WORKLOG — ARCHIVE

Every week.

71 logged days across 17 weeks · back to worklog →

WEEK 31Jul 27 – Aug 02, 20262 done — 2/7 days

FRI 07·31

@reading new VLM/MLLM paper to find research path✓VLM/MLLM start✓Reveiwed journal paper once again, fixed alot>Maibot OC and config>Leetcode and job preparation>Read ADHD related Papers>prepare public repo for journal paper.

TUE 07·28

>Maibot OC and config>Leetcode and job preparation>Read ADHD related Papers>VLM/MLLM start

WEEK 30Jul 20 – Jul 26, 202615 done · 3 decisions — 7/7 days

SUN 07·26

>Maibot OC and config>Leetcode and job preparation>Read ADHD related Papers>VLM/MLLM start

SAT 07·25

@blog English/Chinese grammar fix@preparation for journal github repo public>Maibot OC and config>Leetcode and job preparation>Read ADHD related Papers>VLM/MLLM start

FRI 07·24

✓Updated CLAUDE.md for research repo, refined guildlines.✓paper review✓obsidian private git>Maibot OC and config>Leetcode and job preparation>Read ADHD related Papers>VLM/MLLM start

THU 07·23

✓Refresh the model-comparison section of the prompt post — still the one genuinely stale piece on the site.✓Walk through the backfilled weeks and the site's data files for a sanity pass.✓Notion to obsidian transfer>paper review>obsidian private git>VLM/MLLM startNow for some new research

WED 07·22

@Blog: the "Who Am I?" rewrite.✓The 2026 rewrite went live and took over the main URL; the 2025 original is archived verbatim at its own address, linked from the new ending. Mirror pages, no surgery on the old text.✓Three new photographs shot into the post; the opening borrowed poem is replaced by one I wrote.✓Reconstructed this month's and last month's missing work-log entries from transcripts and commit history across all three machines.✓Month closes are public now: a front-matter layer renders as a card at month boundaries in this very stream — June's is the first.✓The Chinese landing page reached parity with the English writing index (year groups, filters, descriptions), and the language toggle now pairs the two hubs both ways.✓The worklog format itself got a written contract (docs/WORKLOG.md) so future edits know exactly what renders where.=Biographies grow in place; opinion snapshots freeze and get a sequel. That is why the old answer keeps its own page instead of being edited.=Month close stays a hand-written ritual for now — composition, not capture. Console support waits until the habit proves itself.>Refresh the model-comparison section of the prompt post — still the one genuinely stale piece on the site.>Walk through the backfilled weeks and the site's data files for a sanity pass."Who Am I?" has its 2026 answer; the 2025 one stays archived.

TUE 07·21

@Blog: the "Who Am I?" rewrite.✓Wrote the whole 2026 version myself rather than operating on the old one.✓Polish pass, two variants of the opening poem, and a photo selection from seven candidates.>Finalise and ship.

MON 07·20

@Blog: planning the "Who Am I?" rework.✓Mapped what in the 2025 essay is dated and what still holds; drafted the surgery plan.=The heavy parts are concentrated in a few anchors — the body survives. A full rewrite was not obviously needed at this point.>Decide between surgery and a new version.

WEEK 29Jul 13 – Jul 19, 202623 done · 3 decisions · 2 risks — 6/7 days

SUN 07·19

@Journal manuscript; side project.✓Architecture figure redrawn and its caption brought in line.✓IEEE table styling, spelling and terminology consistency pass.✓Extension statement and a clinical-significance sentence added.>Keep tightening the manuscript.

