round-local tools.
    • The yes-recorder that puts your approval on record lives in one round folder (record_brief_yes_1109.py). Each new brand copies it; AIREI did today.
    • The SIMI mask came the same way: PLAN-1108, then t1-t4, then 1110. As you said, MIRRA's slimming pictures were made without one. I have not checked the MIRRA files myself.
Leave a note

4. Words that meant two things (TERMS.md)
  • "steps" had 3 meanings: a pipeline step, a treatment step, and "not a selling point". You used it 22 times. Your ruling was the third meaning, and no check holds it.
  • "S1-S5" were service names, and also S3 storage, shot names, row ids and file names. LC-41 was mislabelled S7, then S9.
  • "skeleton" had 4 meanings, and "approved" had 4 forms:
    • your Use mark;
    • your typed OK on a plan;
    • an /approvals click;
    • a Jev pick. Only one of these opens the drawing gate today.
  • "reskin" meant different things for reels (one fire), statics (edit the skeleton) and promos.
  • The KB never held the working words. 步骤, USP, reskin, skeleton, approved, hook and headline appear in no KB version.
Leave a note

5. New kinds the readers found (not on the starting list)
  • Picture-quality faults with no check (10): a morph, a flash, a seam.
  • The review page could not be used, the work could not be found, or a deploy was never checked live (14).
  • Too slow (3).
  • Over-correcting to your latest note and breaking an earlier one (3).
  • The tab called a skill that does not exist ("Unknown skill: deepen", 3 times).
  • The tab could not view its own output before showing you.
Leave a note

6. My own faults as MANAGER (10 Nov)
  • My 1110d table offered rows that were steps, not end results (S4-127, 129, 143, 145, 147). That is the same fault I was meant to catch.
  • I agreed to SIMI's flatten and mask (1110b, 1110c) instead of asking why a mask was needed at all. Your 16:00 question caught it.
Leave a note

8. What these numbers are measured against
  • Session logs: this Mac's only. 111 tabs, 1 Oct to 10 Nov. Other Macs are not included.
  • Your turns: 909, tagged by six readers.
    • 280 correction rows. A ruling that reversed an earlier one also counts as a row, so this is more than the 243 turns tagged as corrections.
    • Readers tagged each row by hand. 217 rows carry a second kind.
  • Code: read on the runtime copy of master, d8691f27b0 (0 commits behind origin/master at 16:27). The readers read 41ce8e3c05, which was also current when they ran.
  • Jev: not called. The ranking was checked in code with and without the second kind, and the top four did not change (disposition below).
  • Limits:
    • The "came back" links are the readers' judgement, within each reader's slice.
    • The re-make cause is a proxy.
    • Spend is not measured.
    • R6's progress times were made up by that reader; its rows are used, its times are not.
Leave a note

More details

The supporting findings the study is built from.

Day by day (TIMELINE.md)

TIMELINE: UKIYO, 7 Oct to 10 Nov 2026 (this Mac's session logs)

Built by tools/build_docs.py: the six reader timelines, joined in date order, not edited. Times are UTC unless a reader says MYT. Her clock is UTC+8. Caveat: R6's progress.txt carries made-up times (16:25 to 18:35, written before 15:53). Its rows match its files; its times are not used.

R1 timeline: UKIYO before Lady Care, S1-001..S1-212 (7-19 Oct 2026, UTC stamps)

Referent: pack rows S1-p1 + S1-p2 (this Mac's main-project logs only, 3 busy tabs: 9a8ececa social/calendar, 055a5d01 CellLight add-on reels, 6c21af41 new-ads batch; plus 92003618 Sept report, b9368d80 stunt reel). Counts are from my own turns-tag.csv / ledger.csv. Limit: rule_writes in the pack lists only memory/rule/KB/canon writes. A code check written elsewhere is NOT visible, so every STATED/NONE below means "no check visible in the pack", not "proven none". No CHECK rows found.

By day

  • 7-8 Oct (001-002): wants 5+ ugly lofi ads; reel too slow, not from reference deep decode, "we have so many succcesful approved reels". Memory note written.
  • 9 Oct (003-014): maps approved one-fire reels with Jev. Social posts 2-4 plates approved (009). CellLight add-on to reels 08/09/11; she asks for VO/subtitle/end result and a PAS type (012-013), approves voice (014). Asked again for IG/XHS refs she thought were already fetched (005).
  • 10 Oct (015-033): Post 3/4 look notes. VO must start at 0 and give end result in 3 s (017). Hook not "treatment first" (023), BGM missing/weak (024-025). Six ads approved on chain (029). New ad research ordered; "not only reuse" (032).
  • 11-12 Oct (034-062): waits on XHS refs. Handover written for a 30-ad batch (040). Fresh tab asks her 13 questions; she answers, "dont be lazy" (045). Back-body keyframes redone; "Dots still not fver" (046). RM198 offer misread as MSilky (060).
  • 13 Oct (063-087): DPL/milk-light machine questions; "real pain point, not relax USP" (067). Purple manufacturer head rejected (071). "Very not user friendly" review page (072). CL04 reel approved (086).
  • 14 Oct (088-105): "Can we not keep do 1 thing then wait" (088). Pain openers kept mild (092). Outputs "very weak" (093). "Why not propose first" (103). Repeated skeletons, not all her refs in (104). Calendar 1 post/day not met (102).
  • 15 Oct (106-127): Batch 1 "almost all rejected" (119): wrong handheld, raw bg, staged angles, pale, lying down. Viral reels copied by topic not mechanic (120). She asks if negative prompts of failures were recorded: no (121). Still not accurate at 127.
  • 16 Oct (128-149): clothes morphed (131). Her R7 comment lost (132), comments not actioned (140). Statics "all look the same" (134). Sept report ordered (136). Calendar made draggable (149). Commit/push/merge phases ordered.
  • 17-18 Oct (150-193): "Did u use jev skill...before presenting" (157, 158). "show only the one to review" (159). Method failure list (171): handheld wrong, back of head, text and image separate; she asks "why is the issue coming back?". CL-22 locked as correct prompt (176). "i dont want elevenlab" (182). NG-02 method approved (187). Stick on flying machine (193).
  • 19 Oct (194-212): English version of flying reel (199). Device preview vs brand mood (207). Pilot: white mood, no CellLight word, 3 hands (210). Overnight autonomy order, "Bypass all permission" (212).

Counts (my files)

Rows 212. By type: correction 54, question 43, chat 42, command 38, ruling 22, approval 13. Ledger rows 57. By class: 7=12, NEW:visual-quality-defect=10, 6=8, NEW:review-surface-unusable=7, 2=5, 8=4, 3=3, 5=3, 10=3, 1=1, 9=1. Fix recorded: STATED 19, NONE 38, CHECK 0 (see limit above).

