Killing a prototype destroys a running system. It does not destroy what the project learned about what correct behaviour looks like, and that knowledge is expensive, slow to reproduce, and portable.
The evaluation set is where most of it lives. It encodes the cases that mattered, the judgements that resolved the edge cases, and the disagreements someone settled. A team that rebuilds a system can rebuild the code. It cannot rebuild the argument about what the system should do without having the argument again.
The candidates
Salvageable assets are separable outputs that might serve a justified future use. Evaluation cases and labels, schemas, adapters, deployment scripts, and the decision evidence are the usual ones — and candidate status settles neither ownership, usefulness, support, nor suitability for a different population.
Closeout guidance in other disciplines treats this as a defined phase: identify and archive technical information, preserve data, document lessons, and consider closeout options, costs, and risks during earlier phases rather than at the end.
That last clause is the one software programmes skip. Salvage planned at closure is salvage attempted by people who have already moved on.
Rights come apart from artifacts
Rights and restrictions are identified separately for code, data, labels, and third-party dependencies, and unresolved permissions go to the responsible owner before reuse or disclosure.
A public code licence does not automatically cover a bundled dataset. That single distinction accounts for a large share of the assets that cannot actually be reused, and it is discovered at the point of reuse rather than at the point of archiving.
Reuse guidance also asks for clarity about ownership and permitted use, review before opening existing code, and attention to maintenance and security response — which is a set of obligations attaching to the act of preserving something for others.
The package, and what it must retain
The specific version, with provenance, setup instructions, dependency constraints, tests, and known failures.
For evaluation assets specifically: the population, the label definitions, the exclusions, and the exposure history. That last field is what determines whether the set can serve as held-out evidence later, because a set used during tuning has a different role afterwards.
Secrets and live credentials are separated from reusable configuration, and the record distinguishes what was merely archived from what was inspected, successfully exercised, or adopted. Those four states are frequently reported as one.
Salvage is a forward comparison
The decision compares reuse against replacement, against a simpler manual method, and against doing without.
Take a parser that could be salvaged for one defined workflow: six person-days to extract, four to adapt, three to validate — 13 days. The estimated replacement is ten days including equivalent validation. On those assumptions salvage is the more expensive option, and the fact that the prototype consumed 80 days beforehand is not an input.
The 80 days is the trap. It makes salvage feel like recovering value and makes replacement feel like waste, and appraisal excludes already-incurred unchangeable costs from the decision about what to do next.
The same example points at the better answer. If the parser’s labelled edge cases have separate permission and independent utility, preserving those without adopting the parser captures the durable part and discards the rest. The evaluation asset outlives the implementation it was built to test.
What compounds across the portfolio
An organisation that salvages systematically pays the discovery cost once rather than once per project.
The mechanism is specific. Each prototype produces a definition of correct behaviour for its domain, a set of labelled hard cases, and knowledge about which integrations are painful. The second project in that domain starts with all three. The second and third transfers to the same receiving team start with a known evidence format and a team that has done it before.
Nothing here measures that effect. It follows from the assets being separable and reusable, which is a property of how they were packaged rather than of how good the project was.
The rule
What stays fixed is that closure assigns a disposition to every asset: candidate, restricted archive, validated for a stated use, accepted for maintenance, or rejected. What changes is which assets have a named future use, and an asset with none is archived as evidence or retired rather than kept hopefully.
Not to be confused with
A valuation. These are proposed record states, not a model that assigns a number to salvage value. Existing data obligations and closeout restrictions still need a case-specific review.
Rescuing the project. Preserving what a stopped project established is compatible with the project having been correctly stopped, and confusing the two is how salvage becomes an argument for continuation.