SAT 07·18

@Journal manuscript: format and compliance overhaul. Blog in the gaps.✓Found that this journal's review is single-blind, not double-blind — I had carried a same-named conference's rules over to the journal. Full de-anonymisation followed.✓Verified all 46 references one by one against DOIs, with an independent review pass on top: eleven factual errors corrected, including a page range that differs between the open-access and the publisher version.✓The bibliography had no `References` heading at all — the template does not print one.✓Em dashes cut from 63 to 11, with no sentence keeping more than one; every rewritten sentence passed an independent semantic check.✓Cleared every annotation the advisor left on the PDF: citation placement, footnotes, a figure-numbering inversion, and abbreviations that were never expanded.✓Built a local LaTeX toolchain so I can compile and read the PDF myself instead of relying on Overleaf — it immediately caught two problems Overleaf cannot show.✓Blog: reframed the regression note as a quick-reference, added an `updated:` field for honest revision stamps, built the Chinese research page, and audited the whole site for stale content.=Answer-first, theory-after is the standing shape for technical posts here.!Page budget still over target.>Keep trimming; then figures.The manuscript survived a compliance overhaul — every reference verified by hand.

FRI 07·17

@Blog design pass; translation review; journal diagnostics.✓Site-wide compliance sweep against the boring standards — contrast floors, a real focus ring, honest `lang` attributes, minimum hit areas.✓Typography and anchors: killed the wispy light weights at text sizes, made deep-link landings look deliberate.✓Prose spacing ladder so gaps encode hierarchy; nav hover pill; article covers crop to 3:2 and bleed into the page.✓Both Chinese translations reviewed through several independent lenses — terminology, a picky native reader, a mechanical style-contract check, and a faithfulness pass — each finding then adversarially re-checked against the actual files.=One convention I had written down was simply wrong: the byte-level check found no full-width parentheses anywhere in the Chinese files. Corrected the note rather than the files.>Land the translations.

WED 07·15

@Blog typography.✓Weights made honest — the serif has no 500, so nothing pretends it does.✓Full dress-code coverage for prose elements, with the heading ladder capped.>Keep going on the design pass.

TUE 07·14

@Journal manuscript review.✓Eight-dimension review of the draft: 46 findings, 7 of them serious.✓Fixed five outright errors; the rejection-risk items are listed and waiting on my call.!The three that actually decide the outcome are novelty, construct validity, and having only one dataset. None of them are fixable by editing prose.>Decide which risks to answer and which to concede in the limitations.

MON 07·13

@Journal manuscript: the big writing day.✓Asked, seriously, whether this paper can survive review rather than be desk-rejected — and looked for hard evidence either way.✓Related Work written — the last remaining stub. Six-paragraph funnel; references up to 46 with no undefined, orphaned, or duplicated entries.✓Outline migrated from markdown into the journal's LaTeX template.✓Wrote Setup, Results, Discussion, and Conclusion; added the architecture figure and a confusion heat map.✓Switched the proposed configuration after a five-seed run and updated every number and table that depended on it.=Target journal fixed as the first submission, with the narrative spine settled: method leads, reproducibility endorses.>Compile and check formatting.Every section of the journal draft now exists in LaTeX.

WEEK 28Jul 06 – Jul 12, 202616 done · 3 decisions · 1 risk — 6/7 days

SUN 07·12

@Blog console and mobile; journal §III annotations.✓Console put on a text diet — pedagogy off the tool's face — and closing the day now turns the page.✓README stripped to a bare pointer; project reference moved into the agent file.✓Fixed the 375px breakages and made the language toggle context-aware.✓Journal: clarified two habitually confused normalisation steps, changed a parameter claim to "adopt implementation defaults" (which is harder to attack than "we chose"), and verified §III.A against the dataset's own paper line by line.✓Scouted the landscape for the ADHD essay: found three genuinely unclaimed angles, and caught that a widely repeated "23 minutes to refocus" figure comes from an interview, not a paper.=Commit messages go in English from today.>Related Work is the last unwritten section.

SAT 07·11

@Blog V3 worklog access control@J-BHI III.B-C check.@matrix post translation review!Can't finish whole paper before Monday.>Review the V3 pages, then build the entry console — the last big piece.>Finish the paper.Journal WIP....

THU 07·09

@Blog V3 build.✓Homepage shipped in the V3 darkroom design: splash, subject readout, live strip, research panels, writing, proof wall, coda.✓Writing index, article page (ToC rail, reading progress, KaTeX, coda), tag pages, public worklog + archive shipped.✓Content gates live: one validator runs at pre-commit, build, and CI; conflict markers and malformed dates can no longer land.✓Dev/build environment moved fully into Docker; no node_modules on the Mac.=Worklog truth stays in git markdown; SQLite (Cloudflare D1) will only be the intraday working tree for the entry console.=Legacy 2026-04/05 files stay verbatim; the V3 parser reads both generations of the format.>Review the V3 pages, then build the entry console — the last big piece.

WED 07·08

@Experiments on the training box; new site started.✓Second experiment's five-seed set landed.✓Pre-scaling the inputs removed the data-loading bottleneck for good.✓Started the V3 site from the design handoff — Astro scaffold with the MDX and maths pipeline.✓Read both previous generations of the site (the Jekyll one and the Astro one) for migration material.·This is the second redesign in two months. Re-re-design.>Build the V3 pages.

TUE 07·07

@Training throughput.✓Chased low CPU and GPU utilisation — the box was idling between the same two states.>Keep the runs moving.

MON 07·06

@Experiments; taking stock.✓Wrote a long summary of what I am actually doing and where the papers are heading — mostly for my own benefit.✓Scout run for the wide-EDA mechanism; dimension variants parameterised.>Run the variants.

WEEK 27Jun 29 – Jul 05, 202610 done · 3 decisions · 1 risk — 6/7 days

SAT 07·04

@First experiment, core results.✓Core results landed on the repaired pipeline.=The old headline finding does not survive on clean data. It goes.>Pilot the dimension variants.

FRI 07·03

@Getting the runs onto the training box.✓Three problems in a row on the remote box: single-core loading, an out-of-memory failure from pre-loading large images, and an extraction-format mismatch. All fixed.✓First genuine post-repair result in hand.>Full core run.

THU 07·02

@Packaging the retraining bundle.✓Bundle finished; five independent adversarial checks on it caught four documentation errors — the environment claim among them.>Ship it to the box.

WED 07·01

@Full codebase audit.✓The audit overturned "the experiments are done". A library version mismatch had produced a cluster of silent failures: the first stage's pretrained weights were never actually loaded into the second stage, two model parameters had been silently inert for months, and early stopping was reading the outer validation split — an optimistic bias.✓Two of the four physiological features were also near-useless as computed: over-detection at the sampling rate, and a within-window normalisation that erased the absolute level.✓The data pipeline is leakage-safe. Triple-checked. That part holds.✓Fixed all of it the same day and defined the new protocol.=Every earlier result is void. Full rerun, no exceptions — the honest numbers will be lower and that is fine.=Fixes: real weight loading, inner-validation early stopping, literature-grounded features, normalisation removed where it destroyed signal.!Weeks of compute to redo, with the journal deadline already moved once.>Package everything for the training box and rerun.The audit voided every old result. Rerunning everything — honest numbers only.

TUE 06·30

@Research: journal writing (IMRaD+ outline).✓More outline sections drafted and annotated.>Keep filling the outline section by section.

MON 06·29

@Research: journal writing (IMRaD+ outline).✓Long outline session; several sections clarified.>Keep filling the outline section by section.

MONTH CLOSEJUNE 2026

June was supposed to be journal-writing month. A five-day conference deadline landed on the 10th instead — the month became three deliveries back to back, with the journal advancing in the gaps.

A conference paper went from zero experiments to submission in five days — later withdrawn over fees; its audit now motivates the journal.Short paper presented; in-school doctoral and seminar talks delivered.Publication program locked as a chain: preliminary → companion → capstone.Journal by month end: front matter final, methods section half written.

WEEK 26Jun 22 – Jun 28, 20267 done · 3 decisions · 1 risk — 5/7 days

SAT 06·27

@Research: journal writing (IMRaD+ outline).✓More outline work.>Keep filling the outline section by section.

FRI 06·26

@Research: journal writing.✓Confirmed both talks are behind me: the short-paper presentation and the in-school doctoral / seminar talks are done.✓Front matter (title / abstract / intro) finalized; §III.A complete; §III.B half written.✓Made the next-journal direction concrete (vision-language on a larger dataset, adversarial to rPPG artifacts) — parked until this draft is out.=The conference paper is withdrawn — the attendance fee was too high to justify going.=Its rPPG reliability audit folds into the journal as motivation only — not a companion paper, no new contribution, and the retraining-free fusion does not come along (it is not a classification result).=Everything now goes to the journal.!A paper was written, submitted, and paid for in time — and will not be presented. The work survives only as a motivation section.>Finish §III.B, then III.C–E.

WED 06·24

@Presentations.✓In-school doctoral presentation and seminar talk delivered.>Back to the journal.

TUE 06·23

@Presentation slides.✓Heavy slide work ahead of the in-school talks — wording and figures, pass after pass.>Deliver the talks tomorrow.

MON 06·22

@Research: journal writing (IMRaD+ outline).✓Back on the outline between presentation prep.>Slides for the in-school talk.

WEEK 25Jun 15 – Jun 21, 20268 done · 2 decisions — 5/7 days

SUN 06·21

@Research: journal writing.✓Back to the journal after two deliveries.✓Advisor confirmed the baseline author is not a co-author.✓Fold-level statistics computed for the journal.=Publication program locked as a chain, not three separate papers: short paper = preliminary, conference paper = companion, journal = capstone.=The eight-method deep screening leaves the journal and gets cited to the conference paper instead.>Continue §III.B.

FRI 06·19

@Presentation slides for the short paper.✓Slide content and related-work framing.>Finish the deck.

THU 06·18

@Short-paper presentation.✓Short paper presented.>Prepare the in-school doctoral / seminar talks.

WED 06·17

@Conference paper wrap-up; presentation prep.✓Submission-ready LaTeX manuscript and final figures committed for the conference paper.✓Started the slide deck for the short-paper presentation.>Present on the 18th.

MON 06·15

@Conference paper manuscript.✓Conference paper submitted, on the deadline the advisor set.>Presentation slides.A conference paper, written from zero in five days, went out on deadline.

WEEK 24Jun 08 – Jun 14, 202614 done · 3 decisions · 3 risks — 6/7 days

SUN 06·14

@Conference paper manuscript.✓Full writing day on the manuscript. No commits — prose only.!One day left before the hand-in.>Hand the complete paper to the advisor tomorrow.

SAT 06·13

@Conference paper manuscript.✓Experiments aligned onto the complete method set across three window lengths and all subjects.✓All figures and tables regenerated from scratch — no stale artifacts; analysis scripts consolidated down to the paper-output set.✓Results settled: classical family beats the deep family; fusion saturates after two or three good methods; per-subject calibration rescues single methods but not fusion; HRV is not recoverable, which points toward EDA.=The conference paper is now the main deliverable: the advisor wants the complete paper on the 15th to read and submit directly. Same day as the conference deadline.!Writing a full paper in 48 hours.>Rewrite the draft against the final numbers.

FRI 06·12

@Conference paper: experiments and manuscript.✓rPPG reliability audit, classical-vs-deep benchmark, and retraining-free fusion all landed as code.✓Verification scripts written alongside: HR derivation, alignment/tags, fusion no-leakage.✓Results summary, first manuscript draft, deck generator, execution plan.>Lock the numbers and rewrite the draft.

THU 06·11

@Conference paper experiments; journal §III.B in parallel.✓All conference experiments ran overnight. All three contributions hold — GO.✓Kept writing the journal's signal pre-processing section in parallel.=What got carved out for the conference is the journal's rPPG screening step — a result that was already null for the journal. No journal asset or protocol was touched.>Turn the results into a manuscript.

WED 06·10

@Reporting tables; then, from the afternoon, an unplanned conference paper.✓Summarised the reporting statistics into one table (corrected p, effect sizes, naming, descriptions).✓Afternoon: the advisor asked for a brand-new conference paper by the 15th — five days, with no experiments started.✓Spent the rest of the day, straight through to night, turning "nothing" into a plan.=Take the journal's rPPG screening step — the part whose result was null for the journal anyway — and build the conference paper on it. Nothing the journal depends on gets touched.!Five days. No experiments. The journal deadline does not move.>Run every experiment tonight.

