Query a column you never stored
A lot of enterprise columns are exact functions of other columns, total = qty×price, VAT, day-buckets, keys, running indexes, denormalized joins. AT-1 GenDerive discovers the function, verifies it on every row, and stores the column as a ~20-byte formula instead of data. Then it answers predicates on that column by running the formula, an index or aggregate query reads zero bytes.
Discover, store as a formula, query scan-free
at1 genderive build orders.csv -o orders.at1gd # 200,000 rows, 4 derived col(s) stored as formulas, 3 stored # total = qty*price (queryable, ~0 B) # vat = (total*15)//100 # seq = 1*i+0 at1 genderive query orders.at1gd "SUM total" # 6,967,314,136 [scan-free: computed from base, 0 B stored/scanned] at1 genderive query orders.at1gd "COUNT seq BETWEEN 50000 150000" # 100,001 [ZERO bytes read (index formula)] at1 genderive verify orders.at1gd # OK: reconstructed byte-exact (every formula-regenerated column matches)
Discovered, not declared
Finds product / affine / bucket / percentage / index dependencies, and composedones (a column derived from a derived column). Each is verified exact on every row before it’s trusted.
Scan-free answers
Aggregates compute from the base columns; index- function columns answer from pure arithmetic. The derived column is never stored and never scanned. Honest limit: a genuinely independent column can’t be beaten, it’s stored.
Byte-exact
Every eliminated column regenerates byte-for-byte on demand, verified against the original table’s hash. Storage shrinks; nothing is lost.
The general form of GenQuery: it works on any column that is an exact function of other stored columns, the union of generator-recovery and functional-dependency discovery. Pay-as-you-go: first 100 builds/month free, then metered per build. Query and verify are always free.
Most tables carry columns that are not data
A line total is quantity times price. A margin is revenue minus cost. A duration is the gap between two timestamps. Those columns are stored, indexed, backed up and replicated like everything else, and every one of them can be recomputed exactly from the columns beside it. You are paying to store an arithmetic operation.
This finds those relationships rather than being told about them, keeps the rule instead of the column, and recomputes the values on read. On a real table with three such relationships it removed 8.27% of the file, and the recomputed values are identical to the originals rather than close to them.
Where it finds nothing
A table of independent measurements has no derived columns and this will tell you so rather than finding a weak correlation and calling it a rule. It only keeps a relationship that holds for every row without exception, because a rule that is right 99.9% of the time is not a rule, it is a bug that has not been noticed yet.