The research, as an API.

Search everything

MCP tool: search

Full-text search across papers, Research Desk posts, glossary terms, Academy lessons and reference videos, research and free tools, topic hubs, and the site's hub pages (programs, science, milestones, agents, developers, about, …). Returns ranked hits with canonical URLs.

curl "https://neurogenesys.ai/api/v1/search?q=verifier+exploitability"

List papers

MCP tool: list_papers

All NeuroGenesys research papers, newest first: title, area, status, abstract and key findings. Includes pre-registered negative results.

Get a paper

MCP tool: get_paper

One paper's metadata and full text as Markdown (LaTeX math in $…$ / $$…$$), plus related posts, papers, lessons and tools.

curl "https://neurogenesys.ai/api/v1/papers/verifier-exploitability"

List Research Desk posts

MCP tool: list_posts

Research Desk posts (sourced analysis of continual learning, AI agents, neurotech and BCI, photonics, AI silicon and neural data policy, plus Field Watch, daily summaries of new research from other labs), newest first. Field Watch posts (category "Field Watch") report other labs' results, not NeuroGenesys findings; filter with category.

Get a post

MCP tool: get_post

One Research Desk post: full body as Markdown with inline source links, and the source list.

List glossary terms

MCP tool: list_terms

Every glossary term with its one-to-two sentence definition.

curl "https://neurogenesys.ai/api/v1/glossary"

Define a term

MCP tool: define_term

A glossary entry in full: definition, sections as Markdown, sources and related pages.

List Academy courses

MCP tool: list_courses

AI Academy tracks and courses with their lesson lists.

curl "https://neurogenesys.ai/api/v1/academy"

Get a lesson

MCP tool: get_lesson

One Academy lesson as Markdown (LaTeX math in $…$ / $$…$$), with its position in the course.

List topic hubs

MCP tool: list_topics

Topic hubs and the posts, papers, lessons, tools and glossary terms each gathers.

curl "https://neurogenesys.ai/api/v1/topics"

Get a topic hub

MCP tool: get_topic

One topic hub with its member pages.

List research tools

MCP tool: list_tools

The research tools, the papers each comes from, and which API operations compute them.

curl "https://neurogenesys.ai/api/v1/tools"

Search the findings ledger

MCP tool: list_findings

Results from the lab's own experiments, one entry per finding, with evidence level (L0 engineering, L1 toy, L2 small LLM, L3 literature), post-hoc flag, status (result, pending, correction) and date. Includes pre-registered negatives and corrections. Filter by text, study, level or status; with no filters returns the studies and the newest entries. Quote findings with their level.

curl "https://neurogenesys.ai/api/v1/findings?q=selection&level=L1&limit=10"

Get a study from the findings ledger

MCP tool: get_study

One study from the findings ledger with every entry, its preamble, date, status and the posts and papers that report it. Study slugs come from list_findings.

curl "https://neurogenesys.ai/api/v1/findings/civilization-superintelligence-model"

Search the research library

MCP tool: search_library

External arXiv papers by other researchers that NeuroGenesys papers, posts and studies cite (1514 papers), with where each is cited. Filter by text, field, year or citing study; sort newest, oldest or most-cited.

curl "https://neurogenesys.ai/api/v1/library?q=continual+learning&limit=5"

List Academy reference videos

MCP tool: list_videos

Curated external lecture videos and full courses behind the Academy and the research, each with why it is worth watching and where on the site it is used. Filter by Academy course.

List community papers

MCP tool: list_community_papers

Research submitted by the community and approved for publication at /research/community, newest first.

curl "https://neurogenesys.ai/api/v1/community-papers"

List site pages

MCP tool: list_pages

The site's hub and institutional pages (programs, science, milestones, agent selection reliability, compare, developers, about, access, trust, legal, …) with summaries, plus every other public URL from the sitemap. Use read_page for a page's text.