MON 06·08

@Research: journal writing.✓Restructured the IMRaD+ outline; section headings reorganised with detailed comments.✓Updated the old site's work log — as it turns out, the last commit that site ever took.>Keep filling the outline.

WEEK 23Jun 01 – Jun 07, 202612 done · 5 risks — 4/7 days

SUN 06·07

@Research: journal writing.✓more IMRaD+ outline.✓more journal learning.✓more section clarification.!Deadline!I wanna go IEEE>Finish all research for journal.>1 new post>Chinese / English switch for new blog

WED 06·03

@Research: journal writing.✓linked evidence tables/files to claims and contributions.✓did some more IMRaD+ outline.✓Title is mostly locked as "EDA-Augmented Multimodal CCT-LSTM for Stress-Related Task Classification on UBFC-Phys", but subject to change.!Deadline!I wanna go IEEE>J1 skeleton + Related Work; read exemplar before fixing diagnostic-chain length.>1 new post>Chinese / English switch for new blog>Finish all research for journal.

TUE 06·02

@Research: journal writing✓listed the claims and contributions of paper✓did some of IMRaD+ outline.✓Had meeting with my professor, paper title will not has any thing about "leakage safe".

MON 06·01

@Research: journal writing✓Synced job plan and research plan with claude and code repo.✓Drafted a co-author invitation email to the baseline author; sent to my professor for review.✓Learned about the IMRaD+ structure of journal papers and reading exemplar J-BHI papers to understand the style differences vs normal conference papers.!DEADLINE: ~June 15.>J1 skeleton + Related Work; read exemplar before fixing diagnostic-chain length.>1 new post>Chinese / English switch for new blog>Finish all research for journal.

WEEK 22May 25 – May 31, 20267 done · 2 decisions · 1 risk — 1/7 days

SAT 05·30

@Research (Journal 1 / Journal 2 direction) + overall career-research plan consolidation✓Completed a two-round deep strategy review (job market → job family → research alignment). Outputs saved as standalone planning docs (see Decision).✓Locked Journal 1 target = IEEE J-BHI, with a corrected framing:✓J1 writing artifacts (no new training): main results table with paper-friendly names, paired stats, per-class F1, confusion matrices, EDA-feature-by-task table + figures.✓Confirmed statistical handling: paired t + Cohen's dz only; drop Wilcoxon (n=3 → min p=0.25). Stats are supporting; main evidence = protocol + consistent deltas + ablation.✓Cleaned all wording that could read as criticism of the original CCT-LSTM rPPG implementation (red line for the co-author invitation).✓Locked Journal 2 direction (starts after J1 draft): VL-FAU style vision-language joint learning on BP4D++ (main) + a fully public-dataset LoRA demo on UBFC-Phys (companion, to have HR-visible artifacts under BP4D++ EULA limits).✓Drafted (not sent) a co-author invitation email to the baseline author; pending advisor sign-off.=Detailed execution plans now live in standalone docs (the "look-up layer" when this worklog is too coarse):=J1 over J2 over job-prep in priority until the J1 draft is out (~June 15).!- Whole plan is tight under ~8–15 h/week (research-external). If research time shrinks, LeetCode yields first; floor = Hot100 + LoRA-runnable. - J1 timeline depends on the ~June 15 draft pin. If it slips past July, J2 + coding + spring-2027 application state all cascade. - BP4D++ J2 has an un-scoped preprocessing/storage risk (15TB). Do a half-day reconnaissance before committing (confirm VL-FAU likely needs only a subset → avoids moving 15TB).>- Sat advisor meeting (authorship + venue + stats). - Then: J1 skeleton + Related Work; read exemplar before fixing diagnostic-chain length. - 1 new post - Chinese / English switch for new blog - Finish all research for journal.

