Keep the original words.
Every nonempty source line has an occurrence ID. Nothing is silently omitted, merged, or rewritten into an experimental instruction.
THE ATTENTION DESK / EVIDENCE FIRSTKeep the context.
Keep the context.
Find your next step.
Source-linked laboratory notes. Visible uncertainty.
A clear handoff when you return.
Recorded Gemma evidence · live browser decisions · manual photo notes
RESEARCH PREVIEW · NO EQUIPMENT CONTROLSTART WITH A · INSPECT B · EXPLORE C
A source-supported wait. Inspect the evidence and manually confirm a ten-second demonstration.
YOUR PROTOCOL STAYS YOURS.ORIGINAL SOURCES · VERSIONED DECISIONS · HUMAN-CONTROLLED TIMERSNO EQUIPMENT CONTROL
YOUR ATTENTION DESK
NO ACTIVE EXPERIMENTYour attention desk.
The bench notes
ORIGINAL WORDS / NO GENERATED INSTRUCTIONSFocused source-pair analysis preserves headings and original IDs. Unresolved targets and model abstentions require review. See the length-stratified evaluation; inspect every suggestion.
Up to 48 nonempty lines. Only synthetic or approved non-sensitive notes. Input is kept in browser/server memory.
The Attention Map
SOURCE ORDER · NOT A PREDICTED START SCHEDULEGood science.
A little breathing room.
Every source line stays in view. No invented durations. No silent missing steps.
—Source lines
—Fixed windows
—Source minutes
—Review lines
Ask BenchSloth
03 / INTELLIGENCE · EVIDENCE FIRSTRecorded Gemma output
Ask what needs attention, which note belongs where, or what the source leaves unanswered.
Current browser facts. Deterministic, no new model inference.
Gemma selects evidence; the source stays verbatim. Answers cannot start timers or change your plan. A valid citation does not prove a correct interpretation.
KEEP THE EVIDENCE, THE EXCEPTIONS AND THE REASON YOU PAUSED.Reload clears browser state. Use approved equipment timers.
READUNDERSTANDVERIFYPLANRESTRETURN
Keep the evidence.
Make room to think.
Know what needs attention.
Ask Gemma to find the original evidence behind a question. Inspect its selection, uncertainty and real request before accepting an interpretation.
Your judgment stays central.
Track supported numeric source durations after explicit review and manual start. Range choices and monitoring remain visible; a countdown never authorizes unattended work.
Return to the source.
A calm countdown keeps the supporting passage and next source line in view. It never means required monitoring can stop.
Trust is a method.
Not a decorative badge.
This is an experimental research organizer, not a validated protocol or laboratory controller. Prepared results are labeled; real inference uses local Ollama. BenchSloth has not yet demonstrated human time savings. Inspect the frozen evaluations and limitations in the project README before relying on a classification.
Does BenchSloth Actually Work?
NEW / 24 frozen semantic questions: combined candidate15/24 exact evidence answers versus14/24 rules, but13/24 exact QA workflows versus14/24 and8 unsupported evidence selections versus7. Retrieval's top3 contains all required evidence for16/16 answerable questions. No reliable model advantage is established. Model relationships remain experimental proposals requiring review.
Separate unchanged v0.5 planner:20 cases,22 reference links,1 extra unresolved association,16/18 eligible waits,0 observed unsupported offers. Strong scoped rules tie. Independent human n=0; no measured time savings.
Explore every case, answer and failure ↗ · Optional core source ZIP ↗
Human trials completed: 0. Research preview; source interpretation still needs human review.
Remember what you saw, too.
A quick photo can help you return to a complicated task. Attach a picture to the instruction you're reviewing, jot down what matters, and export it for your own records. Photos stay in this browser tab until you export them. No upload, image recognition, or scientific judgment is performed.
Only JPEG, PNG or WebP, up to 8 MB each. Images are resized to a maximum 1200 px dimension, re-encoded as JPEG to remove embedded metadata for a quick preview, and kept in memory—not cloud storage. This is a lossy preview, not an original scientific image. Keep raw instrument files and acquisition metadata in your approved research storage. Exported JSON includes the preview photos, so store it privately. Five photos maximum.
No photos attached yet.