Read a page

MCP tool: read_page

The readable text of any public page by path, for pages without a dedicated operation: /programs, /science, /milestones, /agents, /compare, /about, /access, /trust, /people/…, /tools/…, legal pages. Prefer get_paper, get_post, define_term, get_lesson and get_topic when they fit.

curl "https://neurogenesys.ai/api/v1/page?path=%2Fmilestones"

Selection-error report

MCP tool: reliability_report · Try it in the browser: selection-reliability

From selection/v1 decision events: selection error ε overall and by pool size K with Wilson 95% intervals, the logit-linear ε(ln K) trend, shortlist misses vs confusions, judge-assisted ε (PPI++) when events carry judge verdicts, and, with target_accuracy, the largest library that keeps routed accuracy on target.

curl -X POST "https://neurogenesys.ai/api/v1/reliability/report" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"events":[{"chosen":"refund","target":"refund","pool_size":8,"routed_ok":true,"forced_ok":true},{"chosen":"billing","target":"refund","pool_size":64,"routed_ok":false,"forced_ok":true},{"chosen":"search","target":"search","pool_size":64,"routed_ok":true}]}'

Forgetting or mis-selection?

MCP tool: reliability_diagnose · Try it in the browser: selection-reliability

Compares a baseline and a current window. If forced accuracy (correct candidate forced) fell, the candidates degraded: forgetting. If ε rose while forced accuracy held, the selector is at fault: mis-selection. Forced accuracy is judged only on candidates present at baseline. Returns the verdict, the changes with Newcombe intervals, and an exact split of the routed-accuracy change into selection and capability parts.

Gate a skill, memory or model update

MCP tool: reliability_gate · Try it in the browser: selection-reliability

ALLOW, BLOCK or NEED MORE DATA for an update, from labelled decisions before and after. Non-inferiority on ε: ALLOW if the interval for the rise is below delta, BLOCK if ε rose. Default intervals are anytime-valid (empirical-Bernstein confidence sequences), so CI can re-run the gate as data arrives. paired=true compares the same probe queries before and after (match by event_id or order) and needs far fewer labels.

ε from a few gold labels and many judge labels

MCP tool: reliability_estimate · Try it in the browser: selection-reliability

Prediction-powered inference (PPI++): combine n gold-labelled decisions (y = 1 if mis-selected, f = the judge's verdict 0–1) with N judge-only verdicts into one ε estimate with a valid 95% interval. Reports how many gold labels the judge is worth. A poor judge gets weight λ ≈ 0, so the interval stays valid.

Split a routed-accuracy change

MCP tool: reliability_attribute · Try it in the browser: selection-reliability

Exact decomposition of a routed-accuracy change, A = (1 − ε)·A_forced + ε·A_wrong, into the selector's share (−Δε·(Ā_forced − Ā_wrong)) and the candidates' share ((1 − ε̄)·ΔA_forced + ε̄·ΔA_wrong), from summary rates of two windows.

curl -X POST "https://neurogenesys.ai/api/v1/reliability/attribute" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"before":{"eps":0.03,"a_forced":0.92,"a_wrong":0.1},"after":{"eps":0.09,"a_forced":0.9,"a_wrong":0.1}}'

Labels needed to gate an update

MCP tool: reliability_plan · Try it in the browser: selection-reliability

Labelled decisions needed to detect a rise in ε from eps to eps + change (two-sided α, power), for two independent windows or a paired replay of the same probe queries. Fixed-sample sizes; anytime-valid monitoring typically needs 1.3–2× more.

curl -X POST "https://neurogenesys.ai/api/v1/reliability/plan" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"eps":0.05,"change":0.02,"paired":true,"flip_rate":0.04}'

Upload decision events to a project

MCP tool: reliability_ingest · Try it in the browser: selection-reliability

Append up to 5,000 selection/v1 events to a hosted project (created on first upload). Events with an event_id already stored are skipped, so retries are safe. Costs 1 credit per started 1,000 events.

List your projects

MCP tool: reliability_projects · Try it in the browser: selection-reliability

Your hosted decision-log projects with event counts, tags and the time of the latest event.

curl "https://neurogenesys.ai/api/v1/reliability/projects"

Report on a project

MCP tool: reliability_project_report · Try it in the browser: selection-reliability

reliability_report over a hosted project's events, optionally one tag and/or events since a time.

Diagnose a project between two tags

MCP tool: reliability_project_diagnose · Try it in the browser: selection-reliability

reliability_diagnose with the baseline and current windows taken from a hosted project by tag (e.g. lib-v6 → lib-v7).

Gate an update in a project

MCP tool: reliability_project_gate · Try it in the browser: selection-reliability

reliability_gate with before/after windows taken from a hosted project by tag. Anytime-valid by default, so it can run on every CI build.

