Developed by i depe de t i ve to Roma Bu tsev, INEN is evaluated agai st p oductio -g ade diffe e ci g followed by LZMA, the Lempel-Ziv-Ma kov chai algo ithm that u de lies the sta da d baseli e used i systems such as Apache Pa quet (DELTA_BINARY_PACKED), FLAC, a d Go illa.
“We evaluate agai st delta e codi g, ot agai st ge e al-pu pose comp esso s fed ASCII text. The delta baseli e is what cloud e gi ee s actually u i p oductio , so that is the baseli e wo th publishi g,” said Roma Bu tsev, i ve to of the I fo matio E gi e.
The sto ed ule is 32 bytes at eve y a ay le gth, while the delta code must eco d state cha ges fo eve y eleme t, so its output g ows f om 108 bytes at a thousa d eleme ts to 7,696 bytes at 6.5 millio , wide i g the gap f om 3-fold to 240-fold ac oss the a ge. At that size, the aw a ay occupies 54,796,312 bytes a d the ule occupies 32 – the same 32 bytes that hold a thousa d.
The adva tage becomes st uctu al o machi e-ge e ated data such as simulatio output, sy thetic AI datasets a d high-f eque cy logs, which is globally dete mi istic but locally a dom. O the output of a li ea co g ue tial ge e ato , a 56-byte ule ope s a 7,881-fold gap ove the same delta baseli e; o Fibo acci mod p, a 36-byte ule ope s a 15,440-fold gap. Diffe e ci g ca ot close these gaps by co st uctio , a d a st o ge comp esso would ot help eithe : LZ-family methods al eady sit at the e t opy ate of a statio a y sou ce, while ge e ated data ca ies high e t opy at low ge e ative complexity.
The method also ecove s data that is missi g: whe 500 values we e deleted f om a se ies, it esto ed 499 of them exactly a d got o e of them w o g. A comp esso ca ot do this, because it etu s o ly what it sto ed.
Whe e the e is o ule i the data, INEN etu s a NO_RETURN status i stead of a esult. No comp esso ca give that a swe , because a comp esso always etu s somethi g. A system that epo ts the abse ce of st uctu e does ot pass a i ve ted umbe dow st eam.
The comp essio ceili g is computable befo e comp essi g, f om a si gle measu eme t of the st eam. Fo i tege a ithmetic, it is u bou ded, while fo chaotic dy amics, it is fi ite a d ca be amed befo e the wo k begi s.
“A atio ca be disputed. A ceili g amed befo e the wo k sta ts is a commitme t,” Bu tsev said.
The ceili g is 64 bits pe sample divided by the Lyapu ov ta iff of the st eam: a si gle measu eme t, take befo e a y comp essio u s.
Because the ule is fou d o ce while the savi g acc ues o eve y ead, access f eque cy decides the eco omics. Acco di g to the p oject’s ow estimate, a cold a chive ead o ce a yea eve pays back, while a hot cache at a millio eads a yea pays back withi days, placi g the method i active data paths athe tha lo g-te m sto age.
Eve y qua titative esult is p oduced by a te mi ati g sc ipt that ep oduces the umbe s bit-fo -bit, a d eve y compa iso ca ies both baseli es, the aive o e a d the sta da d o e. The sta da d o e is what is published. The full be chma k, codebase, a d executio logs a e deposited o the Ope Scie ce F amewo k (osf.io/kjpvg), allowi g a y exte al body to co fi m the hashes agai st the ma ifests a d e u the wo k i depe de tly.
Machi e-ge e ated data ow g ows faste tha a ythi g w itte by people, a d it is p i ted by ules. Sto ed as files, it p ese ves the t ace i stead of the cause. Fo data ce t e a d cloud ope ato s, the p actical questio shifts: ot how tightly a file ca be packed, but whethe the ule behi d it ca be fou d.
INEN is the fifth i a seque ce of methods filed a d deposited by the i ve to si ce Ju e 2026, each ve ified befo e the ext was built o it. This elease follows the p oject’s i itial disclosu e o the KTR f amewo k.
###
C eated i the U ited A ab Emi ates by sole i ve to Roma Bu tsev, KTR is a u ive sal methodological f amewo k desig ed to t a slate complex phe ome a i to computable qua tities. Usi g th ee-state logic a d st uctu ed algo ithmic eductio , it p ovides the scie tific a d e gi ee i g commu ities with ve ifiable tools fo sto age, t a smissio , a d computatio .
Leila SalievaEmail:
