Donor
An immutable copy of deployed client artifacts and captured behavior. This is the authority for what shipped and what was observed.
KodeBack process
Recover facts first. Ask for judgment second.
KodeBack keeps source recovery inside a repeatable evidence system. The exact stage set varies by application, but the authority model remains stable.
The three trust surfaces
Donor. Evidence. Candidate.
Evidence
Derived reports, runtime captures, dependency findings, source maps, protected surfaces, parity artifacts, history, and handoff material. Evidence can be regenerated and versioned; it is not candidate source.
Candidate
The editable application engineers will build, run, inspect, and continue maintaining. The candidate is never allowed to become its own oracle.
Intake
What application shape is this? Which entrypoints and routes exist? Are source maps present, live, partial, stale, or absent? Which assets, external origins, states, and adapters matter?
The output is a route decision and evidence inventory, not a generic “minified” label.
Seed
Create the first runnable candidate from the strongest available substrate: trustworthy recovered source, unpacked bundle modules, readable lifted source, or retained donor bytes.
The goal is runnable continuity, not cosmetic perfection.
Dependency and vendor recovery
Join static fingerprints, runtime identity, source-map provenance, version evidence, export shape, and byte ownership.
A region may be externalized, retained as donor authority, classified as modified/private vendor code, separated, or refused.
Runtime observation
Collect constructors, values, lifecycle, DOM and framework ownership, scene identity, request, storage, IPC, worker, message, and state evidence.
The system collects evidence. People or models may interpret genuinely ambiguous meaning.
Scoped cleanup and recovery
Apply covered normalization, source-map promotion, safe vendor externalization, binding-aware renames, asset restoration, selected promotion, or deliberate descoping.
A change is not accepted because it looks cleaner. It must satisfy the gates for its mutation class.
Evidence and parity
Run parse, lint, typecheck, test, build, import/export, module-load, browser, console, network, DOM, accessibility, visual, WebGL, Electron, extension, and retained-flow checks as configured.
No single layer is described as complete proof.
Checkpoint and handoff
Tie accepted changes to the source and evidence state that earned them.
A useful handoff includes fingerprints, commands, verified states, protected surfaces, residual seams, history, and one next action.
Where AI fits
Meaning can be assisted. Authority stays outside the model.
AI is valuable for
Choosing human-readable semantic names, selecting among evidence-backed options, interpreting ambiguous runtime roles, diagnosing mismatches, and proposing bounded implementations where no deterministic actuator exists.
AI should not own
Donor truth, authorization, protected-contract policy, source identity, mutation mechanics when deterministic, required validation, rollback, or final trust status.
Take the next step
See whether the evidence is there.
A recoverability assessment begins by determining what the donor can actually support and where the work should stop.