research-topic — the topic-selection axis

Integration: the microscope, simulation and librarian axes this sketch connects are being integrated, together with a bridge, as soft-matter-agents. The topic axis is not part of it — that system "does not choose research topics" (plan.md §1).

Important

IDEA SKETCH / CONCEPT DOCUMENT

This repository is a design proposal for one axis of a four-part system. It contains design documents only — no code, no knowledge-base entries, and no running pipeline. The value axes, schemas and literature lists written out here are discussion material, not decisions; construction is to be done by someone else (§5), and all of it is subject to change. What is fixed and what is open is separated in §0.

The topic axis. It supplies agentic-microscope (experiment) and Brownian-Dynamics-Agent (simulation) with what to work on, and takes back what those two produce as evidence.

The goal is the four of them running as one multi-agent system — each stage handing the next the knowledge it needs, in a form the receiver can actually consume. The two working repositories already run, and each already keeps a knowledge base. What did not exist is the exchange between them: neither repository references the other, so a result BD establishes is not readable by the microscope, nor the reverse. → §2.1

Custody of that exchange has moved out of this repo. librarian-agent is the fourth axis and owns J1 — the store, the index, and retirement. It is built for the microscope so far: six read tools, no knowledge migrated. This repository keeps J2 (which topic) and J3 (the form of a pass condition), which is where §1 now stands.

Drafted 2026-08-31 · status sketch — no code. An idea sketch only.

Project status

Design documents in design/, and nothing else. No src/, no kb/, no tests — those directories are not empty, they do not exist.

Written	Boundaries between the three repos · six philosophy items · an idea log of 46 ideas, 9 conflicts, 10 questions (6 still open) · the seven exchange points, checked against the live repos → §2.1
Drafted, expected to be replaced	3 value axes / 6 poles · KB formal-object schemas · literature entry points
Empty	The value conditions of every pole. Only the pole names exist
Not started	Everything in Phase 1 and after → §4
Depends on this repo	Nothing. All three other repositories run without it
Verified against the live repos	2026-09-05 — what each of the other two actually keeps, and where the exchange would attach → §2.1. Counts re-measured 2026-09-07 against BD d47c319 and MS 2e0118a: five were stale and one was counting the wrong thing → §2.1

The nearest thing to a deliverable is the 2026-09-03 discussion (§6), whose purpose is to make the four rows of §0 fillable by the person who will build this.

---

0. What these documents are — read this first

This repo is an idea sketch. Nothing is built, and J1/J3 will be built by Prof. Sho Takatori directly (§5).

So the documents below hold three different kinds of thing, and reading them as one kind is the mistake to avoid.

Status	What	Where
Decided	The boundaries between the three repos · the six philosophy items · what this repo does not do	charter.md · §3 · §9
Registered only	Ideas, conflicts, open questions. Not resolved	ideas.md
Discussion material	Value-axis assignments · schema fields · literature lists. Not decisions	lineages.md · kb-schema.md · _common.md

The third row is the easy one to misread. The schemas and axis assignments are written out concretely not because they are settled, but so that the discussion has something specific to push against. The T-013 procedure has never been run once, and much of the literature is unverified by DOI.

What the builder is expected to change:

What	Current state	Who decides
What the value axes are, and how many	Decided 2026-09-05 (T-051 · T-054). Three dimensions — trend · difficulty · impact — each restated to anchor in domain literature: X1 maturity · X2 obstruction · X3 reach. Still two poles per axis	Takatori (agenda 3) — the user, on both. The specific restatement is derived and is still a draft to push against
The value conditions of each pole	Two of six written — T-013 was run on X1 maturity on 2026-09-05 (T-055) → V1 frontier · V2 unfinished. The run came back split: V1's trajectory check is consistent, V2's is not_run. Four poles are still names	Takatori (agenda 1–2)
The fields of the KB's formal objects	Example JSON. Only the principle is fixed	Takatori (J1 · J3)
The form of a rigor axis's pass conditions	A draft back-derived from BD's A1–A10	Takatori (J3, agenda 6)

The only things that should not be changed are the six philosophy items in §3 and the boundary table in charter.md §3 — both were either inherited from the two existing repos or derived from where those two independently converged, so they are not this repo's to change alone.

