algorithm-succession-lab

Reproducible experiments for a narrative-nonfiction paper on the handful of algorithms that ran the world — and what is replacing them. Every figure in the paper is generated by code in this repo, from public data or explicit synthetic models, with fixed seeds. No figure is borrowed; nothing is hand-drawn.

Companion repo to the paper "The Last Legible Machines: Six Algorithms That Ran the World, and the One That Ate Them" (in preparation; DOI to follow). The repo keeps its neutral name for stability.

Quick start

python -m pip install -r requirements.txt
python run_all.py          # runs the 8 self-contained experiments, writes figures/

Each experiment also runs on its own, e.g. python experiments/e1_flybook/run.py. Outputs: a PNG in figures/ and a metrics.json next to each run.py.

The experiments

#	Name	What it shows	Data	Live keys?
E1	Fly-book duel	Two Amazon repricing bots (Eisen's exact multipliers) spiral a book to $23.7M; phase diagram of when coupled optimizers explode	none (deterministic)	no
E2	Supercolony automaton	Founder-bottleneck → one borderless "us"; cue diversity → permanent war borders (the ant→platform-monoculture bridge, runnable)	synthetic	no
E3	Succession study (flagship)	How much of the classical top-10 web survives into LLM answer-engine citations, 2026	live search + answer APIs	yes (demo mode offline)
E4	Red Queen in a bottle	Spam evasion vs filter retraining as a measured arms-race curve; the cost of winning is false positives	UCI SMS Spam Collection (bundled)	no (classical); LLM variant optional
E5	Anternet replication	Harvester-ant foraging regulator vs TCP AIMD: one feedback law, two substrates	synthetic	no
E6	Flash-crash in a bottle	Volume-keyed selling into thinning depth craters a toy market; a stop-logic pause arrests it	synthetic	no
E7	Auction bench	Second-price learns truth-telling, first-price learns to shade to (N−1)/N; revenue equivalence — why the 2019 switch wasn't a revenue grab	synthetic	no
E8	Gale–Shapley bench	Deferred acceptance yields provably zero blocking pairs; greedy accumulates thousands	synthetic	no

Headline results (this seed)

• E1: growth factor 1.2684/cycle → ~51 cycles from $106 to the observed $23,698,655.93.
• E2: at cue-diversity D=2, a mean of 1.0 surviving supercolony and 0 border; at D=64, ~24 colonies and ~1,600 cells of active border.
• E3 (LIVE, 2026-08-22): ChatGPT answered 80% of 50 queries without consulting the web at all (Gemini: 62%). When the web was consulted, only 37% (ChatGPT) / 29% (Gemini) of cited domains came from the classical top-10; the survivors sat at mean classical rank 2.6–3.3. Consultation by query type: informational 5%, ambiguous 5%, YMYL 25%, factual 30%, local 40%, commercial 75%.
• E4: static filter decays from 100% to ~87% catch over 8 rounds; a retraining filter holds ~100% — but its false-positive rate sits at ~17% (the arms race isn't free).
• E6: crash floor 87.5 without a pause vs 94.0 with a 6-step stop-logic pause.
• E7: learned shading → 1.00 (second-price) vs 0.79 (first-price; theory 0.80); revenue 0.667 vs 0.656 (theory 0.667).
• E8: greedy matching at n=800 → ~27,000 blocking pairs; Gale–Shapley → exactly 0 at every size.

Reproducibility conventions

• Every experiment fixes a seed (mostly 20260822) and is deterministic.
• One command per figure; run_all.py regenerates everything in ~15s.
• Each run.py writes a metrics.json — the numbers the paper cites.
• Toy models are labelled as illustrations, not causal claims, in both the docstring and the figure/metrics (see E6's limitation field).

Keyed experiments (E3 live, E4 LLM variant)

These need API access and stay reproducible only up to provider drift, so they run in a clearly-watermarked demo/offline mode by default.

• E3: the classical baseline ships as a dated snapshot (baseline_2026-08-22.json), so no search key is needed to reproduce the comparison — only an answer-engine key. Copy experiments/e3_succession/providers.example.py to providers.py (gitignored), set GEMINI_API_KEY / OPENAI_API_KEY / PERPLEXITY_API_KEY, then python experiments/e3_succession/run.py --live. The runner also accepts an optional answer_citations_detailed(engine, q) provider returning {"cited": [urls], "retrieved": bool, "model": str, "searches": [...]}, which lets it distinguish "never consulted the web" from "searched but cited nothing" — both are measurements, not failures. The 2026-08-22 live run (results.json) used ChatGPT + Gemini. Demo mode (no keys) produces a synthetic figure watermarked "do not cite".
• E4 LLM variant (run_llm.py): optional, gated behind an ethics flag.

Data & ethics

• E4 uses the public UCI SMS Spam Collection (bundled in data/, 5,574 messages). It reports aggregate metrics only, uses mechanical perturbations already documented in the spam literature, synthesizes no novel spam, targets no live filter, and emits no reusable evasion payloads. The LLM variant is held to the same rule and is off by default.
• Toy financial/ecological models (E2, E5, E6, E7) are qualitative illustrations of published mechanisms, not calibrated predictors.

License

MIT — see LICENSE.

Citation

If you use this code, cite the repository (a paper DOI will be added on publication). Data: Almeida, Gómez Hidalgo & Yamakami, UCI SMS Spam Collection.