← all series

Appendix: additional discovery candidates

This is the research backlog for series and analyses that look promising but do not yet meet the repository's inclusion standard. A row here is a lead, not a validated time series: its unit of discovery, historical coverage, source stability and rebuild path still need to be audited before it becomes a problem folder.

There are two importantly different kinds of project here:

1. Direct discovery-event series, where each observation is genuinely a new entity, result, bound or validated performance record.
2. Quasi-experimental analyses of AI exposure, which use treatment timing or differential exposure to say more about causality than a trend line can.

A useful AlphaFold template

Hill and Stein's 2026 paper, “How Artificial Intelligence Shapes Science”, is close to the natural experiment this project should emulate. Experimental structure determination itself barely changed, but downstream research on proteins that lacked structural information before AlphaFold increased by about 15–40% relative to proteins that already had structures. Proteins without prior structures are more exposed to the AlphaFold shock, giving the comparison a clear treatment mechanism.

Candidate discovery-event datasets

Each candidate is scored on the five things that decide whether it can become a problem folder here, replacing the earlier single letter grade. Every column is ✅ (2 points), 🟡 (1) or ❌ (0), and the rows are ordered by the total out of 10; ties keep the earlier ordering. The criteria are:

• Unit — is "one discovery" a crisp, stable event? ✅ crisp · 🟡 needs a dedup or matching rule · ❌ ill-defined.
• History — can events be dated consistently, far back? ✅ clean deep dates · 🟡 recoverable with per-entry work · ❌ hard to recover.
• Rebuild — a public, stable, scriptable source? ✅ bulk or API, clean · 🟡 public but messy or licence-limited · ❌ not readily rebuildable.
• Frontier — a genuine frontier or first-of-kind, not instrument throughput? ✅ yes · 🟡 mixed, needs isolating · ❌ throughput-dominated.
• AI signal — can observations be tied to AI, by attribution or exposure? ✅ strong · 🟡 plausible but indirect · ❌ low exposure. A ❌ here is not disqualifying: several rows are wanted precisely as low-exposure controls, and the note says so.

Deferred algorithm benchmark reconstructions

These three sources have unusually good historical artifacts, but they do not yet provide clean time series of algorithmic progress. Their official annual results or release histories are not comparable measurements: the tasks, machines, rules or software interfaces change over time. They therefore belong in this appendix until controlled retrospective reruns have actually been completed.

MaxSAT Evaluation

The MaxSAT Evaluation archive spans annual solver generations from 2006 onward and often preserves solver source, benchmarks and detailed results. The official scores cannot simply be joined across years because benchmark composition, tracks, time limits, hardware and ranking rules changed.

A defensible series would require substantial extra work:

• freeze one weighted and unweighted benchmark corpus;
• specify an ex-ante rule for selecting each year's solver snapshot;
• recover and containerize roughly twenty generations of historical solvers, including obsolete compilers and dependencies;
• run every snapshot under identical CPU, memory and timeout limits; and
• vendor per-instance logs before calculating solved count, PAR-2 or objective gap.

Until that work is done, edition count measures competition activity and annual rankings mix algorithm progress with changing experimental conditions.

MiniZinc Challenge

The MiniZinc Challenge has annual editions from 2008 onward, with model, solver and result archives. Its official medal and score histories are not a clean time series because the model set, data files, MiniZinc and FlatZinc versions, solver interfaces, hardware, time limits and scoring system all changed.

Reconstruction would require selecting a recurring set of optimization models before examining outcomes, defining how each model is compiled for old solver interfaces, rebuilding historical solvers in pinned environments, and running them on common hardware. Compatibility translations would need to be audited so they do not accidentally give newer or older solvers different problems. This is a valuable project, but it is closer to experimental software archaeology than to parsing an existing ledger.

OR-Tools CP-SAT

OR-Tools release notes and tagged releases provide a dense chronology from CP-SAT's 2018 public launch. Release cadence and prose claims about “performance improvements” are not performance observations, however.

This is the easiest of the three reconstructions, but it still needs a frozen constraint-optimization suite, adapters for old APIs and model formats, pinned compiler and dependency environments, fixed hardware and worker counts, and repeated seeds for nondeterministic parallel search. Only those reruns could support a time series of runtime, solved count, proof rate or objective gap. Without them, a version-by-date chart would measure packaging activity rather than algorithmic progress.

ICARM collaborative record boards

The NSF Institute for Computer-Aided Reasoning in Mathematics (ICARM, Carnegie Mellon, grant DMS 2425401) runs a family of single-problem leaderboards, listed on its resources page. Each board fixes one open question, accepts submissions from logged-in contributors, machine-verifies every submission before recording it, and publishes the whole table at /database.json with a documented API, a /recent activity feed and Apache-2.0 source. Submissions carry a created_at timestamp and a named submitter, so an event series needs no reconstruction — it is the table.