---

1. The jobs, and which of them is still this repo's

#	Job	Product	Consumer	Owner
J1	Accumulate the knowledge base	KB entries as formal objects + a search index	BD · microscope	librarian-agent — moved
J2	Select scientific topics	Topic candidates in falsifiable form, with the basis for their value	BD · microscope · a human	this repo
J3	Manage the definitions of rigor	Axis definitions · the form of pass conditions · the canonical-source registry	BD · microscope	this repo

J1 moved because custody is a job, not a by-product. A store that only accumulates goes stale — BD I-076 — and this repo had no mechanism for the other half. The seventh entry kind, challenge/ (→ kb-schema.md §4.7), is implemented there; the other six schemas stay here, since the form of an entry is J3's business and its contents are not.

What that does not change: this repo still holds form and provenance, and holds neither values nor execution (charter.md §3). The librarian inherits the same prohibition and adds one — it originates no numbers, sets no thresholds, and may not settle a challenge whose falsifier it cannot run.

J3 is BD's I-066 lifted one level.

research-topic   = the source of criteria     what the rigor axes are, the form of a pass condition, the canonical sources
BD · microscope  = the enforcers              each runs those axes as code in its own domain

Definitions here, enforcement there. This is not a migration — BD's A1–A10 and the microscope's G1–G32 stay where they are. What comes here is only the part the two repos arrived at independently. → charter.md

---

2. The loop

                    +-------------------------------------+
                    |           research-topic            |
                    |  survey -> topic candidates         |
                    |  personas ask questions (BD:I-050)  |
                    |  holds form . provenance            |
                    +--+-------------------------------+--+
                       |                               ^
   (1) topics + falsification conditions               |
   (2) literature + conditions of validity      (4) results
   (3) rigor-axis definitions  (J3)             (5) dead ends
                       |                        (6) questions
                       v                               |
     +-----------------+--------+       +--------------+-----------+
     | Brownian-Dynamics-Agent  |       |    agentic-microscope    |
     | simulation               |       |    experiment            |
     | A1-A10 . deterministic   | <---> |    8 gates . G1-G32      |
     | gate                     |  (7)  |    BLOCKED by default    |
     +--------------------------+       +--------------------------+

   (1)-(7) are the seven exchange points -> sec 2.1

   ...................................................................
   challenge/  runs along every edge above, in BOTH directions:

     any node --- cites the TARGET entry's own falsifier ---> any node

     and the falsifier's type -- not the sender -- picks who runs it:
        a measurement -> microscope . a run -> BD . a condition -> here

   So BD and the microscope send work here too.  No owner in
   charter sec 3 changes; only the direction stops being fixed. -> sec 2.2

   ...................................................................
   librarian-agent is ON THE RETURN PATH, not outside it.  J1 moved
   there, so edge (2) originates in it, and (4) and (5) land in its
   custody -- J2 reads them from there rather than from BD and MS
   directly.  It asks no question and returns no result, and it still
   decides what J2 sees, because a search that ranks is a search that
   selects.  -> github.com/kyu-softmatter/librarian-agent

The hypothesis: the four axes pull each other along, forming a positive feedback loop.

Four, and the fourth is on the return path. J1 moved to librarian-agent, so edge (2) originates there and edges (4) and (5) terminate there — J2 reads a result or a dead end out of its custody rather than out of BD and MS directly. It asks no question and produces no result, and it still selects what J2 sees, because a search that ranks is a search that selects. So it is inside C-001 rather than exempt from it, and its own design says so: the retrieval-feedback store is where this loop closes most tightly, and the only damping named for it is human approval → librarian FEEDBACK.md §5.

The objection this repo carries: a positive feedback loop either diverges or confirms itself. If the four axes feed each other, whatever bias they share gets amplified. The loop needs damping, and one piece of it has to come from here — topics go out only in a form BD and the microscope can falsify. → ideas.md C-001

J1 moving added a second place the loop can close, and a second damper. A store that decides which evidence surfaces is an amplifier whether or not it asks anything: what a search ranks, it selects. librarian-agent names the tightest form of that — an agent promoting its own retrieval results to ground truth — and damps it with human approval plus reproduction under two distinct index states. Neither damper covers the other's path, so C-001 now has two answers rather than one, and both are partial.

2.1 What is exchanged at each stage

The arrows above are the goal. What follows is what each arrow would have to carry, and what is actually in place.

Three rows changed hands when J1 moved. Row 2 is librarian-agent's to send, and rows 4 and 5 are its to hold; this repo reads them from there. The research-topic in those three cells is where the knowledge is consumed, not where it now sits.

The counts below were re-measured on 2026-09-07 against BD d47c319 and MS 2e0118a, and they carry their definition so they can be recounted: source/papers/*.md excluding INDEX.md, wiki/**/*.md including the contract, runs/*/record.json (227, in 256 run directories — the directory count is the one that misleads), and folder entries excluding _-prefixed templates and indexes.