WEEK 21May 18 – May 24, 20264 done — 2/7 days

THU 05·21

@Research✓Finished the journal grade diagnostic chain:✓Had new protocol E9 and old E6 E3 running with a minor improvement on E9 at +0.0438 macro-F1.✓We can claim EDA does improve the CCT-LSTM model for stress classification.>1 new post>Chinese / English switch for new blog>Finish all research for journal.

MON 05·18

@New blog✓Now new astro blog is fully working and operational:>1 new post>Chinese / English switch for new blog>Finish all research for journal.

WEEK 20May 11 – May 17, 202618 done · 7 decisions · 7 risks — 6/7 days

SUN 05·17

@New blog backend / Cloudflare Pages staging✓Synchronized old posts, worklogs, and draft coverage from the old blog into the new site.✓Added a release gate for content validation, Astro check/build, artifact verification, and old-blog audit.✓Migrated old post layout semantics into explicit MDX content components and removed the Chirpy/Kramdown compatibility layer.✓Configured GitHub Actions as build-only release validation.✓Disabled GitHub Pages deployment for the new repo to avoid project-path preview conflicts.✓Documented Cloudflare Pages setup as the staging and production path.=Use Cloudflare Pages `*.pages.dev` as the only real staging environment.=Keep `enderjones.com` on the old site until the new staging deployment passes final checks.=Do not use GitHub project Pages for preview because this site is built for the root path.!Accidentally selecting `Ender-Jones/Ender-Jones.github.io` in Cloudflare would keep deploying the old site.!Binding `enderjones.com` before staging verification would prematurely cut over production.>Create a new Cloudflare Pages project connected to `Ender-Jones/personal_blog_astro`.>Verify homepage, posts, worklogs, tags, about, 404, robots, sitemap, images, and Giscus behavior on the `*.pages.dev` URL.>After staging passes, archive/stop the old site and bind the production domain to the new Cloudflare Pages project.

SAT 05·16

@New blog design✓Finalized most of the design.!DO MORE LEETCODE/ACM.>finish research and journal paper.

FRI 05·15

@New blog design / Research✓Astro blog 4 iterations.✓Made a levels experiment final journal-grade diagnostic chain.✓Rewrite README.md for research for journal.=We can not make a levels experiments work using current model.!Should go for lower level place for journal.>Finish the diagnostic chain.

WED 05·13

@Experiment result investigation / Context file manage✓Reconstructed md memory files.✓Added single seed investigation script.✓Made some investigation for low performance issue.✓T3 low/high does not have strong signal, binary feature screening of logreg/linear_svm showed even contact BVP macro-f1 is 0.5596.✓z-score/no z-score in Quantile MTF showed completely identical picture.=Do future investigation for journal writing.!Gonna go fast.>More investigation.

TUE 05·12

@Research monitoring & adjusting.✓Core run complete with lower than baseline performance.=Investigate problem and improve.!Might cost my paper can publish.>Define the problem is in the method or data processing flow.

MON 05·11

@UEvolve implementation: - Scoped a closed-loop RAG workflow connecting indexing, retrieval, decision evidence, and knowledge write-back. - Scoped a self-extension validator to check scaffold, schema, tests, dry-run apply, build, execution, and rollback. - Scoped an export/import package flow for sharing self-generated tools as validated distributable packages. Research experiments.✓RAG part is finished and tested, PR is merged✓Experiment 2 core runs is up and running.=Do more research things!Journal won't be in time at this rate.>EDA-Augmented CCT-LSTM evaluation experiment part 2.

WEEK 19May 04 – May 10, 20269 done · 3 decisions · 4 risks — 5/7 days

SUN 05·10

@UEvolve vibe coding✓Reviewed the project with Meng.✓Had plan for next steps.=Collaborate with Meng on the implementation of the UEvolve.=Defined a parallel plan for multi-agent orchestration.!Research is remain the first priority.!Implementation may take more time than expected, which could delay the research progress.>Start implementing the UEvolve based on the defined plan.>Continue with the research experiments in parallel.

