Precedent dossier
Modeled on a real studio research dossier (four houses, three parallel lanes, 69-source ledger) and that project's facts sheet. Rule zero: cite or don't claim.
1. Folder layout
<project>/research/YYYY-MM-DD-<topic>/
README.md start here: identity table (spoken target -> resolved project ->
architect -> confidence + source ids), read order, known gaps
RESEARCH-BRIEF.md the ask as given, targets verbatim, lanes + id blocks, permissions
SOURCES.csv the ledger. One row per URL. Primary agent owns it.
agents/NN-slug.md one note per precedent (or tight pair), from templates/precedent-note.md
DRAWINGS.md D## index: type, local file, source id + page, scale, limits
ASSET-INVENTORY.json every saved file: path, kind, source_id, url, bytes, sha256
SYNTHESIS.md transferable moves -> proposed tests for our project (inference, labelled)
VERIFICATION.json checks run, results, limits, file hashes
NEXT-AGENT.md ten-line primer for whoever resumes
assets/ evidence/ saved PDFs/images (assets) and page captures/records (evidence),
prefixed by lane: 00- primary, 01-, 02-, ...
Optional: ACCESS.md when records took real work to get (portal, API, request form, what
failed). The original dossier called this PUBLIC-RECORDS-AND-DRAWING-ACCESS.md.
2. Source classes and weight
Use these source_class values (the first six come from the original ledger):
| source_class | Authoritative for | Not evidence for |
|---|---|---|
architect project source | credit, project name, published drawings, architect's stated intent | as-built state, exact areas (often rounded) |
official civic / planning record | legal address/parcel, permitted program, architect/engineer of record, permit dates, written dims on scanned plans | design intent; current condition |
architect firm source | firm-general philosophy | this project's intent (applying it = inference) |
project builder source | credit, completion year, program as built | architect's intent |
publication / corroboration | architect quotes inside interviews; second source for credits | editorial attributions on their own (the original dossier: a broker editorial misattributed one house) |
listing / property corroboration | address, photos, marketing areas | authorship, intent, scale (listing plans lacked title blocks) |
developer / marketing | client/developer identity, unit mix as sold | architect intent, verified areas |
book / academic | analysis, historical context, published drawings with scale | current state |
social / user-posted | a lead to follow | anything, until a stronger source confirms it |
When sources disagree, the higher class usually wins for its own column only: a permit beats a listing on legal program; the architect beats the permit on intent.
3. Citation rules
- Ledger first. Add the row to
SOURCES.csvwhen you open the source, not afterwards. Ids areS###, assigned once, never renumbered. Do not generate ids by sorting URLs; that renumbers everything when a source is added. - Every factual claim carries an id on the same line:
[S012],[S012 p.73],[S012 fig.4; S020 p.3]. Pinpoints (page, sheet, figure, timestamp) go in the citation, never in the ledger label, because one URL is one row however many pages you cite. - A claim with no source is marked
[unsourced]or deleted. Never in a crit deck. - Attributions (architect, year, client, architect of record) need a
architect project sourceorofficial civic / planning record, or two independent sources from other classes. Republished press repeating one release counts once. - Dimensions come only from a drawing or record that writes the number or states a
scale:
[S021 p.3, written]or[S021 p.3, scaled 1/8"=1'-0"]. Never from a photo, render, or image proportions. Published areas are "published figures", not measurements. Our own derived numbers say[derived: method]and cite their inputs (facts-sheet pattern). - Intent layers stay separate: architect-stated / team- or client-stated / firm-general / observed in photos / our inference. Label which layer each move belongs to.
- Status words from the project's evidence vocabulary:
verified-file,verified-live,reported,inference,rejected. Search snippets and blocked pages arereported, neververified. qualificationis specific to that source ("permit names architect of record, not designer"; "plans cropped, no title block"). Never boilerplate like "see report".access_date_localis the day that source was actually opened.
4. Parallel research (what the original dossier did)
It ran three fresh-context agents, one lane each, while the primary integrated:
- Each lane owns
agents/NN-slug.mdand assets/evidence prefixedNN-. Lanes never edit README, SYNTHESIS, SOURCES.csv, or project handoff files. - Give each lane an id block so parallel ids never collide: primary
S001-S099, lane 01S100-S199, lane 02S200-S299. Lanes put their rows in a## Sourcestable at the bottom of their note with the ledger's columns. - The primary merges those rows into
SOURCES.csv. If two lanes ledgered the same URL, keep the lower id and replace the other in the notes. Then it checks the key evidence itself (opens the permit, looks at the plan) and writes README, SYNTHESIS, DRAWINGS. - One precedent? Skip the lanes and write
agents/01-slug.mdyourself.
5. Copyright and assets
- Store links, short attributed quotes (a sentence at most) and your own paraphrase. Never paste long passages or whole articles.
- Save images/PDFs locally only as private study references: record path, source id,
URL, access date, sha256 in
ASSET-INVENTORY.json. Claim no publication licence. - No login or paywall bypass, no buying documents, no contacting architects or owners unless the user says so. A blocked page is recorded as blocked.
- Never substitute AI-generated or redrawn drawings for source drawings. Redraws are ours, labelled as such, and cite the sheet they trace.
6. Contradictions
Every note has a ## Contradictions section. Record both sides with ids, which one you
prefer and why (class, date, specificity), or "unresolved". Rejected claims go in
## Do not reintroduce so the next agent doesn't revive them (the original dossier: a wrong
firm credit and a predecessor address). Area/program figures that drift across permit, approval and
listing stay listed side by side; never average them into one "verified" number.
7. Validate
python3 scripts/check_ledger.py <dossier> # from this skill's folder
Exit 1 = hard errors (bad/missing columns, duplicate or malformed ids, bad URLs, empty required fields, notes citing ids that aren't in the ledger). Warnings: ledger rows never cited, inline URLs not in the ledger, lines with a number + unit and no source on the same line, boilerplate qualifications, duplicate URLs. Markdown links to ledger URLs count as citations, so older URL-cited dossiers can still be checked.
Done means
- README identity table: every architect/year/client cell has ids and a confidence word
-
check_ledger.pyexits 0; each remaining warning is fixed or explained in VERIFICATION - No dimension without a drawing/record id and a stated scale or "written"
- Each note: intent layers separated,
## Contradictions,## Do not reintroduce, search audit - Every saved file is in ASSET-INVENTORY.json with source id and sha256
- DRAWINGS.md says what was not found (no sections? say so) and the best next access route
- SYNTHESIS proposes tests, labelled inference; nothing claimed as adopted design
- NEXT-AGENT.md written; open questions listed, not smoothed over