Delete a project

MCP tool: reliability_delete_project · Try it in the browser: selection-reliability

Permanently deletes a hosted project and all its events.

Best-of-N budget model

MCP tool: verifier_budget · Try it in the browser: verifier-budget

From a verifier's measured exploitability ε, the generator's per-sample accuracy p and the verifier's separation d′: Best-of-N accuracy at each N, the model's safe budget N*, and whether argmax or consensus selection wins. Paper: thurstone-bounds-best-of-n. A model, not a measured prediction: in the lab's pre-registered real-judge study (findings 8.13) no judge made accuracy fall by N = 32 and measured ε did not predict the best budget.

curl -X POST "https://neurogenesys.ai/api/v1/compute/verifier-budget" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"p":0.3,"dprime":1.5,"eps":0.2}'

Estimate ε from a validation set

MCP tool: estimate_epsilon · Try it in the browser: verifier-budget

Operational exploitability ε from verifier scores on a labelled validation set: the fraction of wrong candidates scored at or above the median correct score. Also returns the base rate p.

LLM-judge panel value

MCP tool: verifier_panel · Try it in the browser: verifier-panel

What an N-critic judge panel is really worth from per-critic error e and pairwise co-error P_both: shared-failure share s, error correlation ρ, effective panel size, the majority-vote accuracy ceiling, and critics needed for a target accuracy.

curl -X POST "https://neurogenesys.ai/api/v1/compute/verifier-panel" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"e":0.2,"p_both":0.08,"n":5,"target_accuracy":0.9}'

Panel stats from a miss matrix

MCP tool: panel_stats · Try it in the browser: verifier-panel

e, P_both, s and ρ from an items × critics miss matrix (1 = critic wrong, 0 = right, null = abstained), with seeded bootstrap standard errors.

Tail check before a Φ-based ε

MCP tool: tail_check · Try it in the browser: verifier-panel

Excess kurtosis of verifier score gaps and whether the tails are Gaussian enough to trust a Thurstone (Φ-based) ε.

Audit a verifier from a Best-of-N log

MCP tool: exploitability_audit · Try it in the browser: exploitability-audit

Measured exploitability of a verifier or reward model from scored candidates: problem ε (some wrong answer outscores every right one), candidate ε against the consensus threshold ε* = p/(1−p), the exact Best-of-N curve with its peak N* and the drop beyond it, ε by pool, and seen vs. unseen problems. One credit per 10,000 candidates. Tool: exploitability-audit.

Tune a selector on held-out decisions

MCP tool: selector_tune · Try it in the browser: selector-tune

Repair a router, retriever or verifier selector without retraining what it selects from: per-candidate score standardisation, a deferral margin that keeps ε among accepted picks at or below target_eps, and the smallest top-k shortlist. Decisions alternate between a calibration and a held-out half; every reported number is held-out. Returns the fitted selector. One credit per 5,000 decisions. Tool: selector-tune.

Standard error of a statistic

MCP tool: standard_error · Try it in the browser: threshold-audit

SE of an AUROC (under H0 or at a value), a proportion, a difference of proportions, or a sample mean, for sizing a pre-registered threshold.

curl -X POST "https://neurogenesys.ai/api/v1/compute/standard-error" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"statistic":"auroc_null","n_pos":40,"n_neg":60}'

Check a threshold before registering it

MCP tool: threshold_design · Try it in the browser: threshold-audit

Pre-registration checks: is the threshold at the null, within 2 null SEs of it, beyond the instrument's ceiling, or inside the flat-truth noise band of a range statistic over k points? Returns flags (malformed / unreachable / marginal).

curl -X POST "https://neurogenesys.ai/api/v1/compute/threshold-design" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"threshold":0.55,"direction":"ge","null_value":0.5,"null_se":0.06,"ceiling":0.9}'

Classify a missed threshold

MCP tool: threshold_verdict · Try it in the browser: threshold-audit

After the run: did the result pass, miss within 2 SE (UNDERPOWERED), or miss by more (SUBSTANTIVE)?

curl -X POST "https://neurogenesys.ai/api/v1/compute/threshold-verdict" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"observed":0.53,"threshold":0.55,"se":0.03,"direction":"ge"}'