Three of the corrections are worth naming. source/papers/ was 44 here and is 42 — while BD's own INDEX.md states 40, so the same folder had three counts and no two agreed. wiki/benchmarks/ was 5, which is the number of benchmarks benchmarks.yaml marks blocked; the row describes the ones already running, and that is 7. And wiki/findings/ was the file count, which includes _TEMPLATE.md.

The Exists today column is read off the two live repositories on 2026-09-05; everything in Missing is unbuilt. The two working repos have moved well past the snapshot charter.md §2 was built from (T-041), so this table supersedes that one on the question of what BD keeps.

#	Stage	The knowledge that has to cross	Exists today	Missing
1	research-topic → BD · MS	A topic candidate in falsifiable form — with falsifiable_by naming which of the two can kill it	The receiving slots exist. BD takes a case as intake/<case>/{sketch,system.yaml,observation.yaml} (8 cases); MS takes a research goal	Everything upstream of them. Nothing produces a topic. Both are fed by hand
2	research-topic → BD · MS	A published number with its conditions of validity — the part that appears only in the original paper	BD: knowledge/source/papers/ — 42 distillations carrying doi · read_depth · provides · used_by · lab_authored. MS: kb/literature/ — schema, template, and zero entries	The two formats do not meet. BD has 42 entries MS cannot read; MS has an empty folder BD cannot fill. This is the sharpest gap in the system
3	research-topic → BD · MS	Rigor-axis definitions and the form of a pass condition (J3)	Each runs its own: BD .claude/rules/ (4 files) + A1–A10; MS 8 gate modules + G1–G32	No shared form, so the same axis can be written two ways and neither notices → T-004's revisit condition
4	BD · MS → research-topic	Results and counterexamples	Dense on both sides. BD: 227 runs/*/record.json post-mortems + l4.json/metrics.json; MS: kb/calibrations/ (4), kb/decisions/ (20)	No consumer. Nothing reads them for topic selection
5	BD · MS → research-topic	Dead ends — cause, not symptom	BD already writes them: knowledge/wiki/findings/dead-end-*.md, with a mandatory ## Prevention section	Custody here is empty, and BD's dead ends are not indexed anywhere outside BD
6	value personas → the gate	A question plus what would close it	BD already implemented this — knowledge/wiki/questions/: "open questions, with what would close them." Independent arrival at T-008	Two entries, BD-internal. No cross-repo question store
7	MS ↔ BD	A number measured on one setup, reused on the other	MS solved this within itself: docs/03-cross-system-transfer.md separates what transfers as-is from what needs recomputation	Nothing crosses between repos. The hardest row, and the one neither repo can do alone

Row 2 is where to start. It is the only row where one side holds the content and the other holds an empty slot shaped to receive it, so it can be tested without either repo changing what it does.

The convergence that makes row 2 tractable

The three repos independently arrived at the same rule: a published number is unusable without the conditions under which it holds, and it never counts as a measurement.

Repo	How it says it
research-topic	literature.conditions — "when this expression holds — the part that appears only in the original paper" → kb-schema.md §4.1
microscope	Transfer conditions is a mandatory section, and a literature value is always assumed, never measured, so it cannot advance a verdict. "An entry that cannot say what would have to hold for its number to apply here is not usable, however good the paper is"
BD	Provenance is from_paper · from_knowledge · assumed · derived plus a tier, and "a derived value is recomputed and compared, never trusted as written." The cost of getting this wrong is recorded: T = 300 K mislabelled as measured propagated a −4 % to −14 % error into every τ_B downstream

This is a fourth convergence, and it was not in charter.md §2. It matters more than the count: §2's table established that the three agree on shape, and this one shows they agree on the field the exchange actually turns on. Row 2 does not need a format negotiated from scratch — both sides already demand the same thing of a literature entry.

What this does not settle. BD keeping a real knowledge base removes one of the three grounds for J1 (T-003 argued a third place was needed because the microscope's KB is instrument-bound and BD had none). The ground narrows rather than falls: BD's knowledge/wiki/ is simulation-bound the way MS's kb/ is instrument-bound — systems/ cards, concepts/ in BD's own conventions — but source/papers/ is not, and neither is MS's kb/literature/. The shared layer is the literature layer, and that is exactly row 2. → Q-002

2.2 Collaboration, not dispatch — challenging a claim

The seven rows above still run one way at the criteria level: definitions and topics go down, results come up. That is management, not collaboration. charter.md §5 forbids this repo from becoming an enforcer — but nothing in it says the other two cannot send work here.