FRI 05·08

@Paper writing✓Whole paper review and polishing.✓Paper sent to professor for submission.>Submit the paper.>Prepare for next experiments.>Review past exam question sets.>Plan next research projects.>Practice English.>Get Korea's visa.(After passport renewal)

THU 05·07

@Paper writing✓Paper method section EDA-argumentation part.✓Paper evaluation protocol description part.>Finish the paper.

WED 05·06

@K.I.C.S. Paper writing✓Paper method section paper reproduction part.!Paper might be a little too long.>Finish the paper.

TUE 05·05

@Paper writing / experiment results review.✓Paper related work section.✓Reviewed all clean p128 results for P1-P8.=K.I.C.S. paper should only include P0/3/4.!Deadline is close.>Finish method and experiment sections.>Prepare for submission.

WEEK 18Apr 27 – May 03, 202623 done · 11 decisions · 4 risks — 4/7 days

FRI 05·01

@Paper writing✓Paper introduction.!Intro and abstract are little too long and may need to be shortened.>Finish the related work section, then move on to method and experiment sections.

WED 04·29

@Paper writing / experiment monitoring.✓P0/P3/P4 clean runs finished and generated results.✓Paper abstract.✓Paper/Journal writing plan.=The paper should be written for K.I.C.S. first, then extended for a journal.=K.I.C.S. paper only need P0/P3/P4 results form experiment 1.=Journal paper can include more protocol variants and experiment 2.!I don't see any major risk at this point, but the paper writing process can always be more time-consuming than expected.>verify split metadata for all runs and results for P0/P3/P4.

TUE 04·28

@Paper preparation / experiment monitoring / job preparation.✓2/3 seeds of P0/P3/P4 finished and generated results.✓Estimated time of finish: Apr.29th 06:00 JST.✓Estimated f1 between baseline and best method: 0.1x.✓Bought custom domain and set up DNS for this Jekyll site.✓Made an intro animation for the homepage.=Switch job preparation to low, about 1 to 2 hours per day until K.I.C.S. paper is submitted.=Use the task prediction protocol for the paper, but clearly note the differences from the original CCT-LSTM paper.=Use the full reproduction for journal submission later, but do not delay the current paper for it.!I forgot all the experiment protocol are for tasks prediction, which means:>Finish paper's prior work and method sections.>Do some ACM coding questions for job prep.

MON 04·27

@Public worklog site / local Jekyll development / clean p128 experiment queue.✓Installed Docker Desktop and verified `docker`, `docker info`, and `docker compose` were usable locally.✓Turned off unnecessary Docker Desktop features for this workflow.✓Reviewed the practical Docker model for this site:✓Added a Docker Compose Jekyll workflow using the `ruby:3.3` image, a mounted project directory, and a Docker volume for bundle dependencies.✓Confirmed why the container listens on `0.0.0.0`, while the host port can be bound to `127.0.0.1:4000` for local-only development.✓Moved Work Log out of normal posts and into the `_worklogs` collection.✓Reframed the old About tab as a Work Log archive entry point.✓Left the personal introduction role to the pinned “Who Am I?” post.✓Added a homepage Work Archive card before pinned posts.✓Polished the Work Archive card toward a cleaner terminal-style surface with dynamic latest-month and entry-count metadata.✓Polished post-card hover behavior and sidebar social links.✓Removed RSS from the lower-left social links and kept GitHub, Twitter, and Email.✓Confirmed the three-seed rPPG Stage1 clean run completed and passed split-metadata checks.✓Resumed the main clean queue; seed `333185252` correctly skipped completed Stage1 work and entered Stage2 P0.=Keep Work Log as a public archive collection instead of mixing it into the normal post stream.=Keep the site development workflow reproducible through Docker Compose rather than depending on host Ruby state.=Bind local Jekyll preview to localhost for normal development.=Keep exact experimental scores out of the public worklog unless they are part of a deliberate public result table.=Treat the current clean p128 queue as the formal path and keep old exploratory results separated.!Stage2 is now running, but `training/results_p128_clean.csv` has not been generated yet. Do not treat the clean protocol as result-bearing until Stage2 finishes and writes verified rows. The public worklog is useful as a resume surface, but it should not leak private notes, unpublished-sensitive tables, or messy job-process details.>Watch Stage2 P0 for seed `333185252`.>Verify generated result rows and split metadata before using them.>Let the main queue continue into the remaining clean protocol work.>Keep the Work Log archive card and sidebar polish in a small visual-change lane.>Continue appending date-stamped public notes while keeping raw private notes elsewhere.