Patterns that Lady Care later repeats

  1. Her locked method treated as optional / re-asked each time (class 7, 3). She says "use jev skill" or "as what we done before" in 002, 003, 030, 044, 053, 058, 062, 108, 128, 137, 145, 157, 158, 160, 164, 168, 176, 200, 212. Ledger: 002, 003, 103, 108, 120, 151, 157, 158, 182. A rule she must restate ~20 times was never a gate.
  2. Reskin vs making a new picture. Handheld machine wrong: 071, 119, 127, 171, 206, 207, 210 (same fault 7 times in 9 days, only memory notes in between: 071 memory feedback_ukiyo_keyframe_hand_direction). Text and image generated apart again 008, 171 ("This is a simple task that we do all the time in multiple brands. why the issue is comming back?"). Repeated skeletons 104, 106, 134.
  3. Approvals and the board: review page unusable 072, 090, 107, 135, 159, 173, 205 (7 rows, same complaint, each fixed in the page only). Comments not actioned or lost 129, 132, 140, 057. Her "approved" read as GO for bulk when she meant the plan (103, 151).
  4. Rules written to memory but never checked in code: 17 of 19 STATED rows are memory notes (e.g. 045 feedback_check_workspace_and_logins_before_asking; 120 feedback_viral_reskin_copy_mechanic_and_flow_1to1 then 127 again). 121: she asks whether failed prompts were recorded; answer "No, I didn't".
  5. Context lost at compaction/handover/limit: 005 refs re-asked, 045 questions already answered in workspace (a handover wrote 13 questions for decisions she had given), 060 RM198 offer misread despite 039 ruling ("CellLight 3.0 first trial change to RM198"), 092 two rulings collided (067 real pain point vs 13 Oct mild-skin rule), 154/155/177/178/184/196 resumes after account limits, and 035 a password written into a log (secret rule memory only). Bare "Resume" turns (24 of 212) show limit stops cost turns.
  6. Pacing: one thing then wait (088, 211), generating before proposing (103), "whats taking so long" (073, 098, 156, 167, 202).

NEW classes

  • NEW:visual-quality-defect (10 rows: 008, 015, 025, 054, 093, 131, 174, 176, 193, 210): the output itself is bad (AI face, morphed clothes, stick, low contrast, pale). Not a rule gap; no fire-time check sees it. Example 131 "client clothes morphed", 193 "stick coming out".
  • NEW:review-surface-unusable (7 rows: 072, 090, 107, 135, 159, 173, 205): the board/page she reviews on hides the item (buried among cards, no playable video, local file link). Examples: 159 "show me the one need to review instead of scroll all", 205 "link unclickable". Possible third (single row, low): reviews of her board comments lost (132) is class 8 + 10; kept there.

Not readable / caveats

  • Long her turns are cut in a few rows (e.g. 030, 040, 045, 119, 127, 137); full text is in corpus/turns.jsonl, not pulled. Verbatim in ledger.csv carries the cut marker where it was cut.
  • 5 corrections have second_class only; no Jev scoring done (not allowed).

R2 timeline: S1-213..S1-371, 19 Oct 22:48 UTC to 28 Oct 04:08 UTC (UKIYO before Lady Care)

Referent: pack rows S1-p3 + S1-p4 (this Mac, main-project sessions only). Counts are of these 159 rows.

By day (UTC)

  • 19-20 Oct: statics round 8 done (18 on board). NG-01 handheld "too small, thin, angle wrong"; she asks if reels are one-fire (S1-217): agent admits shot-by-shot, 4-23 Wan fires. She locks "mostly Wan 3.0". 20 Oct: double nose in NG-01 missed by Jev (S1-223). Social tab starts c-drama skit: script rejected twice (S1-231, 232), then "wait, plan first" (S1-236), "decoded fully?" (S1-237).
  • 20-21 Oct: 30 board comments (10 approve, 18 change). She asks "exactly ref 1:1? get sued" (S1-249/250). Fresh-composed stills gave "AI angle" (S1-254/255); she confirms stills from the reference's own frame (S1-256). Statics/ads: wrap + Meta upload, chat-builder prefill set to Chinese only (S1-279), 21 ads approved (S1-280), schedule 12AM (S1-282).
  • 21-22 Oct: skit continuity faults (mask back on, her position jumps, uniform skirt/pants, bed) S1-258, 273, 275, 284, 286, 308. Sunburst / Jev / production-house review asked (S1-303, 304). Jenn rule on people images on fal: she says "bypass Jenn rule" (S1-307), later Codex used (S1-313).
  • 23 Oct: 黑金 removed from all ads (S1-316). "shortfilm skill should include all this, why still happening" (S1-320); gate coded (review_receipt.py). Push gate red for unrelated reasons, bypass with reason (S1-330). Manager role set up (S1-341). Treatment video blocks still wrong (S1-344).
  • 24 Oct: blur over-applied then "fake" (S1-346, 348); skin noise (S1-349); "dont be lazy" on Higgsfield API (S1-347). Memory index missing a pointer (S1-350).
  • 25-28 Oct: Disney: "why ask million times" (S1-352). Social reskin: s02 keyframe rejected, not one-fire reskin-autopilot (S1-367, 369); "i dont see any changes" while waiting on MANAGER GO (S1-370/371). She approved: NG-02 (v8), 21 ads 21 Oct, CL-16, test fires "ok lets test" (S1-252, 274, 281, 284, 298, 350, 360).

Counts (turns-tag.csv)

159 rows: correction 46, question 43, command 35, ruling 15, approval 14, chat 6. Ledger 50 rows (includes 4 reversal/restatement rows).

Patterns (row ids)

  1. Locked method treated as optional / lost: one-fire Wan (S1-217, 225, 254, 255, 367, 369 six times over 8 days); "use shortfilm skill + Jev" (S1-304, 320, 338); video blocks of treatment (S1-214, 297, 344, three asks); 1:1 vs blend (S1-249, 250, 267, 292).
  2. Reskin = edit the reference vs drawing new: stills composed fresh gave AI angle (S1-254, 255, 256); s02 keyframe too polished, 2 ppl differ from reel scene (S1-367, 369). Rule confirmed S1-256 ("memory, lane rules updated") but no write visible in pack, and it recurred 27 Oct.
  3. Continuity faults repeat shot after shot: mask on/off (S1-273, 275, 284), her position (S1-258, 308, 315), uniform (S1-286). Fixed by code gate only late (S1-339 positions line check; S1-320 review_receipt.py). 3 CHECK rows total out of 50.
  4. Overcorrection from one note: close-ups (S1-264 to 265), blur (S1-335 to 346 to 348).
  5. Stored not consumed: memory written (28 of 50 rows STATED or NONE with memory only) yet same error returns: face gate (S1-223) then skin noise (S1-349).
  6. Context loss / tab coordination: she pastes notes between tabs (S1-319), pastes agent text back in (S1-257, 260, 310), memory index lost a pointer (S1-350), Disney ruling re-asked (S1-352), UKIYO waiting on MANAGER GO unseen by her (S1-370/371).
  7. Gate bypass becoming routine: people-image fal rule bypass (S1-307, reusing a 21 Oct template); zen-ci bypass with reason (S1-330).
  8. Board/approval: duplicate card (S1-272), links not accessible (S1-291), jargon "0.07" to her (S1-276), board filter for unapproved (S1-342).

NEW classes

  • NEW:overcorrection_from_one_note: one note over-applied. S1-265 (all close-ups), S1-346 (even blur), also ties to S1-348.
  • NEW:absence_claim_without_search: S1-347 agent said no Higgsfield image route without searching ("you dont be lazy"; agent: "I stopped too early").
  • NEW:taste_bar_miss: S1-245 male lead not good-looking enough (also S1-246); no check exists for this.

Unreadable / gaps

Pack text is head/tail only; where_recorded for rows relies on pack rule_writes. Memory notes from MEMORY index (S1-348, 349, 352) are not in the pack rule_writes: confidence low-med. Rows S1-302, 351, 353, 356, 362, 371 have empty replies in the pack.