The missing primitive is already in all three repos: every KB entry carries its own falsifier. That is philosophy ③, and charter.md §2 records it as one of the six things BD and the microscope arrived at independently.

Where	The falsifier that is already there
MS kb/expertise/ (6)	each entry carries the observation that would retire it
MS kb/literature/	## Falsification conditions is mandatory — "the first is always the local measurement that would replace this" — and "what sits in this folder is exactly what is worth measuring next"
BD knowledge/wiki/findings/ (22)	## Scope / limits — where this stops being true
BD knowledge/wiki/benchmarks/ (7)	known-answer systems already running as regression tests — benchmarks.yaml also lists 2 ordinal and 5 blocked, and 5 was the number this row used to carry
BD knowledge/wiki/questions/ (2)	open questions, with what would close them

A falsifier is a work order that nobody executes yet. Raising doubt about a claim is asserting that its falsification condition may now be reachable; verifying it is running that condition. And the type of the falsifier decides who does the work — which is what turns the hierarchy into a graph:

The doubt	Falsifier type	Routes to
the published number does not transfer to this setup	a local measurement	microscope
the claim is contradicted by what the model actually does	a run	BD
the conditions of validity were never checked against the source	literature	research-topic
the claim was used outside the scope it stated	none available	a human (not_resolvable)

BD and the microscope become senders. No row of charter.md §3 changes — the microscope still owns instrument values, BD still owns run provenance, this repo still holds form and custody — but the direction of a work order is no longer fixed.

Three rules it has to keep.

1. A challenge is a work order, not an argument. It must cite the target entry's own falsification condition. An entry with no falsifier cannot be challenged, and that is a defect in the entry rather than a limit here. This is also what keeps the mechanism clear of _common.md §3, which forbids a persona rebutting another persona's basis: a challenge is settled by a run, a measurement or a search, never by the better argument.
2. The verdict is a result, not a vote — BD I-052. upheld / rejected is set by what the run or the search returned, and not_run is the default (I-055).
3. It is bounded. A challenge carries resolvable_by (T-034's damping device), a cost, and a depth — a challenge against a challenge increments it. Without the bound the loop circulates doubt instead of topics, which is C-001 wearing a new costume.

This is the damping the loop was missing. §2 raises the objection that a positive feedback loop amplifies whatever bias the four axes share, and asks for damping without naming a mechanism. A channel in which any axis can invalidate another axis's input is negative feedback, and it is the first mechanism in this design that is one.

This already happens, by hand. MS kb/expertise/oil-objective-trapping-in-water is a challenge that was upheld. A human observed on 2026-08-18 that oil objectives do trap in an aqueous sample; the entry records that "that observation is correct and it corrected this project: lens 7 had been refusing those objectives outright." It carries evidence: measured, review_after: 2027-08-18 and supersedes: null — a resolution, a retirement date, and a supersession slot. The mechanism is not being invented here. What is missing is only that it runs inside one repository instead of between three.

Not to be confused with question/. A question/ entry is a value persona asking something about a topic candidate, and it dies when the literature turns out to answer it. A challenge is aimed at a claim already in the KB, and it dies when the claim's own falsifier is run. Different target, different death condition. → kb-schema.md §4.2

Where the challengeable surface actually is. BD's entries/ store is 135 entries distributed 52 tooling · 45 method · 25 handbook · 10 intake · 3 paper — BD's own note on that table is that the tooling-to-paper gap "is widening." So most of what is recorded is our machinery misled us, not the literature said something. The scientific claims worth doubting are in source/papers/ (42), wiki/findings/ (22) and wiki/benchmarks/ (7) — not in entries/. A challenge system pointed at the wrong store would find almost nothing to do.

Where this belongs in the three jobs. Not a fourth job — it is the half of J1 that the word "accumulate" leaves out. A knowledge base that only accumulates goes stale, and BD I-076 already named that failure: all the files are there and search finds nothing. Custody has to include retirement. kb-schema.md needs a seventh entry kind, challenge/, and it is not written.

---

3. Philosophy

Six items. The first four are inherited from the two existing repos; the last two originate here.

① The default is failure, not passing — inherited

BD I-055, and the microscope's BLOCKED. With no basis for a verdict, there is no pass. Here that reads: the default for "this topic has value" is "unknown," and no code path exists that moves to "yes" without a basis.

② The LLM does not originate numbers — inherited

BD: personas name the criteria, code settles them. Microscope: the subagent supplies only the half of a judgment that has no closed form, and originates no number. Two repos reached the same rule independently. Same here — a value persona says where to look, and the search result says what is there.