The blocker is shared by all of them: the oldest board opened on 2026-05-27, so no board carries a pre-2026 baseline against which a 2026 rate could be compared, and the opening weeks are dominated by seeding rather than discovery (the elliptic-curve board took 131 of its 275 curves on 2026-05-27 and 2026-06-24/25). They are worth revisiting once each has a year of post-seeding activity; the note per board is what would have to be resolved first.

Elliptic-curve rank records

Built as problems/math-elliptic-rank/: the folder carries Dujella's lower-bound and exact-rank frontiers, the ICARM leaderboard snapshot, and the 2026 rank ≥ 30 step, whose AI attribution is held as a dated quote from the board's editable commentary field rather than as a settled credit.

Cross-domain analyses

1. Build a discovery index rather than an output index

Convert perhaps 30–50 series into dated discovery events. Fit the pre-AI trend separately for each series, estimate whether 2021–2026 observations lie above that trend, and meta-analyse the deviations. This avoids the central ambiguity of arXiv and Crossref counts: cheaper writing can increase output without increasing discovery.

2. Test whether acceleration is proportional to AI exposure

Assign each domain an ex-ante AI-exposure measure. Protein structure, medicinal chemistry and combinatorial optimization are computational and search-heavy; telescope-limited astronomy, field taxonomy and some experimental physics are less exposed. Estimate a post-AI × AI-exposure interaction and include the low-exposure domains as controls. A common 2023 kink across both groups would point to a secular data, investment or reporting effect rather than AI.

3. Use staggered AI shocks

“ChatGPT happened in 2022” is too crude. Candidate treatment dates include modern deep learning around 2012, AlphaFold2 and AlphaFold DB in 2021, generative protein design around 2022, broadly used LLMs in 2022–2023, GNoME in 2023, and increasingly capable scientific agents in 2025–2026. The Hill–Stein design shows how to combine a discrete shock with cross-sectional exposure.

4. Extend the AlphaFold design with BindingDB

Split protein targets by whether they had an experimental PDB structure before July 2021. For each target, measure time to first potent ligand and the frequency and magnitude of affinity-record improvements. Test whether previously structureless targets catch up after AlphaFold. BindingDB supplies publication dates, while AlphaFold DB supplies prediction coverage. This gets closer to “did AlphaFold increase drug-discovery output?” than a paper count does.

5. Apply the same logic to enzyme discovery

For each newly characterized EC reaction, recover the underlying publication date from BRENDA. Classify papers or labs as explicitly AI-assisted, or classify enzyme families by how much sequence and structure information AI can exploit. Then test whether the creation rate of genuinely new experimentally demonstrated catalytic functions changes.

6. Separate new data from better inference on old data

For each exoplanet, distinguish the observation date from the announcement date, then plot discoveries made from archival observations separately. AI that improves inference or search should create findings from fixed existing data without requiring a better telescope. The same design can be used for old code, sequencing datasets, collider data and other archived evidence.

7. Measure frontier magnitude, not only event count

Use Starrydata to reconstruct quantities such as global experimental ZT(t) or best battery capacity at a standardized cycle count. NREL already maintains this kind of frontier for photovoltaic efficiency. AI may create many trivial events without moving an important frontier—or one large frontier jump without many events—so both count and magnitude should be retained.

8. Measure problem-opening-to-solution latency

Candidate queues include newly identified drug target → first potent ligand, disease gene → validated mechanism, protein sequence → functional annotation, exoplanet candidate → confirmation, and open mathematical question → resolution. AI may appear more clearly as shorter queues than as increased gross output.

9. Run scientist-level event studies with NIH and OpenAlex

NIH ExPORTER provides bulk project data, legacy records, publications and patents. Combine it with OpenAlex authors, works, topics and citations. Identify researchers' first explicit AI use, match them to non-adopters using pre-adoption histories, and compare externally validated discoveries, patents, topic distance and citations after adoption.

A 2026 Nature study of 41.3 million papers, “Artificial intelligence tools expand scientists' impact but contract science's focus”, already reports higher individual publication and citation output alongside a narrower collective range of topics. The valuable extension is to replace papers and citations with externally validated discovery outcomes.

10. Treat low-AI-exposure series as falsification tests

Gravitational-wave detections, detector-limited astronomy, some taxonomic discoveries and experimental structure deposition dominated by microscope throughput can serve as explicit controls. If these accelerate in the same way and at the same time as BindingDB or computational optimization, the common cause is unlikely to be AI exposure alone.

Highest-value next investigations

The current priority order is:

1. BindingDB target–ligand frontiers
2. BRENDA new enzyme functions
3. NREL and Starrydata materials frontiers
4. PDG precision histories

Together they cover medicinal biology, basic biology, materials science and fundamental physics with unusually objective notions of progress. They also provide a useful mix of AI-exposed and lower-exposure domains for the cross-domain event-study framework.