Breaking

Fuga Decides Whether Two Model Identifiers Are the Same Model Using Catalog Evidence Only

September 16, 2026 at 15:06 EDT

AI Agents
Foundation Models
Software Dev & Coding

Anyone who keeps a library of recorded agent runs eventually hits the same wall: the recording is still valid, but the model identifier inside it is not. Fuga's replay harness spent this window building an answer that does not involve guessing. A pair of functions added to the replay model environment module now decide, from catalog evidence alone, whether a model identifier observed on a live command line is the same model as the one captured in a recording.

2026-09-17 · FUGA · LOCAL DEVELOPMENT

When the recording outlives the model name

A replay harness now decides whether a live model identifier and a recorded one are the same model using catalog evidence only, and writes down every translation it had to make.

declared encodings the harness will reverse

conditions required before a retirement alias is accepted

unique resolution required, or the match fails

Evidence in, guesswork out

Both columns count admissible matching rules. The harness accepts three rules that the catalog itself declares, and none based on a shared prefix or family name.

catalog-declared rules accepted

prefix or family guesses accepted

Three ways a match can pass

EXACT

The two identifier strings are character-for-character identical

ENCODING

The requested string reverses, through a declared encoding, to the recorded identifier

RETIREMENT ALIAS

A retired entry names the current entry as its replacement, and both display names match exactly

CONDITIONS FOR THE ALIAS PATH

All four must hold, or the comparison returns unproven equivalence.

The recorded entry is marked retired and the current entry is neither retired nor deprecated
The retired entry explicitly names the current entry as its replacement
Both entries carry a non-empty, identical display name
The current entry encodes its command-line identifier as its own identifier

RECORDED, NOT SILENT

Any non-exact match is appended to the case result with the worker, the timestamp, the catalog digest and the full mapping source.

SCOPE OF THE CHANGE

This is replay bookkeeping only. The module states it never mutates the catalog, the model pin or user policy, and never changes model admission or selection.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.