③ Every judgment carries the check that would overturn it — inherited

BD I-065, and the per-entry falsifiers in the microscope's kb/expertise/. Here that reads: each question carries "if this already has an answer, where would it be?" Without it a literature survey has no stopping condition, and then the search budget becomes the value judgment. → T-008 · T-023

④ Natural language is not state — inherited

BD's I-133 antipattern. This is exactly where J1 dies. KB entries are formal objects of the shape {topic, source, conditions of validity, falsification condition, state} — not summaries written by an LLM.

⑤ In rigor, disagreement is a bug; in value, disagreement is the product — new

BD divided by axis rather than by person on the grounds that "different axes give orthogonal verdicts, so conflict never arises in the first place" (I-067).

Value is not orthogonal. A theory/minimal-model lineage holding that something is uninteresting because it does not reduce to a dimensionless group, and an experiment/emergence lineage holding that it is interesting because a new state shows up in the lab, are true of the same system at the same time.

So this repo inherits seven of BD's eight prohibitions and drops one: "no rebutting another persona." In its place it fixes the form — not rebuttal but a divergence record: when two personas attach opposite value to the same target, the point of divergence itself becomes a KB entry. Topics with large divergence are the first-rank candidates.

Structural consequence: BD has one persona per axis (orthogonality is the goal); here there are two per axis — the two poles (disagreement is the goal). Convene one pole and no divergence exists. → T-018

This is the most dangerous decision in the repo. → C-002 · C-005

⑥ The canonical source of value is not a textbook — new

BD's I-068 moved the distillation entry point from names to 27 textbooks, on the grounds that "a school's pass criteria are written down explicitly in the textbooks." Value has no such grounds — "what is interesting" is not written in textbooks.

Where it is written, in three places:

Source	What it gives
The Outlook / Open questions section of a review	What that lineage holds to be still unsolved
The trajectory of the body of work (in time order)	What that lineage actually picks
What was not done	An adjacent area left untouched — the most valuable signal, and the hardest to read

The third is this repo's best product and is indistinguishable from a distillation error. → Q-005

---

4. The plan

Phases are ordered, and each carries its exit condition. A phase without one does not end.

The corpus arrives after the system. Papers and books are to be supplied later, so every phase below has to be completable without them — which is the order kb-schema.md §0 and T-030 already argued for on independent grounds (fill before the form is fixed and you throw away what you filled). Two consequences.

① The supply layer's first job is ingest, not crawling. Accepting a hand-supplied document is the path that exists on day one; the forward-citation index (Phase 3, Q-006) is what a corpus needs after it arrives, not what gets it there.

② A book never enters the repository. BD already solved this — an origin: handbook entry's source reads distillation#section ← [short-name] p.page, so "a claim can be walked back to the page it came from without the book being in the repository." That is the contract for the books to come, and it is already running for two.

Phase 0 — sketch (now)

Task	Exit condition
Fix the boundaries between the three repos	charter.md §3 is filled with no item owned twice
Register ideas, conflicts, questions	ideas.md — register, do not resolve
Read the microscope's kb/expertise/	✅ 2026-09-05 — Q-002 closes, on neither branch it offered. The contents are instrument- and lab-bound (n_medium = 1.333 for this lab's samples, this nosepiece's oil, applies_to_systems: [current, …]), so §2③ holds and J1's basis survives. But the frontmatter is domain-neutral — question · source · expert · confidence · evidence · scope · applies_to_systems · review_after · supersedes. The form transfers although no value does, and the form is J3's business, not J1's
Re-derive charter.md §2 from the BD remote	◐ partially, 2026-09-05. T-041 was right: one row is now false — BD's knowledge base is built, not raw (44 wiki pages · 135 entries · 227 run post-mortems) → §2.1. The other rows are not re-derived, and §2's caveat 2 still stands

Phase 1 — prototype the two poles of one axis

Why a pair and not one: philosophy ⑤ (disagreement is the product) and C-005 (the divergence schema) cannot be tested without a pair. Same method BD used when two worked examples validated its design.

Choosing the axis is not this repo's call (§0) — agenda 3.

Task	Exit condition
Pick one axis and fix its two poles	One real case exists where the two poles attach opposite value to the same system
Write each pole's value conditions	Per T-013 — value conditions → review Outlook and trajectory → check consistency
Draft the divergence-record schema	The closing condition for C-005
Run the trajectory check once	T-014 — does a pole's claim match the actual trajectory of its literature?

Phase 2 — interlocks (before any KB content)

The order looks backwards on purpose. BD's I-075 and I-076 already diagnosed the two ways a KB dies — ① nobody fills it (recording is always the last step and always the skipped one) ② the index goes stale, so the files are all there and search finds nothing. This repo is the KB, so both failures are fatal.

