State a Model Cannot Hold Belongs in Version-Controlled Files

# State a Model Cannot Hold Belongs in Version-Controlled Files ## Claim Any long-lived AI-assisted project should externalise its load-bearing facts into small, individually addressable files under version control, because the context window is not storage and a diff is the only honest record of what an agent changed.  ## Evidence **Source**: [[3. Resources/Readwise/Articles/Vibe Coding 2.0 I Wrote, Voiced, and Published a Full Novel in 60 Days|Vibe Coding 2.0]] — Zsolt Viczián The failure that motivates version control is not backup or collaboration. It is silent deletion: > [!quote] > it starts to forget stuff. It starts to remove stuff. Maybe it didn't even forget it. Maybe just deleted it A model discarding material it judged unimportant looks identical to a model that never had it. A diff distinguishes the two, and nothing else does. The externalisation half is the same insight applied to inputs rather than outputs: > [!quote] > unless I give a calendar and each of the scenes that you're reading in the book is pinned to a date AI gets confused Facts the model must reason *from* — dates, distances, who knows what, travel times — belong in files. Supplied, they get computed against. Withheld, they get invented, and the invention is consistent-sounding and wrong. ## Scope & conditions - Holds where work is long enough to exceed a single session and correctness depends on internal consistency. - Does not hold for short, disposable, single-turn tasks where the overhead exceeds the benefit. Premature structure also suppresses the divergence you want during exploratory drafting. - Larger context windows do not remove the need. Capacity to hold input is not capacity to sustain coherence across it, and the deletion problem is orthogonal to window size. - The categories worth externalising appear to be time, space, identity, and knowledge-state — though nobody has enumerated this properly, which is an open question rather than a settled list. ## Links - [[AI & the Future of Knowledge Work MOC]] — the context-engineering discipline that makes long-lived AI-assisted work survivable - [[3. Resources/Permanent Notes/An Agent That Reports on Its Own Work Verifies the Change Not the State]] — why the diff rather than the agent's account is the evidence - [[3. Resources/Permanent Notes/Intent Written on the Outside Compounds While Re-Derived Intent Is Debt]] — extends: the same argument about externalising what would otherwise be re-derived - [[3. Resources/Permanent Notes/Capture Seeds Before You Know Their Use]] — applies-in: the note-taking form of the same discipline - [[3. Resources/Permanent Notes/Notes Are Files Not Database Rows - Durability Beats Features]] — why files specifically, rather than any store