R3 timeline: S2-p1 (S2-001..S2-115, 28 Oct to 2 Nov 2026, this Mac, local logs)

1. Her FIRST Lady Care command (verbatim)

Row S2-019, session 60ab9400 (full id in corpus), ts 2026-10-28T13:35:23.503Z (UTC, as in corpus/turns.jsonl). Full text, 2444 chars, not cut:

<pasted_content id="fc98">
这边给你介绍 🔎
Hifu Vaginal Tightening 私密处超声刀

私密处Hifu技术是一种安全有效的治疗方法 可以改善私密处的外观和功能 🪬 如果Dear正在考虑进行私密处整形手术 Hifu技术可以作为选择对象 ✨

☁️ 私密处Hifu技术的优点 
 - 非手术
 - 无需麻醉
 - 恢复时间短
 - 效果持久

☁️ 可改善
 - 松弛的阴唇
 - 阴道松弛
 - 阴道分泌物过多
 - 阴道干涩
 - 性交疼痛

以下人群不适合进行Hifu技术
 * 怀孕或哺乳期的女性
 * 近期有过阴道炎、宫颈炎等妇科炎症的女性
 * 有阴道或阴道周围肿瘤的女性
 * 有心脏病、高血压、糖尿病等慢性病的女性
</pasted_content id="fc98">

<pasted_content id="fc98">
🔍 Thinking About Feminine Rejuvenation? Try HIFU Vaginal Tightening!

Looking for a safe, non-surgical way to refresh and tighten your intimate area? ✨
HIFU (High-Intensity Focused Ultrasound) is a gentle yet powerful treatment that helps improve both appearance and comfort — no surgery needed! 🪬

🌸 Why Women Love HIFU:

No surgery, no needles

No anesthesia required

Little to no downtime

Long-lasting results

🌸 It Can Help With:

Loose or sagging labia

Vaginal laxity after childbirth

Excessive discharge

Vaginal dryness

Pain during intimacy

🚫 Who Should Avoid HIFU:

Pregnant or breastfeeding women

Women with recent vaginal or cervical infections

Anyone with tumors in or around the vaginal area

Women with chronic health issues like heart disease, high blood pressure, or diabetes

✨ Feel confident, comfortable, and in control — naturally.
</pasted_content id="fc98">

<pasted_content id="fc98">
*私密修复黏膜护理*
私密黏膜是什么：
-私密黏膜是内壁最柔软的组织 提供水分和养分 具有自我清洁 抵御细菌 维持酸碱平衡的重要作用

*黏膜受损常见问题*
-干涩紧绷
-敏感瘙痒
-白带异常/有异味/红肿
-容易复发感染

*黏膜受损的危害*
-妇科炎症
-宫寒
-宫颈糜烂
-不孕不育
-流产
-HPV
-性病
-宫颈癌

*黏膜的护理过程*
-先会清理毛发⚠️如果有的情况
-会用阴道扩张器检查阴道的情况
-用生理盐水来清洁
-放自家的产品进去阴道 （主要的作用细菌 消炎 让阴道黏膜回到健康的ph）
-外阴按摩

*原价一次RM599*
</pasted_content id="fc98">

<pasted_content id="fc98">
*私密紧致护理-徒手*
主要是修复我们阴道里面的肉芽让达到紧致水润的状态

❌无刀无💉无痛 无副作用 可以安心护理

*适合大多数女性*🌷
-产后
-干涩
-松弛
-性冷感

*功效*
-紧致
-饱满
-湿润
-提高性生活

*过程*
-清洁阴道
-放入我们的产品（主要是激活肉芽组织自我修复 从根源改善松弛）
-做阴道精雕按摩（让产品更好吸收 更加紧致饱满）
-外阴按摩
-敷私密膜（美白 保湿 紧致）

*原价RM899*
</pasted_content id="fc98">

<pasted_content id="fc98">
*私密SPA护理-徒手按摩（也成为激活海底轮护理）*

*主要功效*
-激活海底轮
-穴位按摩
-紧致
-激活性腺

*做的过程*
-如果有毛发需要处理掉
-做一个深层清洁
-按摩激活穴位30分钟左右
-敷私密膜

*原价一次RM499*
</pasted_content id="fc98">

私密spa-原价499 体验价格299
私密黏膜- 原价599   399
紧致徒手-原价899 699

WhatsApp Image 2026-10-16 at 14.50.33 in downloads

This is the new line service UKIYO wanna run ads for, these are all the knowledge base we have and information. u can also scrape in ukiyo website to see if any information on this. 

we need to build the DNA for this line, and KB first. brand design all same as UKIYO current one.

(Rows S2-001..S2-018 are other UKIYO tabs: SOCIAL, AD HOC, MANAGER. S2-021 is the first MANAGER mention: "ukiyo intimate care need to get it ready onboard for ads production ready".)

2. How the agent first turned the KB into working files (60ab9400, Write/Edit in corpus/actions.jsonl)

Measured against corpus/actions.jsonl, session 60ab9400, 28 Oct 14:18-14:44 UTC. Worktree: ~/zennith-work/wt-ukiyo-lady-care-1028

  • 14:18 intel/ukiyo-lady-care-2026-10-28/SOURCES.md (Write, then Edit)
  • 14:19 brands/ukiyo/lady-care/KB.md (Write; head: "Why this exists ... which of those words may go in an ad")
  • 14:24 brands/ukiyo/lady-care/LINE-DNA.md and LINE-DNA.json (Write; "inherits all of UKIYO and overrides nothing")
  • 14:44 brands/ukiyo/lady-care/READINESS.md
  • 14:45 memory project_ukiyo_lady_care_line_2026_10_28.md (Bash heredoc, not the Write tool)
  • 29 Oct 00:46 intel/.../refs/REFS.md; 29 Oct 00:51-03:17 memory feedback_new_line_skeletons_from_its_own_category.md This is where the first terms and rules were set: KB (client words and which may go in an ad), LINE-DNA (brand rules, incl. L4 "no anatomy / no below-waist body"). FINDING: the pack shows rule_writes=0 and other_writes=0 for S2-019 although the KB files above were written in that turn. The pack misses writes under the worktree path and every Bash heredoc/python write. I used actions.jsonl instead.

3. Her early rulings (first 5 days) that I can see written down

  • 29 Oct "remove that never. for intimate care series we need to show" (S2-033): memory append 08:07 (Bash), LINE-DNA.md + LINE-DNA.json edited 08:48-08:49 (Bash python, her words quoted). Later LINE-DNA L4 reads "Drawn or covered only ... Lifted by her 29 Oct".
  • 30 Oct "can add some pink element inside, only for this intimate care series ..." (S2-043): LINE-DNA.md/.json edited 04:22-04:23 with her words as a quote; Jev RECEIPT.md 04:32; PORTRAYAL-DNA.md 04:24. Also test5/_shared/RULES-LADYCARE.json written 30 Oct 08:51 (Write).
  • 29 Oct PORTRAYAL-DNA.md first written 14:17 (after S2-040 "the ones u present is not on point").
  • 2 Nov after S2-092: KB.md 05:35, LINE-DNA 06:18, HANDOVER-2026-11-02-LADYCARE.md 06:19 (Bash).
  • NOT visible written: 28 Oct "20 ads, statics" (S2-023), 30 Oct "20 ads, skeletons from whole index" (S2-061), 30 Oct "dont jump too freelance and patchwork" (S2-058), 31 Oct pink leaf not on all (S2-068), 2 Nov "no need all woman" (S2-097). I found no write for these in the windows I checked; they may be in files I did not search.
  • Engineered? Only LINE-DNA.json (and RULES-LADYCARE.json) might be read at fire time. I did not verify any reader. Rows are marked CHECK, not proven.