Task	Exit condition
capture-gate — block session exit with no record	One week in warn mode, then blocking
Force the search index to refresh	A cycle cannot end without an index refresh
Settle the KB formal-object schema	kb-schema.md — zero places where a natural-language field is used as state
What J1 may write into a public repo	T-044 · T-019① — the rule exists and is enforced before the first topic assessment lands. This repo is public, and deleting the file afterwards does not undo the disclosure → §8

Phase 3 — the J2 pipeline

Task	Exit condition
Keywords → literature survey	Inherit BD I-080 (author-adjacent, then system-adjacent) and I-082 (citation graph)
Wire up a forward-citation index	Q-006 — OpenAlex / Semantic Scholar. J2 cannot run at all without this
The question gate	BD I-052 — a question's fate is decided by whether a search returns anything, not by an LLM vote
Emit topic candidates	The output is in a form BD and the microscope can falsify (T-034, damping)

Phase 4 — integration

→ §7.

Not in the plan yet

• Retrospective validation of value. C-003 — whether a topic was worth doing is known years later. The topic scorecard (T-033) is the only grounding, and early in the loop there is none. This is not deferred; there is no answer yet.

---

5. People and roles

Axis	Participants
BD agent (repo)	Dr. Takuya Kobayashi · Saksham Malik
agentic-microscope (repo)	Kyu Hwan Choi · Saksham Malik
research-topic — KB · rigor definitions (J1 · J3) — built directly	Prof. Sho Takatori · Kyu Hwan Choi
Overall management	Kyu Hwan Choi
Funding · advising	Prof. Sho Takatori

Why there is no lead column

Everyone named in a row is a participant in that axis, and how the work divides inside it is theirs to settle rather than this document's to assign. The table records who is on which axis — nothing about seniority, and nothing about who reports to whom.

It also carries no commitment status. An earlier draft tagged every row with one, which was the right hedge while the rows were a proposal. They are not a proposal now.

Where the roles overlap — a structural warning

If the person authoring J3 (the rigor definitions) is also a canonical-source author of a value lineage, the one automatic check in T-014 weakens.

trajectory_check — does a lineage's value claim match the actual trajectory of its literature? — carries information only when the value conditions were back-derived from the literature independently. Written directly, the check approaches guaranteed-pass and its information content drops.

This shape is already named. BD personas/_common.md §8:

If a benchmark's input overlaps the model's input, that check is guaranteed to pass and carries zero information. → trivially_satisfied: true

Removing names (T-018) cut this risk sharply — a persona is no longer a specific individual, so the overlap is not structural. It did not disappear: the V1 (reduction pole) literature of axis X1 contains his own canonical papers.

The split:

What	Who	Why
J3 — the rigor definitions (axes · the form of pass conditions · conditions of validity from the sources)	Authored directly	Rigor is strictly better authored by the authority. No overlap problem
V1's value conditions	Back-derived from literature independently	His answers are used as a held-out comparison only

So there are now two reasons, not one, to write the V1 control draft before Thursday: ① so the discussion is a test (§6) ② to avoid trivially_satisfied. → T-016

Saksham Malik spans two axes

This is the real pressure behind the integration question (Q-008). The cost of two sets of criteria has been abstract so far; with a person on both repos, a person pays it — learning the same thing twice in two formats. It may pull §7's decision point earlier.

---

6. Discussion with Prof. Sho Takatori — 2026-09-03 (Thu)

Two reasons this discussion is part of the design.

① Handing over construction. J1 and J3 get built directly by him (§5). This repo is a sketch, so the first purpose of the discussion is to make the four rows of §0's "what the builder is expected to change" fillable.

② The only cheap attack on C-003.

A value lineage has no ground truth — there is no way to check whether the distillation is right. The exception is a lineage whose canonical-source author can be asked.

The canonical sources of A5.E8 (swim pressure, PRL 113, 2014) and A6 (nonequilibrium quantification) sit in the V1 (reduction pole) literature of axis X1. So for exactly one of the six poles a comparison path exists. The distillation method itself gets tested on one case; the other five poles have no such path.

