AI Research Tech Tree

The AI Research Tech Tree is a public, evidence-linked atlas of artificial-intelligence research developments, open directions, landmark works, and recorded relationships. It combines a chronological Timeline, a structural Network, a capability-oriented Opportunity View, and an accessible List.

Public/live release: v1.2.1 Stable. Release-prep PR #14 was squash-merged into durable main as 59d4d90ebcb8bb8b230d9cea2286214a3d2011a5; annotated tag v1.2.1 (object 49ddf1d74def95a4582cf2c3f9a391b4d5e99619) targets that commit. GitHub Pages run 33715374446, job 100523342659, passed and deployed on 2026-09-03. The live manifest reports version 1.2.1, tag v1.2.1, commit 59d4d90, Stable state, clean source, required-clean source, matching committed inputs, and data digest 865174514ba64e20d6f2a90471a6766b6d5fa18f5b0e62c85d9601de077a50f2. The citable dataset identity remains v1.0.0 in CITATION.cff, edition 2026-08-21-stable-1, because this UI release makes no semantic dataset changes. The full exact merged-tree hosted browser/Lighthouse suite was not rerun; that limitation does not invalidate the successful deployment or live byte reconciliation. Each staged build records its exact commit, tag state, and checksums in release-manifest.json; pull-request previews and untagged builds remain non-release artifacts. The atlas remains a curated research aid, not a complete literature review, a peer-reviewed historical account, or a claim that every displayed relationship is causal. Records expose their available sources and review state so that gaps can be inspected and improved in public.

The bounded v0.1.1, v0.2.0, and v0.2.2 source checkpoints are preserved in PLAN.md. The authorized v1.0.0 release freezes the citable dataset contract described below; the public v1.2.1 UI release preserves that contract while updating presentation and navigation.

• Public site: https://neb6dav.github.io/ai_tech_tree/
• Repository: https://github.com/neb6dav/ai_tech_tree
• Maintainer: @neb6dav
• Corrections and additions: open a structured issue or submit a pull request

What the views mean

• Timeline emphasizes chronology and research branches.
• Network reveals the structure of recorded associations, including hubs, bridges, leaves, and cross-branch links.
• Opportunity follows one historical development into documented capabilities, refinements, applications, constraints, competing approaches, and explicitly hypothesis-grade research opportunities. The first bounded map covers diffusion models.
• List provides a searchable, keyboard-friendly representation of the complete filtered result set.

Network positions are computed from recorded associations. Visual proximity, centrality, cluster membership, and node position do not by themselves establish chronology, importance, similarity, consensus, intellectual influence, or causality. Consult each record and relationship's evidence and review state.

The Opportunity View is a separate, time-oriented capability graph rather than a restyling of the historical atlas. Its paths use a uniform width: width does not encode volume, importance, value, certainty, or remaining opportunity. Solid, dashed, and colored paths communicate declared relationship and evidence categories. Opportunity status is local to the named branch and context; the project does not claim that a technology is globally “fully exploited.” Typed cross-links connect records back to their corresponding Timeline and Network entries.

The current diffusion alpha has an auditable boundary: 60 nodes, 94 relationships (93 report rows plus one explicitly contextual editorial connectivity edge), 78 source URLs, eight branches, eight constraints, eight hypothesis cards, and nine unresolved claims. It remains imported_unreviewed pending source-by-source human review.

Evidence vocabulary

The atlas deliberately separates a relationship from the strength of the evidence attached to it. Current records may use states such as:

• Direct: a source explicitly supports the stated relationship.
• Partial: a source supports part, but not all, of the stated relationship.
• Contextual or indirect: sources support the surrounding history but do not explicitly establish the edge.
• Editorial association: a curator-added connection useful for navigation or interpretation; it is not presented as a proven causal lineage.
• Unassessed: the record or edge has not yet received the indicated review.
• Hypothesis: a deliberately speculative research connection or open direction.
• No source found in the current audit: the review did not locate supporting evidence in its stated source set. This is not proof that the claim is false.

See METHODOLOGY.md for the inclusion, sourcing, relationship, and review rules.