WEEK 17Apr 20 – Apr 26, 20269 done · 7 decisions · 1 risk — 3/7 days

SAT 04·25

@Repository cleanup / documentation roles.✓Reorganized git, `.gitignore`, and resource directories.✓Clarified the roles of several project files:=README should explain the project to a future reader.=Temporary notes and current state should not be dumped into README.=Current state belongs in `TODO.md`.=Experiment history belongs in `run&conclusion.md`.>Keep README stable and readable.>Move temporary notes elsewhere.

FRI 04·24

@Phase 0.5 / Phase 1 monitoring.✓Monitored Phase 0.5 / Phase 1 experiments.✓Reviewed preliminary R0-style results.✓Checked the effect of scheduler patience.✓Continued evaluating whether 128-dimensional vs 256-dimensional embeddings should be compared under a cleaner protocol.·Training hyperparameters that look like minor engineering settings can meaningfully change conclusions in small-data deep-learning experiments.!One run is not decisive.>Continue comparing protocol variants under cleaner conditions.>Avoid over-interpreting any single result.

THU 04·23

@Project context system cleanup.✓Migrated and cleaned up the project context system.✓Reduced overlap between memory files, paper notes, and temporary context.✓Defined a simpler structure:=`TODO.md` should be the current dashboard.=`run&conclusion.md` should be the experiment ledger.=Long-term rules and stable project assumptions belong in `CLAUDE.md`.·Research context is infrastructure. If project memory is scattered, future work becomes slower and less reliable.>Keep `TODO.md` short and current.>Keep run results out of chat memory and inside the run ledger.

WEEK 16Apr 13 – Apr 19, 20263 done · 2 decisions · 1 risk — 1/7 days

WED 04·15

@Configuration cleanup.✓Checked whether training parameters were actually controlled by YAML configuration.✓Reviewed possible drift between YAML, CLI arguments, and code defaults.✓Moved more important training settings into config.!If scheduler, patience, and other training parameters are scattered across YAML, CLI, and hard-coded defaults, experiment provenance becomes fragile.=Important training parameters should be visible in YAML.=Future runs should be reproducible from config files.>Keep config files aligned with actual training behavior.>Avoid hidden defaults when running formal experiments.

WEEK 14Mar 30 – Apr 05, 20266 done · 4 decisions — 2/7 days

FRI 04·03

@CCT-LSTM multi-modal fusion.✓Implemented auxiliary losses for unimodal branches:✓Tested the idea that auxiliary supervision may help prevent fusion from suppressing useful unimodal signals.·The auxiliary-loss version looked more stable than the older baseline in preliminary runs. Exact scores are intentionally omitted here. This page is a public work log, not the final experiment table.=Keep auxiliary loss as a serious candidate for the formal protocol.=Do not judge it from a single run.>Re-run with multiple seeds.>Check whether the improvement is stable across folds and seeds.

THU 04·02

@CCT-LSTM reproduction / protocol review.✓Reviewed the current state of the CCT-LSTM reproduction.✓Re-checked the original paper’s likely training flow.✓Noted that some previous experiments may not be strict faithful reproductions.✓Started moving run results out of scattered memory/context files into a dedicated run ledger.=Treat faithful reproduction and method extension as separate tracks.=Keep experiment results in an explicit result ledger instead of assistant memory or temporary notes.·The freeze vs no-freeze question matters because it changes both interpretation and compute cost. If the encoders are frozen after unimodal pretraining, the reproduction means something different from initializing them into the multi-modal framework and continuing training.>Compare freeze vs no-freeze behavior.>Keep run conclusions in `run&conclusion.md`.