4. By day

  • 28 Oct: Social tab: board shows no change (S2-002), keyframes wrong (S2-003/005), machine no label, room cheap (S2-012). Lady Care starts 13:35 with KB paste; skeleton check 23:57 "your skeleton shows facial" (S2-022).
  • 29 Oct: needs 20 ads (S2-023). Test 5 posts (S2-025). She sees refs show illustrated body part; lifts "never" (S2-032/033). Rates not 9 (S2-039). "ones u present is not on point" (S2-040).
  • 30 Oct: new client scripts (S2-041). Pink + CellLight illustration style allowed (S2-043). Approves LC-01, LC-07 (S2-044), LC-10, LC-02 (S2-046), "ok lets try" LC-08/09/11 (S2-048). Asks why no organ shape (S2-054), why not reskin high-impression ads (S2-057), no patchwork, lay all skeletons first (S2-058), 20 ads (S2-061). Board messy (S2-065).
  • 31 Oct: LC-15/LC-12 product shots, radiant face, test reskin autopilot (S2-068/069). Social: Halloween, Deepavali; "so long only done 2 post?" (S2-079).
  • 1 Nov: "so u do so long only reskin 3?" (S2-082). Asks if Jev picked the method (S2-083). No card finished in two days (agent's reply, S2-082).
  • 2 Nov: LC-15 blocked by disk 7.6 GB (S2-085). Re-sends prices, asks if KB stored (S2-087). Asks why tightening not a PAS point (S2-088), no BOFU/promo (S2-091), towel/water not Lady Care (S2-092). Social: rejects 8 Nov posts (S2-094). Skit: patchwork (S2-096). "No need all woman" (S2-097). Board unclear (S2-100). Orders reskin all statics (S2-108). Overnight 8.5+ target (S2-115).

5. Patterns

  1. A rule the agent wrote hid what the references show. L4 "no anatomy" was in the 28 Oct LINE-DNA; she corrected it 29 Oct (S2-032/033), again 30 Oct for pink and style (S2-043), and the same gap came back as "why cannot reskin like the ref" (S2-054), "towel and water dont relate to ladycare" (S2-092). Class 3/6. Each lift was a text edit in LINE-DNA; no check seen that compares output with the ref's portrayal.
  2. "Patchwork / freelance / operator heavy": S2-058, S2-092, S2-096, S2-083, S2-094. Her fix each time: lay the whole plan first, use Jev, use the locked method. Class 7.
  3. Delivered but she cannot find or see it: S2-002, S2-013, S2-065, S2-100. Four times: live check passed or card exists, but her view differs. NEW class.
  4. Too slow: S2-045, S2-053, S2-079, S2-082 (14 rounds on LC-15, 2 days no finished card). NEW class; part is disk (S2-085) and picture engine returning nothing (S2-053).
  5. Picked wrong refs for the line: S2-022 (facial skeletons), S2-040 (not on point), S2-057 (not top ads), S2-094 (social, 7 of 8 posts wrong ref per diagnosis agent). Class 7.
  6. KB questions: she re-sent prices on 2 Nov (S2-087) though they were in the 28 Oct paste (S2-019); she asked for a Jev onboarding review twice (S2-089/090).
  7. Jev named repeatedly by her as the method picker (S2-044, 058, 083, 086, 091, 092, 096, 104, 106, 108, 113-115): the ask recurs because method choice was not engineered.

6. NEW classes

  • NEW:delivered-but-not-findable: S2-013 (card 76th at bottom), S2-065 (messy board), S2-100 (too much info). Also S2-002 (check passed, she saw nothing). Check checks the file, not her view.
  • NEW:throughput-too-slow: S2-045, S2-053, S2-079. Cost, not a fault in content; causes seen: engine empty draws, disk floor, check loop.
  • NEW:infra-guard-blocks-ask: S2-017 (disk guard + hook blocked a merge she asked for).
  • NEW:ask-repeated (S2-090) and NEW:overcorrection-to-latest-words (S2-097): both low confidence.

7. Limits

  • Pack rule_writes is incomplete (see section 2). Where_recorded for rows after S2-040 comes from actions.jsonl keyword search, ties to a given ask are by time only.
  • Reply text in the pack is short; "agent_had_done" is "not visible" where I could not see it.
  • Local Mac only. Amend rows S2-032-amend and S2-033-amend are appended; originals kept.

R4 timeline: pack S3-p1, rows S3-001..S3-135 (UTC stamps; MYT = +8h). Referent: slices/packs/S3-p1.jsonl only.

Tabs seen: UKIYO SIMI statics (ec9ff1dc, then 3ddfe4cd, then fec86d6b), UKIYO SIMI REELS (e5a30782), UKIYO SOCIAL (e296ae26, then e23b8a27), UKIYO SKIT (14820dea), MANAGER (69ab6b3a, then 49f4bb30), ADS (97c8cf62).

By day

  • 3 Nov (UTC 00:07-13:52), rows 001-071.
    • Asked: skit reel (Wan 3.0 pasted prompt), statics review rounds, reels 14/17/22 reskin, Nov calendar, services/promo recap for Lady Care ads (S3-028).
    • Wrong: statics "not reskin", close-ups only, refs only from Behyskin (S3-002..004); reels plan not one-fire (S3-021, 032, 034); Brazilian light line missing (S3-026); skit take had no sound, flicker, blurred word (S3-024); Behyskin "static" refs were reel stills (S3-057); god-audit penalised AI hero from an old DESIGN.md (S3-065).
    • Approved: skit v6b (S3-043, final S3-049). Ruled: AI hero is valid (S3-065), do not block uterus drawings (S3-066), 9% weekly limit so only SIMI runs (S3-070).
  • 3 Nov late to 4 Nov (UTC 22:54-07:59), rows 072-135.
    • Asked: include held ads (S3-072), promo price only on LC-08 (S3-081), r22 light illustration + clean machine (S3-085), review every skeleton ref before making (S3-099), reels LC-27/LC-28 test, bulk image runs (S3-121, 127).
    • Wrong: held ads left out; r14 bg morph; r22 no sound; own-built social board with filters broken (S3-088); reels plan 58 keyframes / 7-8 fires (S3-111, 115); r14 got a second spliced fire (S3-124); social still one picture at a time (S3-127); "dropped" misread (S3-135).
    • Approved: reels go (S3-105), r14 fire and LC-28 test "Go, no need to wait for me" (S3-128), review reel go (S3-134). Ruled: RM198 with no 黑金 word (S3-132), first scene no client lying down, only LC-27 and LC-28 first (S3-131).
  • About 20 rows are MANAGER/infra turns about disk, Mac mini, Owen's sheet, Pinxin/POSH (S3-080, 084, 086-094, 096, 100-103, 107, 108, 110). They are not UKIYO. S3-089 holds a pasted login; not copied.