Agenda — in priority order

#	Item	IDs	Why it has to be him
1	"What does a problem need before you pick it?" — can the value conditions be stated explicitly?	T-013 · C-003	The answer key for distillation. Comparable against what the literature yields
2	Test philosophy ⑥ — is it true that value lives in review Outlooks and work trajectories? Do his own choices actually show up in those two places?	T-012	If this is false, all of Phase 1 collapses
3	What actually divides active matter — throw away the 3-axis / 6-pole draft in lineages.md and set it again	Q-004 · philosophy ⑤	The real dividing lines inside a field are not on the surface of its literature. §0's first two rows get filled here
4	C-004 — how do you tell "nobody asked" from "nobody asked because it is boring"?	C-004 · T-007	This is senior tacit knowledge itself. No automation path is visible
5	Q-005 — distillation error, or a real gap?	Q-005 · T-014	When a pole says X is interesting but X is absent from its literature
6	Review the pass conditions of rigor axes A5 and A6 — existing BD output, checked by the canonical author	BD A5 · A6	The actual starting point for building J3. The conditions field of the source registry comes from here
7	Q-009 — will he take J1/J3, and what happens with funding?	Q-009 · §5	Authoring J1/J3 and the funding line rest on the same person. What stays valid if declined is not written down
8	Is removing names enough? — lineages replaced individuals, but V1's literature overlaps the group's own output, so the reading list points anyway	T-019 · §8	The real ceiling on disclosure. Feeds directly into the C-006 integration call

To prepare beforehand

• A draft of V1's (reduction pole) value conditions, back-derived from literature alone. The control for agenda 1. It has to exist before asking, or there is no comparison.
• Two or three candidate axes for agenda 3 — on the assumption they get discarded.
• The BD A5 and A6 persona files (existing output, something real to show). Material for agenda 6.

Careful: ask agenda 1–2 first and the answer drags the design with it. Walking in without the control draft turns the test into dictation.

Agenda 3 is the opposite — the axis assignment is not this repo's to decide (§0), so the draft is not defended. It is carried in to be thrown away.

---

7. Integration with the two existing agents — under consideration

The current decision is "index, do not merge" (T-004). The grounds are BD I-057 — a boundary is drawn at is it reversible and checkable. Indexing is reversible; migration is not.

What changed on 2026-09-05: the seven exchange points are now written down against the live repositories (§2.1), so this section is no longer choosing between three shapes in the abstract. Row 2 is testable under option A — BD holds 44 literature distillations, the microscope holds an empty folder built to receive exactly that, and indexing the two is reversible. Whether the remaining rows need B or C is a question about rows 3 and 7, not about all seven at once.

But integration is open. Three candidates:

Option	Shape	Gains	Loses
A. Three repos + indexing (current)	This repo points at the other two	Reversible. Neither repo is touched	Risk of two sets of criteria. Cross-references are manual
B. Monorepo	All three in one	One set of criteria. References never break	The disclosure obstacle is gone — all three repos are public, and this one stays public by decision (§8). What is left against B is timing, not scope
C. This repo as the orchestrator	The other two as submodules, the loop runs here	The loop becomes real	This repo becomes an enforcer and violates its own structural prohibition (charter.md §5)

B's obstacle was disclosure scope, not engineering — and that obstacle is gone. agentic-microscope became public on 2026-08-28 after the vendor material was stripped, and this repo is public by decision as of 2026-09-04 (§8). "Put them in one repo and the stricter side wins" needed two different scopes. There is one. C-006 is dissolved, not relaxed. → T-044

Which does not make B the answer — it exposes the quieter argument. Disclosure was the loud objection, and it was doing the work of looking like the only one. T-004's actual grounds never mentioned it: indexing is reversible, migration is not. Nor does the decision point below.

The sequence, for the record. Removing names on 2026-08-31 (T-018 · C-006) took away the largest argument and moved B from impossible to discussable. T-019's two remaining arguments held the ceiling there. The 2026-09-04 decision removed the ceiling by settling the scope question, not by answering T-019 — both of its reasons survive as constraints on content (§8). So B is now possible on disclosure grounds and undecided on every other one.

Decision point: after Phase 1, once it is clear whether divergence records come out in a form BD and the microscope can actually use. Integrating before that fixes the structure without knowing what is being integrated. → Q-008

---

8. Disclosure scope — public, and what that costs

Decided 2026-09-04: this repository stays public. It has been public since it was created on 2026-09-01, and an earlier draft of this section described it as private, which was never true of the repository as published. The decision settles visibility. It does not retract T-019's two reasons — each one becomes a constraint on content instead, and one of them is not yet written down. → T-044

The original reason for private, and how it was already removed

In the first draft a persona distilled a named researcher's research philosophy. That produces sentences like:

"To Prof. X this system would look like a problem whose answer is already known."

A value judgment attributed to a named researcher becomes a matter of reputation once public. BD's roster of ~200 is entirely real people, many of them living. agentic-microscope could go public because its lenses were subsystems (optics, detection, compute), not people.

