ZEILX.AI — Source Ledger Design
← Verification
Verification · Source Ledger
Design document · not yet built

A ledger to watch
the sources over time

Two linked tables. One row per source, one row per use. A scheduled workflow would re-check every source on a cycle and raise a hand when something changes — never deciding what the change means. This page documents the design; the workflow does not exist yet.

§ 01The one rule that shapes everything

Every column belongs to exactly one of two classes. This is not a formatting choice — it is the human gate expressed as a schema. If a workflow could write to a tier, the gate would leak.

Machine-written

Facts a workflow can establish without judgment: does the link resolve, is there a retraction notice, has the version changed, when was it last looked at. Overwritten on every run.

Human-only

Judgments a workflow must never touch: the tier, why that tier, whether the source still supports your claim, and whether it is cleared. Changed only by you.

§ 02Tab one — Sources

One row per source, forever. A source cited in six reports appears here once. This is what gets re-checked on a schedule, and what makes the checking compound instead of repeating.

SOURCES keyed on source_id · one row per source · never duplicated
source_id citation type doi / url link status retraction version last checked tier tier rationale still useful
KC-01 Institutional report, 2024 institutional 10.xxxx/kc.2024 200 OK none v1 · current 2026-07-14 T1 Primary; figures in source PDF yes
USAT-02 Outlet salary database journalism https://…/db 200 OK n/a updated 2026-06 2026-07-14 T2 Editorial standards; corroborated ×2 yes
MTA-07 Journal article, 2019 peer-reviewed 10.xxxx/j.2019 200 OK RETRACTED 2026-07-14 T1 Peer-reviewed at time of use needs review
BLG-03 Expert blog post commentary https://…/post 404 n/a 2026-07-14 T3 Named expert; not independent needs review
Machine columns refresh every run. Gold columns stay exactly as you left them — the workflow only sets "still useful" to needs review, never to yes or no.

§ 03Tab two — Usage

One row per use of a source. This is the tab that makes a retraction actionable: it knows every report and every claim a source is holding up.

USAGE joined to Sources on source_id · one row per citation instance
usage_id source_id report claim supported load-bearing impact flag status
U-014 KC-01 The Broken Bargain Wealth share figures, §3 yes clear cleared
U-031 MTA-07 The Manufactured Contest Behavioral effect claim, §2 yes source retracted review required
U-046 MTA-07 Before the Algorithm Supporting context, §5 no source retracted review required
One retraction in Sources → two flagged rows here, in two different reports, without you searching for them.

Why the split matters

A single source is checked once. The consequence lands everywhere it was used. Flatten this into one sheet and the same source gets duplicated per report — then updated in one place and not the others, which is exactly how a ledger rots.

MTA-07 · retraction detected ──▶ The Manufactured Contest · §2 · load-bearing Before the Algorithm · §5 · contextual

§ 04What the workflow does on a cycle

Scheduled — weekly or monthly. It reads, checks, writes to machine columns only, and stops at the gate.

01 · auto
Read

Pull every row from Sources.

02 · auto
Check

Resolve the URL. Query Crossref/OpenAlex for retraction and version metadata.

03 · auto
Write

Update link status, retraction, version, last checked. Nothing else.

04 · auto
Propagate

Any change → flag every Usage row for that source and mark it for review.

05 · human
Decide

You judge: re-tier, replace the source, soften the claim, or publish a correction.

§ 05What the workflow may never do

Hard limits

These are not preferences. Each one, if violated, converts the ledger from a record of verification into a source of false confidence.

  • Assign or change a tier
  • Mark a source "still useful" — only "needs review"
  • Clear a usage row for publication
  • Delete a row, ever — retired sources are marked, not removed
  • Fail silently: a run that errors must notify, or you will assume it ran

Illustrative rows; no live data. Once built, the ledger becomes the enforcement mechanism for the four-pass method — an appendix entry with no ledger row behind it would not count as verified, however confident it reads. Until then, the four passes are applied by hand and recorded per report.