Every claim needs an address
A useful research record connects each formula or result to a source file, version, section or line range, and recovery date. A GitHub commit pins code. An archive checksum pins bytes. A dataset revision pins upstream evidence. A page title or remembered description helps locate material but does not replace those identifiers.
Original formulas are preserved before notation is cleaned for display. The transcription, any normalization of symbols, and any interpretation are separate fields. If source text is incomplete, the page names the missing definition. It does not invent an equation to satisfy a remembered count.
Use more than one evidence axis
A source can be recovered without having been executed. A numerical result can be internally consistent while its measurement interpretation remains unsettled. A program can pass a local unit test without establishing that a historical multi-device deployment is healthy now.
Each page therefore records implementation evidence, runtime evidence, and scientific interpretation separately. Dates matter: the dashboard rendered in the October 14 recovery, but its offline indicator and changing counters did not verify a running Pantheon backend. Historical logs remain dated records rather than live system status.
| Evidence label | What it establishes |
|---|---|
| Recovered source | The stated code or document was obtained and inspected |
| Verified artifact | Specific bytes, structure or saved arithmetic passed a stated check |
| Recorded experiment | The supplied evidence records a run with the stated protocol |
| Observed runtime | A particular process or behavior was directly checked at a stated time |
| Documented design | A source describes intended architecture or a method |
| Project hypothesis | A scientific interpretation is proposed for testing |
| Remembered history | A conversation record supplies a recovery lead |
| Unresolved source gap | The needed source, definition, observation or relationship is missing |
A complete equation explanation
Start with the source formula and a plain-language sentence about what it computes. Define every input and output in that equation's local context. Give the domain, shape, units, indexing convention and any denominator restrictions. Explain how the operation behaves on simple cases, and identify exactly where the software uses it.
Then distinguish the source's intended meaning from properties that follow mathematically. A clipped score is bounded; that does not make it a calibrated probability. A mean cosine is a numerical statistic; that does not establish general synchronization. A state update can be deterministic; that does not establish equivalence to a biological circuit.
Finally, show what evidence exists: a derivation, a source implementation, a test, a saved experiment, or only a proposal. Where a formula repeats, preserve all occurrences and explain whether they are unchanged copies, renamed symbols, coefficient variants, or genuinely different rules.
Counts need a unit
An equation total may count unique mathematical forms, numbered statements, array components, code assignments, or repeated appearances across files. Those totals answer different questions. The archive should expose both occurrences and deduplicated forms, with generated members linked to their generator.
The same care applies to AI builds: a repository, a branch, an agent family, a deployment, a proposed generation and a manuscript are different entities. A coverage report can list all recovered entities and unresolved leads without presenting a remembered number as a verified inventory.
What replication would add
Checksum verification detects byte changes. Recomputing saved table arithmetic checks consistency. Rebuilding a program checks compatibility under a specified environment. Running the experiment again checks whether its outcome is reproducible under a protocol. Independent replication adds new execution and judgment rather than relabeling the same original output.
For the current MIP review, all saved primary outcomes were crosschecked and selected objective calculations were recomputed for every network. For the JAIC supplement, saved data integrity and selected summary arithmetic were checked. Neither operation reacquired the complete upstream adapter release or tested a current distributed runtime. The evidence pages retain those limits next to the verified results.