Patterns

  1. Reel method (one Wan 3.0 fire, true reskin of the ref reel) came back at least 7 times, in 3 tabs: S3-021, S3-032, S3-034, S3-044, S3-111, S3-115, S3-124. The rule was written in the social handover METHOD LOCK (HANDOVER-2026-10-27.md L8-26, from S3-036 and S3-115 briefs) and in HANDOVER-1102-REELS.md L18 and L41 (S3-115 brief). A text rule, no check. The one fire is also limited by the tool (driver 15 s, driver+output 30 s, S3-116); she asked for "as long as Wan 3.0 can" (S3-032), a clash the agent had not told her about until S3-032 and S3-116.
  2. Reference handling: wrong ref type or source came back 5 times: S3-003, S3-004, S3-051, S3-057, S3-058; then she repeated the demand after the fix (S3-099, S3-106). A check was built (S3-062 reply says all brands, 9 test groups; S3-082 reply says folded into ad-reference-research). The file path is unverified here, and S3-101 reply says her "use this" marks sit inside the pick page and the tool cannot read them (dead consumer, class 4, agent's finding).
  3. Stored lessons not applied: agent's own answer S3-012 (about 90 UKIYO lesson notes hand-picked per round); service split decided but only LC-10 carried it, other 15 SPECs old (S3-026, S3-042); MANAGER sent GO on a plan breaking the written one-fire rule, found only when she asked (S3-113 -> S3-115).
  4. Cross-tab (class 10): SIMI REELS (e5a30782) did not know the social tab's METHOD LOCK (S3-021, 032, 034 vs S3-015, 036). SIMI was told "4 at once" (S3-121) and social was not (S3-127, same ruling, 3 rows apart). The SIMI statics tab (fec86d6b) re-fired the social N1/N3/N5/N7 statics after the social tab's reel-frame refs (S3-055, 072). S3-060 (MANAGER) says two tabs worked the same line and each paid the start load.
  5. Old rules read as live: god-audit DESIGN.md "never AI-generate a hero" (S3-064/065, retired 3 Nov by editing brands/ukiyo/DESIGN.md in the working tree; a later memory says god-audit reads a stale copy); one-at-a-time lock from 27 Oct (S3-121).
  6. No sound, flicker, bg morph, blurred word shipped before she saw them: S3-024 (skit), S3-082 (r14), S3-083 (r22 sound). Nothing in the pack shows a check for these.
  7. Review surface: no approve/reject buttons (S3-053), too many tabs (S3-104), own board with broken filters (S3-088). Also she keeps asking for a standard proposal board "following master and system" (S3-083, S3-085, S3-088).
  8. Cost: UKIYO ran about 22,000 steps instead of 4 (S3-052, agent measure from session logs); weekly limit 9% left (S3-070); token guard armed (reference_token_guard_local_hook.md, S3-070 rule_writes: code/guard).

Fixes that became code or a gate in this window (CHECK, unverified by me)

  • S3-053/054: approve page test, proven red by removing the Reject guard (S3-054 reply).
  • S3-058/062: reference-intake check, 9 test groups (S3-062 reply).
  • S3-065: brands/ukiyo/DESIGN.md hero rule retired (text file read by the audit).
  • S3-070: token guard hook (memory note names it). Everything else in the ledger is STATED or NONE.

NEW classes

  • NEW:over-engineered workflow (S3-052). She says the job is plan, hook, skeleton ref, reskin; agent measured about 22,000 steps. Also S3-111 (58 keyframes), S3-115 (7-8 fires per reel).
  • NEW:board review surface (S3-104). Items she must review are not findable or not actionable. Same family: S3-053 (no approve/reject), S3-088 (filters broken, too many posts), S3-083 (not side by side).
  • Wording clash (class 1): "dropped" read as "removed" in S3-134/135, while the same word means "posted" in S3-133 ("Done dropped my review"). Also "dropped" in her S3-134 note line.

R5 timeline: pack S3-p2, rows S3-136..S3-245 (4 Nov 08:00 to 5 Nov 14:04 UTC)

Referent: slices/packs/S3-p2.jsonl (110 rows, this Mac). Times are UTC from the pack; her clock is UTC+8. Tabs: MANAGER 49f4bb30, SOCIAL e23b8a27, SIMI fec86d6b, EDIT d6ea670d, one MIXED row 1b2cdb53 (S3-199).

By day

  • 4 Nov (S3-136..S3-198): MANAGER cuts LC-28/27 to 30 s in 2 fires (S3-136). Board not updated, link not opening (S3-137/138/145). SIMI LC-28 part 1 not reskinned, real staff face used (S3-146..148). She says "taking too long", board first (S3-153). Headlines rejected: "relax, charge", wants pain-to-face (S3-160, 163, 166). R17/REVIEWREEL not amended (S3-156). She approves LC-28 part 2 (S3-168). R22 rebuilt: VO, machine first, PIP (S3-172, 173, 183). LC-28 must be 9:16, VO mismatch (S3-175, 190, 191). Tabs hit the 150k token guard from S3-184 on. R22 script v2/v3: no end result, machine-buyer wording (S3-193..197). She approves script v3 (S3-197).
  • 5 Nov (S3-199..S3-245): ads-system audit brief pasted to MANAGER (S3-199). Headline edit 28 of 29 done (S3-201); N3/N5 morphed (S3-203). ACCA pictures regen (S3-209, 217). SIMI LC-28 not in UKIYO uniform/font (S3-210). She approves LC-28 (S3-216) and R22 (S3-227). Walnut overdone (S3-230). LC-27 paid Wan fires approved (S3-231, 240). RM99 vs RM198 (S3-233, 236). She asks to view reels first (S3-241). NEW1/2/4 direct reskin answers (S3-243).

RC7-DIAGNOSIS-1105.md (headings read only)

Headings: 1 what rc=7 / model_not_pinned is; 2 what the lane passed for "model"; 3 gap behind "headline cards not on the spine"; 4 secondary fleet finding; 5 Jev consult; 6 proposed diff; 7 what is left.

  • No row in ledger.csv is a correction of hers whose cause it names. It names a spine/gate fault, not anything she corrected.
  • Rows that touch its subject (not ledger rows): S3-174 error "LC-28-K9d c30 rc=7"; S3-187 reply "Code 7 is harmless" plus grep of model_not_pinned; S3-194 reply "ACCA isn't on the spine yet, just like the headline pictures".

Patterns

  1. Board not reflecting the work. S3-137, 138, 145, 149, 156, 241 (6 times, 3 tabs). Her board-first rule was written to memory in the SOCIAL tab (S3-143, feedback_board_first_audit_after.md), then she repeats "put whats done on board" (S3-153) and the SIMI tab (S3-241) had no sight of it. Class 10 pair: S3-153 and S3-241. Also S3-149: SOCIAL re-asked R14 she had already approved, MANAGER knew (class 10).
  2. Wrong method on LC-28/R22 repeated across tabs: S3-146, 147, 148 (no reskin, real face), S3-175, 190, 191 (aspect ratio, VO source), S3-172, 183, 204 (R22 structure). Fixes went into handover or briefs only (STATED).
  3. Source read was the wrong or stale one: S3-195 (machine-buyer ACCA content used for a customer reel), S3-209 and S3-210 (her pictures and the UKIYO finishing preset existed, not used), S3-233 (canon DNA.json says RM99, KB says RM198), S3-230 (walnut). One worth noting: the memory for RM198 already existed when she corrected (S3-233 error text), so a stored memory did not stop the RM99.
  4. Token-guard burn: S3-180, 184..186, 221, 224, 228, 237 and about 20 further rows show the guard error. She repeats "no need reread, use Jev to pick" 5 times (S3-186, 221, 224, 228, 237). The guard blocks; nothing makes a resumed tab load less.
  5. Her ruling repeated until it sticks (end results in script): S3-193, 195, 196.