Optical recirculation pass budget

MCP tool: photonics_recirculation · Try it in the browser: photonics-budget

Passes through a loop with per-pass transmission η before precision falls from start to floor ENOB, for thermal-noise-limited, shot-noise-limited and amplified detection.

curl -X POST "https://neurogenesys.ai/api/v1/compute/photonics/recirculation" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"start_enob":8,"floor_enob":4,"eta":0.9}'

Nonlinear material figure of merit

MCP tool: photonics_material_fom · Try it in the browser: photonics-budget

FoM = n2 / (β_TPA · λ) for a Kerr material; below 1, two-photon absorption eats the signal before a π phase shift. Pass a known material name or n2 and beta.

curl -X POST "https://neurogenesys.ai/api/v1/compute/photonics/material-fom" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"material":"Silicon nitride"}'

Logarithmic-number-system energy check

MCP tool: photonics_lns_energy · Try it in the browser: photonics-budget

Energy per MAC in linear vs. logarithmic arithmetic (Horowitz 45 nm figures), and the share of weight-bound decode energy that free arithmetic could save at a given bytes-per-weight.

curl -X POST "https://neurogenesys.ai/api/v1/compute/photonics/lns-energy" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"precision":"fp16","lut_pj":5,"bytes_per_weight":2}'

Pre-ADC aggregation collapse

MCP tool: photonics_aggregation_collapse · Try it in the browser: photonics-budget

Numerically shows a bank of K fixed-gated d×d operators equals one precomputed matrix: returns the residual ‖bank·x − W_eff·x‖ against the output scale.

curl -X POST "https://neurogenesys.ai/api/v1/compute/photonics/aggregation-collapse" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"k":8,"d":64}'

List free tools

MCP tool: list_free_tools

The free in-browser calculators at /tools, with the API operation that computes each one where there is one.

curl "https://neurogenesys.ai/api/v1/free-tools"

LLM API cost

MCP tool: llm_api_cost · Try it in the browser: LLM API Cost Calculator

Cost per request and in total for Anthropic, OpenAI and Google models from list prices (checked Nov 09, 2026), including cached-input and long-context rates. Omit model to compare every model.

List model prices

MCP tool: list_model_prices

The list prices behind llm_api_cost, in USD per million tokens, with each provider's pricing source.

curl "https://neurogenesys.ai/api/v1/tools/model-prices"

Training compute and cost

MCP tool: training_compute · Try it in the browser: LLM Training Compute Calculator

Training FLOPs (6·N·D), GPU-hours, wall-clock days and cost on an H100 or B200 at a given model FLOPs utilisation, plus the Chinchilla-optimal split of the same budget. Tokens default to 20× parameters.

curl -X POST "https://neurogenesys.ai/api/v1/tools/training-compute" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"params":7000000000,"gpu":"h100","gpu_count":256}'

Eval sample size

MCP tool: eval_sample_size · Try it in the browser: AI Eval Sample Size Calculator

How many eval items you need: a confidence interval of a given half-width (margin), a difference between two models on separate (independent) or the same (paired, McNemar) items, or the smallest difference n items can detect (mde).

curl -X POST "https://neurogenesys.ai/api/v1/tools/eval-sample-size" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"mode":"margin","p":0.8,"margin":0.02}'

Neural data rate

MCP tool: neural_data_rate · Try it in the browser: Neural Data Rate Calculator (EEG & BCI)

Raw data rate and storage for a neural recording setup (channels × sample rate × bit depth ÷ compression), compared with common wireless and wired links.

curl -X POST "https://neurogenesys.ai/api/v1/tools/neural-data-rate" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"channels":1024,"hz":20000,"bits":10}'

US neural data privacy laws

MCP tool: neural_data_laws · Try it in the browser: Neural Data Privacy Law Checker

State laws that regulate neural data: definition, consent rule, enforcement, effective date and source. Filter by state.

BCI company comparison

MCP tool: bci_companies · Try it in the browser: BCI Company Comparison

Brain-computer interface companies: device, placement, electrode count, reported performance, regulatory status and trial participants, with what each claim rests on.

GPU memory to run an LLM

MCP tool: llm_memory · Try it in the browser: VRAM Calculator: Can I Run This LLM?