Repository map

The public v1.2.1 UI release and its v1.0.0 dataset citation keep a single-file application shell while the historical atlas is authored in a strict 15-lane canonical dataset. The build projects that canonical data into the application and the machine-readable publication artifacts.

Maintained source

• ai-research-tech-tree.html — maintained application shell; its embedded historical-atlas projections are generated from the canonical atlas and are not a second authoring source
• src/data/atlas/manifest.json, catalog.json, sidecars, and the nodes/ and relationships/ lane shards — canonical historical-atlas authoring source
• canonical-atlas.js — strict built-in loader and assembler for the canonical atlas
• src/compatibility/ai-research-tech-tree.html — state-preserving public compatibility redirect to the root application
• src/network-view.js — 2-D WebGL network-view source
• src/opportunity-view.js and src/opportunity-layout.cjs — accessible Opportunity View renderer and deterministic layout logic
• src/data/opportunities/diffusion-models.alpha.json — maintained diffusion-models Opportunity map, published canonically at data/opportunities/diffusion-models.alpha.json
• src/data/opportunities/opportunity-map.schema.json and validate-opportunity-data.js — canonical schema and evidence-aware validation rules, published at data/opportunities/opportunity-map.schema.json
• src/compatibility/opportunity-map.schema.json — delegating schema retained at the former public src/data/... URL
• generate-knowledge-graph.js — canonical-data projector and machine-readable graph exporter
• build, layout, and injection scripts — deterministic generation of the network layout and deployable page
• release-gate.js, ui-layout-gate.js, accessibility-gate.js, network-gate.js, and opportunity-gate.js — validation gates
• publication, methodology, governance, and contribution files in the repository root and .github/

Generated files; do not edit directly

• index.html — GitHub Pages entry point
• network-atlas.bundle.js — locally bundled @cosmos.gl/graph network renderer
• opportunity-atlas.bundle.js — locally bundled Opportunity View renderer
• network-layout-v1.json — deterministic, versioned network coordinates
• ai-research-tech-tree.json
• ai-research-tech-tree.jsonld
• ai-research-tech-tree.ndjson

The JSON, JSON-LD, and NDJSON exports currently describe the historical atlas. Their dataset human URL resolves to ./, and each of the 339 atlas records resolves directly to ./#node=<stable-id>; the former published HTML filename remains only a compatibility entry point. The Opportunity map remains a separate, schema-linked JSON graph and is advertised from the page as an alternate machine-readable representation; this prevents capability assessments and hypotheses from being silently recast as historical lineage.

Use with LLM research systems

The machine-readable exports and the 339 static node pages delivered in the public v1.2.1 UI release are designed to be ingestible by retrieval systems and LLM-assisted research workflows. Stable IDs, typed relationships, chronology, lanes, evidence grades, review states, provenance, and durable source links can help a research agent retrieve relevant lineages, compare traditions, follow supporting or contradictory context, and identify documented evidence gaps. The static pages are a crawlable representation of the same canonical records, not a second semantic dataset; later combinatorial tools are intended to expose underexplored intersections without silently promoting them to facts.

This is a navigation and retrieval substrate, not an authoritative scientific knowledge base, a training corpus warranty, or evidence that an automatically suggested connection is true or novel. LLM workflows should preserve the atlas's evidence distinctions, treat contextual, editorial, unassessed, and hypothesis-grade relationships accordingly, and inspect the cited primary sources before making scientific claims.

ai-research-tech-tree.original.html is an archival pre-repair baseline retained for provenance. It is not the current application source.

Generated artifacts remain committed through the public v1.2.1 UI release so releases can be inspected, downloaded, and served without a build service. Change maintained source, run the build, inspect the resulting diff, and commit the regenerated artifacts with the source change.

v1 public contract

The citable dataset identity remains version 1.0.0, release state Stable, edition 2026-08-21-stable-1, dated 2026-08-21. The public application/UI release is v1.2.1; it changes presentation and research navigation without changing the canonical dataset. The historical-atlas review cutoff remains 2026-08-04; the edition date does not imply that every historical record was re-reviewed on 2026-08-21. The diffusion Opportunity map remains an alpha dataset dated 2026-08-19 with import state imported_unreviewed. Stable application code therefore does not imply publication-level validation of that separate map.