NEW classes

  • NEW:board-deploy-unverified: deploy or link reported done, but her view (live board) shows old version or nothing. Examples S3-137, 138, 145.
  • NEW:copy-register-miss: copy tone or wording that follows house style instead of the client sales register she wants (pain to face, end result). Example S3-160 (only one row).
  • NEW:session-start-load-cost: a fresh tab burns its budget on required reading before any work. Example S3-180 (related S3-221, 228, 237 as class 5/3).

Notes

  • Referent for "tab did it": each ledger row names the tab in its session id (49f4bb30 MANAGER, e23b8a27 SOCIAL, fec86d6b SIMI, d6ea670d EDIT).
  • Low-confidence rows: S3-152, 168, 170, 183, 204 (agent action not visible in pack).
  • Jev not called; no two-class rows scored.

R6 timeline, S4-p1 (6-8 Nov, 103 rows) + S5-p1 (9-10 Nov, 75 rows). 178 rows, 60 ledger rows.

Referent: packs as read in study-1110/slices/packs; rule writes mostly not visible in pack, so NONE/STATED is a floor, not proof.

Days

  • 6 Nov (S4-001..~030): reel R14 walk-out keyframe and bg-morph rejected repeatedly (S4-002,003,012); client note "too AI / words with no point" (S4-005,007,020); "Unknown skill: deepen" x3 (S4-007,013,040).
  • 7 Nov (S4-031..~075): statics too AI -> remake; pink rule (S4-054); view links fail again and again (S4-018,019,030,039,059,071); weekly limit hit (S4-074,075); intent ledger / check_intent (S4-051,074).
  • 8 Nov (S4-076..103): b2 batch verify (S4-097 verify_b2.py, proven red once, lives in non-durable scratchpad); status asked repeatedly (S4-050,053); steps-as-point again (S4-092,093).
  • 9 Nov (S5-001..040): Kuching copy identical to Bukit Jalil (S5-015,054); client end-results found never ingested (S5-018) then "pending/restricted" challenged (S5-023); manager hand-offs between tabs (S5-031..037); overnight statics ask (S5-040).
  • 10 Nov (S5-041..075): night produced 0 of 42 statics (S5-041,045); reversal "use all client terms" (S5-062); Kuching advertiser borrow and paused build (S5-043..053); reels scripts still steps (S5-072,073).

Patterns (row ids)

  1. STEPS NOT END RESULTS, 6+ times: S4-005,S4-045,S5-014,S5-022,S5-072,S5-073. Fix each time was memory text (STATED). Recurrence after fix at every repeat. DIAGNOSIS s2 names deep-clean/step-1 as stored nowhere; it does not name the recurrence count.
  2. CLIENT RESULTS / BANS: S4-007 -> S5-018 (never ingested) -> S5-020 -> S5-023 -> S5-062 (reversals). 6 Nov "no own claim bans" memory existed; agent in S5-023 says "I've done this before". C1 explains it. S5-058: unlock did not reach the reskin tool (routing gap), so the unlock was stored, not consumed.
  3. PINK AND SAMENESS: S4-054 -> S5-035. Pink was ruled 3 Nov; memory only (DIAGNOSIS s2); template line still forced pink outfits on 9 Nov.
  4. UKIYO RUNS ITS OWN FLOW, not the shared engine: S4-012,S4-038,S4-055,S5-047,S5-066,S5-072 (box-fit only UKIYO). Memory "use reskin-autopilot never own flow" (24 Oct) predates all of them.
  5. CANNOT VIEW OUTPUT: S4-018,019,030,039,059,071 (5th+ view-link fault).
  6. R14 keyframe/bg-morph: S4-002,003,012,026,027,033,036 (4th round).
  7. Tab coordination: S5-034,036 (SIMI vs UKIYO same work), S5-063 two tabs share a title.

Rows DIAGNOSIS-1109b already explains

S4-005,007 (C1 never-say list does two jobs), S5-020,023,062 (C1), S5-035 and S4-054 (pink memory only), S5-072 and S5-022 steps part (s2 stored nowhere), S4-015 / fit rows (paragraph vs cover), S4-051/074 (intent ledger gap), S5-018 partly (ingestion).

Rows it does NOT explain (written 9 Nov, after or outside its scope)

S4-018/019/030/039/059/071 view links; S4-002/003/012/026/027/033/036 R14 morph; S4-074/075 weekly limit; Unknown skill deepen; S5-015/054/069 Kuching place swap; S5-034/036 tab split; S5-041/045 overnight 0/42 output; S5-047/066 UKIYO flow vs other brands; S5-058 routing gap after unlock; the recurrence COUNT of the steps fault.

Recurrence after a "fixed" error (matter most)

S5-072 (6 Nov step rule), S5-023 (6 Nov no-bans), S5-035 (3 Nov pink), S5-073, S5-022, S4-071 (view link), S4-036 (R14), S5-047 (24 Oct own-flow).

NEW classes

  • NEW: named-skill-does-not-exist (agent calls a skill that is absent): S4-007,S4-013,S4-040.
  • NEW: cannot-view-output (agent hands her output she cannot open): S4-030,S4-039,S4-059,S4-071.
  • NEW: capacity-stop (weekly limit, disk, Gemini quota halts a lane): S4-074,S4-075,S5-038.
Leave a note

Words in detail (TERMS.md)

R7 TERMS and ontology map (P2) - UKIYO Lady Care, 10 Nov 2026

Referents. Her counts: 909 human rows of turns.jsonl (in-scope sessions, task-notification rows excluded), this Mac only; per-day detail in term_counts.json. Runtime = ~/.zennith/runtime/os at 41ce8e3c05 (rev-list HEAD..origin/master = 0). STATE = ~/.zennith/state/ukiyo-lady-care-2026-10-28. Full table: terms.csv (40 rows). Jev: code-only, no call (barred by brief).