Memory to run a model at a given quantization, context length and batch: weights, KV cache (GQA, sliding-window and MLA aware) and overhead, and how many of each GPU it takes, or the smallest Apple Silicon configuration that fits. Model and GPU data checked Nov 09, 2026. Same model as /tools/vram-calculator.

curl -X POST "https://neurogenesys.ai/api/v1/tools/llm-memory" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"model":"llama-3.1-70b","quant":"q4km","context":8192}'

LLM tokens per second

MCP tool: llm_inference_speed · Try it in the browser: LLM Inference Speed Estimator (Tokens per Second)

Roofline estimate of decode speed (memory-bandwidth or compute bound), aggregate throughput at a batch size, and time to first token for a prompt, on a listed GPU or Mac or a custom device. Efficiency assumptions are inputs. Same model as /tools/inference-speed-estimator; data checked Nov 09, 2026.

curl -X POST "https://neurogenesys.ai/api/v1/tools/llm-inference-speed" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"model":"llama-3.1-8b","quant":"q4km","device":"rtx4090"}'

Wrong-tool rate as an agent's tool library grows

MCP tool: agent_tool_selection_error · Try it in the browser: Agent Tool Count Calculator

Expected selection error ε (wrong tool picked, or the right one missing from the shortlist) at a library size K, from logit-linear fits to published benchmarks: all tools in the prompt, embedding or BM25 shortlists. Says whether K is inside the measured range, and the largest library under a target error. These are other teams' systems; measure your own with the reliability_* operations.

curl -X POST "https://neurogenesys.ai/api/v1/tools/agent-tool-count" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"k":300}'

LLM judge vs human agreement

MCP tool: llm_judge_agreement · Try it in the browser: LLM Judge Agreement Calculator

Agreement between an LLM judge and human labels: percent agreement, Cohen's kappa with a 95% interval, the confusion matrix, and for two labels sensitivity, specificity and how many human labels the judge is worth (PPI++). Pass label pairs or 2×2 counts.

curl -X POST "https://neurogenesys.ai/api/v1/tools/llm-judge-agreement" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"counts":{"tp":40,"fn":10,"fp":5,"tn":145},"unlabelled":5000,"unlabelled_positive_rate":0.22}'

BCI information transfer rate

MCP tool: bci_itr · Try it in the browser: BCI Information Transfer Rate Calculator

Wolpaw information transfer rate for a brain-computer interface: bits per selection, per minute and per second, typing-speed equivalents (log2 26 ≈ 4.7 bits per letter, or ~1 bit with a language model), and with trials, the accuracy needed to beat chance. Includes published reference systems.

curl -X POST "https://neurogenesys.ai/api/v1/tools/bci-itr" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"targets":36,"accuracy":0.9,"trial_seconds":2,"iti_seconds":0.5,"trials":40}'

EU AI Act risk tier

MCP tool: eu_ai_act_check · Try it in the browser: EU AI Act Risk Checker

EU AI Act (Regulation (EU) 2024/1689, as amended by the 2026 Digital Omnibus) risk tier for an AI system from yes/no answers: prohibited, high-risk, transparency or minimal, plus general-purpose model obligations, each with its article and application date (checked Nov 09, 2026). Call with no answers to get the questions. Not legal advice.

curl -X POST "https://neurogenesys.ai/api/v1/tools/eu-ai-act-check" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"answers":{"s_eu":true,"t_chat":true,"role":"provider"}}'

Count tokens

MCP tool: token_count · Try it in the browser: Token Counter & Context Window Checker

Token count of a text with the o200k_base encoding (GPT-4o and later; within a few percent for Claude and Gemini), and what that many input tokens cost on each model at list prices (checked Nov 09, 2026). Same as the /tools/token-counter page.

curl -X POST "https://neurogenesys.ai/api/v1/tools/token-count" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"text":"How many tokens is this sentence?"}'

Test-time compute cost

MCP tool: test_time_compute · Try it in the browser: Test-Time Compute Cost Estimator

Accuracy and cost of sampling N answers per query: coverage (any sample right), majority-vote accuracy given how often wrong answers agree, and cost per query and per month with the prompt cached after the first sample, plus the curve over N = 1…128. Same model as /tools/test-time-compute-cost.

curl -X POST "https://neurogenesys.ai/api/v1/tools/test-time-compute" \
  -H "Authorization: Bearer ng_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{"p":0.6,"n":8}'