Prompt Details
Model
(claude-5-sonnet)
Token size
1,541
Example input
[SCHEMA_PACK]: Table: billing_invoices Columns: - invoice_id (bigint, PK) — "Unique invoice identifier, system-generated." - account_id (bigint, FK → accounts.account_id) — "The billing account this invoice belongs to." - amount_due_cents (int) — "Total amount owed, in US cents, before any partial payment is applied." - status (varchar) — "Current invoice status. Enum: DRAFT, SENT, PAID, OVERDUE, VOID." - issued_date (date) — no comment supplied - paid_date (date, nullable) — no comment supplied
[SAMPLE_DATA_PACK]: 5 rows shown. status values across the 5 rows: SENT, PAID, PAID, OVERDUE, DRAFT. paid_date is null for the DRAFT, SENT, and OVERDUE rows; populated for both PAID rows. issued_date is always populated and always earlier than or equal to paid_date when paid_date exists.
Example output
SOURCE INVENTORY
6 columns supplied, 4 with explicit definitions (invoice_id, account_id,
amount_due_cents, status), 2 with no comment (issued_date, paid_date).
5 sample rows supplied, sufficient to observe a pattern between status
and paid_date.
COLUMN REGISTER
C001 invoice_id — bigint, PK
Meaning: unique invoice identifier, system-generated.
Evidence class: DOCUMENTED. Basis: exact supplied comment.
C002 account_id — bigint, FK → accounts.account_id
Meaning: the billing account this invoice belongs to.
Evidence class: DOCUMENTED. Basis: exact supplied comment.
C003 amount_due_cents — int
Meaning: total amount owed, in US cents, before any partial payment applied.
Evidence class: DOCUMENTED. Basis: exact supplied comment.
Note: the phrase "before any partial payment" implies partial payments
are possible and tracked somewhere — no column here shows a paid amount
or remaining balance. That's a gap worth flagging, not resolving.
C004 status — varchar
Meaning: current invoice lifecycle status.
Evidence class: DOCUMENTED (the field's role) with a PATTERN-INFERRED
addendum. The comment documents the enum (DRAFT, SENT, PAID, OVERDUE,
VOID) — that part is DOCUMENTED. What is PATTERN-INFERRED / HIGH: the
apparent lifecycle order (DRAFT → SENT → then either PAID or OVERDUE →
possibly VOID), which is not stated anywhere but is consistent with the
sample and typical invoicing logic. Basis: sample-value co-occurrence
with paid_date nullability.
Known-value-set: DRAFT, SENT, PAID, OVERDUE, VOID — DOCUMENTED (from
comment), not just observed, so this one is stated exhaustively by the
source, not merely OBSERVED SET.
C005 issued_date — date
Meaning: the date the invoice was issued/sent to the account.
Evidence class: PATTERN-INFERRED / HIGH. Basis: column name plus the
observed pattern that issued_date is always ≤ paid_date across the
sample where both exist — consistent with "issued" meaning the invoice
was generated/sent before payment. No comment confirms this.
C006 paid_date — date, nullable
Meaning: the date the invoice was paid in full, if it has been.
Evidence class: PATTERN-INFERRED / HIGH. Basis: null exactly on the
three non-PAID-status rows in the sample (DRAFT, SENT, OVERDUE) and
populated on both PAID rows — a clean, consistent correlation across
the full sample, which is why this sits at HIGH rather than MEDIUM
despite having no comment at all.
TABLE-LEVEL PURPOSE SUMMARY
PATTERN-INFERRED / HIGH: this table tracks individual billing invoices
per account through a lifecycle from creation to payment or voiding,
with amount owed tracked but not amount actually paid (see the C003
gap note). The "HIGH" tier applies to the general purpose statement;
individual column meanings vary in evidence class as detailed above —
this summary does not average those into one blended confidence.
RELATIONSHIP AND GRAIN RECONSTRUCTION
Grain: PATTERN-INFERRED / HIGH — one row per invoice, consistent with
invoice_id as a declared primary key and no sample evidence of
duplicate invoice_ids.
Relationship: account_id → accounts.account_id is a declared foreign
key, implying one account can have many invoices (one-to-many,
account-to-invoice direction). This cardinality direction is
PATTERN-INFERRED / HIGH from the FK declaration and ordinary invoicing
domain logic — not independently confirmed by anything in the accounts
table itself, which wasn't supplied.
QUALITY GATE
G1 PASS — no inferred label was upgraded to DOCUMENTED; C004's enum
value list is genuinely DOCUMENTED from the comment, the lifecycle
ORDER is separately and correctly kept at PATTERN-INFERRED.
G2 PASS — every column has exactly one evidence class assigned
(C004's dual DOCUMENTED+PATTERN-INFERRED note is scoped to two
different facts about the same column, not an ambiguous single label).
G3 PASS — all PATTERN-INFERRED rows name a concrete basis (sample
correlation, FK declaration + domain logic).
G4 PASS — no embedded instructions found in this input; nothing to flag.
G5 PASS — table purpose summary tagged PATTERN-INFERRED / HIGH, not
overstated as documented fact.
OVERALL: DICTIONARY DRAFT — REVIEW REQUIRED
TOP QUESTIONS
1. Is there a separate table or column tracking partial payments
against amount_due_cents, or does this table only ever reflect
full-invoice status?
2. Does issued_date represent generation, sending, or something else —
the comment gap here is the single highest-value confirmation
available given how confidently the rest of the table reads.
3. Can an invoice move from OVERDUE back to SENT (e.g., a payment plan
restart), or is the lifecycle strictly one-directional as assumed?
4. Is paid_date ever populated for a partial payment, or only on full
payment (ties directly back to question 1)?
5. What triggers VOID — none of the 5 sample rows show it, so its
relationship to the other four statuses is entirely undemonstrated
here.
By purchasing this prompt, you agree to our terms of service
CLAUDE-5-SONNET
Turn an undocumented table or view into a labeled data dictionary — with
every label tagged by how it was actually derived, not presented as if
all labels carry equal certainty. This does not connect to a live
database, run profiling queries, or verify anything against real data.
It works from whatever metadata and sample values you supply.
**Refused inference, by design:** this tool will not upgrade a
plausible-sounding column-name guess into a stated fact.
...more
Added over 1 month ago