Top collisions

1. Brand never-words vs the client's own words (strongest)
  • runtime/os brands/ukiyo/ad-pain-hook-intel/pack.yaml:73 never_say (32 CN + 7 EN) bans 紧致, 修复, 精雕, 激活, 异味.
  • Those are the client's own names/results: S3 "私密紧致护理", S2 "私密修复护理", S3 "(徒手精雕)", S4 "激活海底轮" (KB.bak-1030:54, KB.md).
  • KB section 5 "Never (27)" (KB.md:199) was SUPERSEDED 6 Nov (KB.md:139-145): use client end results. Her ruling 10 Nov: "you may use all terms used by client" (remake-1110/build_banks_1110.py:8).
  • Still live, different lists: pointfix-1107/fit.py:11 HARD (仪器, 女性, bare 宫|性); remake-1107/scorer.py:54 same; bulk-1109/mgr_check_s4_copy.py:20-21 PENDING flags 紧致|闷热|血液循环, REJECTED flags 激活穴位 - while the 1110 bank lists these as usable_as_stated. mgr_check reads stage-a-1109, not remake-1110.
  • fit_1110.py (remake-1110/test-1110) reads NO word list: it checks row width only. So no 1110 script enforces any never-word; the old lists conflict with each other.
  • pack.yaml:29 points at facts.json that does not exist on this Mac.
2. steps (3 senses)
  • Pipeline step: "dont skip steps" 9a8ececa 14 Oct 06:13; "from step 0" 7375bb0b 22 Oct.
  • Treatment procedure: "UKI-NOV-NEW1-STEPS 细胞光 3.0 护理步骤" e23b8a27 5 Nov 08:35; SERVICES.md lists steps per service.
  • Not a selling point: "i noticed u miss out the part where client highlighted that the steps is not a point of sales ... Why your verification didnt catch this?" fec86d6b 7 Nov 09:35. Again 4308feb8 9 Nov 11:51: "the supporting copy is focused more on the steps. instead of after treatement, what customer will experience?"
  • Her counts: 步骤 1 row (5 Nov), "steps" 22 rows. Enforcement: scorer.py:12 mentions it in a doc line only; no regex.
3. end result / mechanism / points / USP / objective
  • "end result" 12 of her rows (9 Oct to 10 Nov): ICP x Persona x Scenario x End Results x Pain Point (7375bb0b 19 Oct 09:10); "point their problem to their face, give them the end result" (49f4bb30 4 Nov 11:59).
  • Bank: client END RESULTS are typed mechanism (build_banks_1110.py:116, 90 of 103 anchors); RESULT_TYPES = ("intended result","benefit") at :33; objective hardcoded "awareness" for every anchor (:76-78).
  • "USP": selling point (055a5d01 10 Oct 13:33) vs video end block (6c21af41 18 Oct 13:50). runtime/os has brands/ukiyo/usp-bank.json (built 2026-09-05, Mirra schema reskin) and zhongdian-bank.json (重点, built 2026-09-21). 卖点: 0 of her rows. No script reader of either bank found by grep over skills/.
  • "results" (22 rows) = client end results; bank record type "intended result" 73 vs "benefit" 13.
4. S1-S5 reused for five things
  • Service id (S1 HIFU .. S5 Brazilian laser, S5 added 3 Nov: S5 term count 0 in KB.bak-1030/pre-1103/1102, 11 in bak-1107).
  • AWS S3 ("push to S3"); Mirra static ids; reel shot names ("NG-01's third shot (S3)"); statement ids S2-022, S4-146; bank files S2-bank.
  • Real damage: LC-41 still mislabelled S7 in receipts -> S9 (14820dea, 3 Nov).
  • Her S-number use: 9 rows (16 Oct to 9 Nov); S2 5, S3 6, S4 3, S5 4, S1 3 (S2-S5 hits all 9 Nov).
  • Name drift: S4 "by-hand intimate spa" (28 Oct) -> 私密spa (29 Oct); site has External SPA RM588 and V-Spa under laser add-ons (KB.md:113-124: S2 no site match, S3 probable Internal SPA RM688, S4 probable V-Spa RM499).
5. skeleton / Use-mark / approved / reskin
  • skeleton = base photo (Image 1) needing use_mark.by+at (skeleton_gate.py:2-27, :94-100); reel = reference video, SKIPPED (:90); = OCR rows of the posted ad in fit_1110.py:82; = "slot to ICP/persona/pain" (7375bb0b 25 Oct 07:21).
  • Her words: skeleton 23 rows (14 Oct to 10 Nov); "for ukiyo i need to be like hold the agent hand step by step and all need reference" 7c5a4dba 7 Nov 07:59.
  • Use-mark: 0 of her rows; lives in db collection skelpicks; 3 Nov: only 6 of 21 slide skeletons were hers; 8 Nov: a base also counts if she saw it in a plan she OK'd.
  • approved (68 rows): Use mark, typed OK of a plan, /approvals click, Jev pick. approval-decision-atom is written by pillar-bulkfire/scripts/queue_decode_on_approve.py:13,18 and named in brief-composer/scripts/compose.py.
  • reskin (87 rows, 10 Oct to 10 Nov): reel = Wan 3.0 one fire (ac81bcc9 26 Oct 06:28); statics = edit skeleton (ec9ff1dc 2 Nov 09:34); promo change (055a5d01 12 Oct 06:26: "reskin and changing the promo aint much of a difference"); autopilot decode+keyframes (5dbeef0a 27 Oct 13:18). skills/reskin-autopilot/SKILL.md covers batches only.

Other collisions

  • fit: fit.py = claims + budget (pointfix-1107/fit.py:11); fit_1110.py = budget only.
  • hook: card CN headline (HEADLINE-REWRITE-1104.md:29-30) vs reel first seconds; headline vs board title (:9). Her counts: hook 19, headline 19 (30 Oct to 10 Nov), body 8 (body copy vs body photo at the intimate area).
  • point: pain point / winning points / 重点 / "what is the point" (833ab602, 9a8ececa 15 Oct 07:41). 34 rows.
  • SIMI: her label for the UKIYO tab/lane (ec9ff1dc 29 Oct 05:42; 49f4bb30 4 Nov 13:50); no definition found in turns or STATE md.

KB version facts (python difflib, by mtime)

  • KB.bak-1030 (30 Oct 10:59, 192 lines) -> pre-1103 (3 Nov 15:45, 258): added "client's scripts and menus (sent 29 Oct)"; class counts Ad-safe 1->4, Reframe 12->15, Consult-only 3->6, Never 18->27.
  • bak-1102-pre-prices (30 Oct 12:24, 254) -> bak-1107 (3 Nov 15:45, 298): added S5 and "Her list, 3 Nov". bak-1107 -> live (6 Nov 15:30, 304): only the SUPERSEDED block.
  • Terms with 0 hits in every version: 步骤, USP, 卖点, reskin, skeleton, approved, hook, headline. "end result" 0 until live (1). The KB never held the working vocabulary.

Not-found transcripts

  • grep -rn 卖点 over her 909 rows (count_terms.py) -> 0 hits.
  • grep -rIn skelpicks skills .claude in runtime/os -> empty output.
  • ls ~/.zennith/workspace/ukiyo/wa-sales-agent-2026-11-01/pack/facts.json -> No such file.
  • brands/ukiyo/lady-care is not in runtime/os (unmerged worktree), so no runtime copy of the line files exists to read at fire time.
  • LINE-DNA.json/.md live copies are not in STATE; only the .pre-1103-services.bak copies.

Limits

Counts are from this Mac's turns only. Fire-time reader column = file:line where found; "not read" rows mean grep found no script consumer, not proof of none elsewhere (runtime/os only searched under skills/, sidecar/, zennith_core/, .claude/, brands/ukiyo).

Leave a note

Where the rules live, and who reads them (RULE-STORES.md)

R8 P3 - Lady Care rule-store map (10 Nov 2026)

Referents: engine = runtime pin ~/.zennith/runtime/os HEAD 41ce8e3c05 (0 behind origin/master). ~/zennith-os = ad97be0b65, 4369 behind, NOT used for engine claims. STATE = ~/.zennith/state/ukiyo-lady-care-2026-10-28. All counts taken 10 Nov.

