research-note · a record framework

research‑note

The agent computes and writes the record. You read it and decide.

01

Every calculation counts.

A paper reports the one tunnel that reached gold. A record keeps them all.

Tunnels are objectives, gold is a fact. The gold was always in the rock — the tunnel only reached it. A dead end is a result, and it stays on the map.

02

You judge. It does the rest.

It runs, watches, collects, concludes — then stops one step short.

read what is known→ run what is missing→ record the evidence→ choose the next step

an entry

### Entry: 3×3 TFIM ground-state scan
- Id:          01.002
- Status:      closed
- Motivation:  how does energy error change with D?
- Recipe:      open TFIM; dt=0.05; second order
               D in {2, 4, 8}; QuSpin ED reference
- Facts:       D=2  6.77e-4
               D=4  1.34e-4
               D=8  1.51e-4
- Conclusion:  fivefold gain at D=4,
               then a shallow plateau

the objective that asked for it

## Goal
Compare Quimb simple update with QuSpin ED.

## Plan
- [x] P1 - do both codes use the same model?
      01.001 convention gate - passed
- [x] P2 - does the error shrink with D?
      01.002 ground-state scan - closed

## Verdict
The scan improves from D=2 to D=4, then
becomes non-monotonic at D=8. This recipe
supports an error floor near 1.5e-4.

These numbers come from the included 3×3 TFIM example, which records both the ground-state comparison and real-time dynamics.

03

Distilled for reuse. Lossless for history.

Validated code can move into the project library, and reviewed conclusions into the wiki. The original record stays intact.

THE COMMONS where it goes next knowledge and tools distilled — one merged copy, travels far src/ wiki/ skills/ motivation, fact and verdict kept whole — one lane each, context attached entry log code@sha entry log code@sha YOUR PROJECT where it settles first src/ lift wiki/ wiki-compile skills/ memory-lift what actually ran — entries, objectives, and the code that produced them merge — lifted by a person append — reviewed by a person
Knowledge and tools — built in common, for the next investigation. Motivation, fact and verdict — kept whole, for the next analysis, and a seed for questions nobody has asked yet.

04

Every close is a label.

Question, procedure, evidence, verdict — and a person's acceptance of it.

open→ running→ analyzedagent stops→ closedyou

Each reviewed calculation adds a judgment the next investigation can build on.

A reviewed entry is finished work — hand it over, cite it, or train on it.

a typical run

One run, end to end.

YOU hand over a goal correct it by hand, or ask the agent close it ask for a lift any step, not just the end THE AGENT plan objective · plan run hpc-submit analyze analyze · correct distil lift · wiki-compile memory-lift the page decides what runs next what is running, unclosed, needing attention; refreshed on every record write status — DASHBOARD (OPTIONAL) look any time check every page on each edit and commit, and keep one format LINTERS · FIXER (OPTIONAL)
You set the goal and accept the result. The agent plans, runs, analyzes, and records the work. Linters keep the record consistent; the dashboard and background fixer are optional. At any point, you can correct the plan, close an entry, or ask to promote a useful result.