On 2026-08-31 this was removed at the definition stage rather than the output stage (T-018 · T-035) — a persona is not an individual but a lineage defined by a body of literature, and the rule is citations stay, attributions go. lineages.md contains no names.

The two reasons that argued for private, and what each one becomes

T-019:

1. The topic assessments that will accumulate in the KB are unpublished research directions. Nothing to do with names. This one is not resolved, it is relocated. Nothing of the kind is exposed today — J1 has not started and kb/ does not exist — so it stops being a question about who can read the repo and becomes a rule about what J1 is allowed to write into it. That rule does not exist yet, and it has to before the first topic assessment lands: in a public repo, deleting the file afterwards does not undo the disclosure.
2. The V1 literature of axis X1 overlaps the group's own output. Without naming the lineage, the reading list points at it anyway. So how much T-035's name-stripping actually strips has not been measured — and staying public means carrying that cost while it stays unmeasured. Measuring it no longer gates the repository; it decides whether V1's entry points need thinning.

Therefore:

• A content rule for J1 is now a Phase 2 interlock, alongside the other two ways a KB dies. It is a writing rule, not a disclosure question, so it belongs with the schema rather than here. → §4 Phase 2
• C-006 dissolves (§7). Its whole force was that two repos had different disclosure scopes; all three are public and this one stays public, so there is one scope. What survives in Q-008 is the timing argument, which never depended on visibility.
• Consent from the people involved is a separate matter, and this decision sharpens it. They are no longer persona targets, but the literature overlap remains and is now public by decision rather than by oversight. → §6 agenda 8.
• The prohibition in _common.md matters more, not less. "This person would find X uninteresting" was forbidden as attribution without a basis — reasoning that never rested on the repo being private. What has changed is only the size of the audience for a slip.

---

9. What this repo does not do

Not done here	Done where
Running or judging simulations	BD
Proposing instrument settings, running experiments	microscope
Originating numbers	Nowhere (philosophy ②)
Setting thresholds and tolerances	BD and the microscope, each for itself. Values are domain-owned
Settling a conclusion ("this topic has value")	A human. Irreversible judgments are not on the automatic path (BD I-057)

---

10. Documents

Document	Contents	Status
design/charter.md	Boundaries between the three repos. The evidence that two of them converged — §2 re-derived against the live remotes 2026-09-05	sketch
design/ideas.md	Idea log (T- · C- · Q-)	sketch
design/kb-schema.md	J1's formal objects. The answer to BD I-133. Seven entry kinds — challenge/ added 2026-09-05	sketch
design/personas/_common.md	The common contract for value personas	sketch
design/personas/lineages.md	Value lineages, 3 axes / 6 poles, plus the roadmap entry points. No names	sketch
design/personas/V1-frontier.md	X1 maturity, positive pole. The first output of T-013, ever. Trajectory check consistent	draft
design/personas/V2-unfinished.md	X1 maturity, negative pole. Trajectory check not_run — and that is not a pass	draft
src/	Does not exist	—

---

Reference convention

IDs from other repos carry a prefix: BD:I-050 · MS:G27. This repo's own IDs are T- (idea), C- (conflict), Q- (open question).

Quotations from BD's design documents are translated; the Korean originals are findable by ID in that repo.

Indexing targets (T-004 — index, do not migrate):

Prefix	Repo	Visibility
BD:	https://github.com/kyu-softmatter/Brownian-Dynamics-Agent	public
MS:	https://github.com/kyu-softmatter/agentic-microscope	public

Q-007 (what does this repo point at for BD) is closed — both repos are on GitHub, so references pin commit SHAs. Local paths are not used.

---

Licence

MIT, settled 2026-09-04 — the same as agentic-microscope and Brownian-Dynamics-Agent. A documentation licence would have been the more exact instrument for a repo that is currently all prose, and it was rejected on one ground: this is a software repo that has not got its code yet (src/ arrives in Phase 1), and with C-006 dissolved (§7) a monorepo is now possible on disclosure grounds. Choosing differently here would rebuild a smaller version of the seam that just came down.

Two things it does not reach. The literature these documents cite — the entry points in lineages.md are citations, and a citation carries no right over the work it points at. And the quotations from BD's design documents, which are translated here; those are the same author's and MIT there, so nothing is in conflict, but they are that repo's text and not this one's.

And the thing a licence cannot settle, stated here because §8 is the section people will not read: MIT grants reuse of these documents. It does not grant anything about the researchers whose literature the value lineages were back-derived from. That is a consent question, not a copyright one → §8.