1. Stores

# store count duplicates contradictions / stale
1 memory dir (931 .md) 328 match ukiyo/lady care/私密: feedback 162, reference 105, project 57, user 1, untyped 1 operator-memory copy: 326 match, 2 originals missing, 6 older MEMORY.md 5 lines cut at load; archive 75 lines match
2 rules ledger ukiyo.json 45 rules (30 active/14 cand/1 retired); 4 name Lady Care feeds brand pack last_seen max 5 Nov; 3 rules carry a check
3 brand pack pack-c238ba6a 30 active rules in compiled copy of #2 built 9 Nov at older repo_head c489daf65f; taste input timed out
4 brand-requirements.json:634 Sep decisions only, 0 Lady Care - batch-local, not repo canon
5 Lady Care canon (KB.json 103 facts, LINE-DNA.md 333 lines, READINESS) branch worktree only KB.md copies not in master, not in runtime pin; STATE KB.md cites a KB.json STATE lacks (ls: No such file)
6 STATE KB.md / SERVICES.md 304 / 123 lines; 52 class rows worktree KB.md (4 Nov) KB.md:140 STATE "SUPERSEDED 6 Nov" vs worktree :140 "4 Nov 2026 override ... c47 异味 MAY be named"
7 DESIGN.md 2 shas: bbb1cd656020 (pin, 237 l) vs 9808f3d92f03 (~/zennith-os + ~/.zennith/brands, 227 l) 9 worktrees = bbb 9808 "AI hero allowed (ruling 2026-11-03)" vs bbb decision 40 "never AI-generate"; both "NO second accent", neither has pink/lady
8 tab handovers/plans/diagnoses 112 doc-like files - 10 Nov ruling supersedes earlier word lists
9 tab scripts 988 scripts, 199 mention a store token hardcoded word lists see case 1

2. Reads: what fires

script opens at
reskin-autopilot/readiness.py brand-requirements.json :514
reskin-autopilot/brand_pack.py compiled pack (4 Lady Care rules) :78-111
reskin-autopilot/learnings.py batch-local --decisions file :97
codex-image-gen/gen.py none of the rule stores (grep ukiyo lady
atom-router/check-run-gate.sh none (no ukiyo token) -
god-mode-audit/god-audit.sh DESIGN.md ~/zennith-os first, then ~/.zennith/brands; head -80/-40; DNA.json :801; RULES.md :1291 (absent for ukiyo) :826,:1315
tab check scripts hardcoded regex, KB.md price line (split_s4.py:52), KB.json from worktree (tests/) -
scorer NOT PINNED - not located; open item -

Net: the engine reads 4 Lady Care rules (via the pack) and a stale DESIGN.md; the canon (KB.json, LINE-DNA) is read by no engine script.

3. Three stale cases

  1. CONFIRMED. bulk-1109/mgr_check_s4_copy.py:20 PENDING=紧致|闷热|不适感|远离异味|异味|血液循环 and :21 REJECTED=深层清洁|深层护理|激活穴位|疏通|激活性腺, both turned into FAIL checks (:59-60). Superseded by her 10 Nov "you may use all terms used by client" (reels-acca-1110/STALE-SOURCES-1110.md:9-11); DIAGNOSIS-1109b.md:228-229 allows 异味. The same REJECTED list is also copied in split_s4.py:20,64.
  2. CONFIRMED. god-audit.sh:826,1315 read ~/zennith-os/brands/ukiyo/DESIGN.md (9808, locally modified, stale checkout) before the pin's. Neither has the pink-touch rule; it lives only in the worktree LINE-DNA.md section 6. 292 audit runs in STATE logged "loaded DESIGN.md from" that path. Effect on record: 7 Nov pink marked down as a second accent.
  3. CONFIRMED, partly. KB.md class tables (Ad-safe L155, Reframe L168, Consult-only L188, Never L199; 52 rows) were superseded by the banner at KB.md:140-144 (6 Nov: her words against self-made claim bans, points to WORDS-1107j.json) and by the 10 Nov all-client-terms ruling. Tables never rewritten (STALE-SOURCES-1110.md:37,41 propose edits, routed to MANAGER). No script parses the class tables: KB.md readers only take the price line (split_s4.py:52) or provenance sha (build_plan.py:27,730). Stale in docs, not in code.
Leave a note

How much room a tab had to work (FINDING-CONTEXT-BUDGET.md)

FINDING: the working room each tab had (context budget)

Source: slices/context-budget.tsv (built from this Mac's session logs, 111 sessions in scope) and LEDGER-WRONGS.csv (280 rows). "Tokens" = the context size the logs report on each turn. A compaction = the log shows the context drop sharply (the tab summarised itself and lost detail).

In plain words

  • Every tab starts with about 97,403 tokens of fixed load (rules, memory, hooks) before she types anything (median, all 111).
  • 85 of the 111 tabs (2026-10-04 to 2026-11-10) ran in the small window (~166k). That leaves about 58,595 tokens of room per cycle (median of 85 sessions). Then the tab compacts and comes back at about 107,582.
  • Those 85 small-window tabs compacted 4,450 times. The 14 large-window tabs (~1M, up to 21 Oct) compacted 55 times.
  • Each compaction is a point where her exact words can turn into a summary line. This is the "context loss" class (5).

Busiest tabs

session started (UTC) compactions work room per cycle
fec86d6b 2026-11-03T13:45 374 56,478
e296ae26 2026-10-28T08:17 332 49,396
ec9ff1dc 2026-10-29T05:42 288 52,017
0104f8c6 2026-10-29T07:10 210 56,439
7c49523a 2026-10-26T07:05 167 54,267
077efd30 2026-10-25T12:22 152 54,392

Corrections by window

window sessions her corrections (ledger rows) rows with class 5 (context loss), first or second class
small (~166k) 85 211 27
large (~1M) 14 69 4
no compaction 12 0 0
session not in table - 0 0

What this does NOT prove

  • The large-window tabs were mostly before Lady Care (to 21 Oct). Later work was harder and had more tabs, so the split is not a clean test.
  • It is this Mac only.
  • It shows where detail can drop. It does not show that a given correction came from a compaction. The ledger rows tagged class 5 are the cases where a reader saw it.
Leave a note

Re-makes and checks (REWORK.md)

REWORK: how much was drawn again

Source: rework-summary.json, rework.csv, rework-keys.csv (tools/rework.py). Measured against: corpus/actions.jsonl (this Mac's session logs) + readers R1-R6 + ~/.zennith/state/ukiyo*/**/*.log. 35 days. Day = Malaysia day.

In plain words

  • Commands found in the logs: 79 free picture runs, 258 video-engine runs, 910 check/audit runs, 837 board runs. For each picture or video made there were about 2.7 check runs and 2.5 board runs.
  • 126 picture/video runs name an ad (LC-28, R6, NEW4, ...). 71 of them (56%) re-made an ad first made on an earlier day.
  • Codex picture logs read: 108. 93 of them had at least one empty draw (the tool returned no picture), and 15 were dropped for the wrong shape.
  • Her turns in the readers' tags: 909. Of those, 243 are corrections and 72 approvals.

Ads re-made most often

ad picture/video runs on how many days
R6 34 7
R5 24 4
NEW4 16 3
LC-28 14 5
R3 13 6
LC-27 11 5
NEW1 7 3
R4 6 2
LC-29 6 3
R1 5 1

What came just before a re-make (proxy)

For each re-make, the class of her last correction in the same tab before it:

  • class 7: 15
  • class 3: 14
  • class 6: 9
  • class 8: 7
  • class no-prior-correction-in-session: 7
  • class 1: 6
  • class 9: 4
  • class 10: 2
  • class NEW:board-deploy-unverified: 2
  • class 5: 2
  • class NEW:review-surface-unusable: 1
  • class NEW:overcorrection_from_one_note: 1
  • class NEW:board review surface: 1

Caveats

  • A "run" is one command line. One command can draw several pictures, and one picture can need several commands.
  • "Video-engine runs" counts commands that call a video engine; it is not a spend figure. Money spent is not measured here.
  • Only runs that name an ad in the command count toward re-makes. Many runs name only a folder, so re-makes are under-counted.
  • The cause column is a proxy (the last correction before the run), not a proven cause.
  • This Mac only.
Leave a note

Anything else