The following publication paths are stable:

• / for the application;
• /ai-research-tech-tree.html as the state-preserving compatibility alias;
• /ai-research-tech-tree.json, /ai-research-tech-tree.jsonld, and /ai-research-tech-tree.ndjson for the historical exports;
• /data/opportunities/diffusion-models.alpha.json and /data/opportunities/opportunity-map.schema.json for the Opportunity map;
• /src/data/opportunities/diffusion-models.alpha.json and /src/data/opportunities/opportunity-map.schema.json as compatibility paths;
• /CITATION.cff, /release-manifest.json, /social-card.png, /robots.txt, and /sitemap.xml for publication support.

The historical export keeps schema version 2, generator version 1.3.1, dataset IRI urn:uuid:7d0547f2-6239-5a56-82a3-1c846701c866, and the matching #vocab- vocabulary IRI. The Opportunity map and its canonical schema remain at schema version 1.0.0; the schema $id is https://neb6dav.github.io/ai_tech_tree/data/opportunities/opportunity-map.schema.json. The deterministic Network coordinates retain schema 1.0.0, layout network-v1, algorithm deterministic-lane-force-v1, and seed ai-research-tech-tree-network-v1.

Every existing exported record ID is frozen at the v1 candidate baseline, including historical lanes, classifications, nodes, relationships, evidence assessments, papers and links, landmark works and links, and Wikipedia sources, plus Opportunity records and its map and visual-band IDs. The ordered identity inventories are locked by SHA-256 in the publication-compatibility test:

• historical export: f5cff253d7a70641cf1f9a9058561f6d69bbae2d365166a7883694b3ef90241a;
• Opportunity records: 65afb35bc56b6d771312cc59a2ab3d4a7b48828b54a2262b7efb18d845702b33.

Correcting prose, evidence, dates, or relationships must not silently rename an existing record. Any exceptional identity correction or future inventory addition requires explicit compatibility review and a deliberate update to the locked baseline.

Canonical historical authoring uses schema 1.0.0 and the fixed lane order roots, symbolic, search, rl, neural, training, language, vision, generative, prob, alt, robotics, safety, systems, science. Nodes live in src/data/atlas/nodes/<lane>.json; relationships live in src/data/atlas/relationships/<target-lane>.json; manifest.json fixes those paths and the six approved sidecars. Embedded HTML records and machine-readable exports are projections, not alternate authoring sources.

Build and validate

The repository declares Node.js 24.x and npm 11.x as its artifact-producing toolchain families. Those declarations do not freeze a Node or npm patch release, and the ubuntu-24.04 Actions label does not make the evolving hosted runner image immutable. Calibration and validation reports therefore record the exact observed runtime and browser versions. Reproduce the checked-in artifacts with:

npm ci
npm run build
npm run test:fast
git diff --exit-code

npm run test:fast is the ordinary pull-request tier: it runs the data, accessibility, layout, Network, Opportunity, canonical-data, staging, contract, and deterministic artifact-budget gates without a browser or preview upload. Maintainers can install Chromium and run npm test for the manual release-candidate tier; that full suite assembles _site, exercises all four views in headless Chromium at desktop and mobile sizes, and runs Lighthouse. The browser gate blocks external requests, console errors and warnings, missing runtime fragments, broken deep links or focus restoration, and active-DOM drift from the reviewed platform peaks (7,244 on Windows and 7,244 on Linux) as well as any breach of the unchanged 8,000 ceiling.

At the v0.2.2 checkpoint, Lighthouse is a blocking regression signal against the staged application on a controlled, uncompressed, no-store local origin. The source calibration used five independent mobile-profile runs on Windows x64 with Node.js v24.14.1, Lighthouse 13.4.1, Playwright 1.62.1, and Playwright Chromium 151.0.7922.34 revision 1234. Each gate uses the independent median of three runs.

Metric	Five-run Windows median	Blocking limit	Local three-run gate
Performance score	53	at least 42	50
First Contentful Paint	22,728.84345 ms	at most 27,500 ms	22,734.979 ms
Largest Contentful Paint	22,900.34345 ms	at most 27,500 ms	22,893.525 ms
Total Blocking Time	166 ms	at most 550 ms	248 ms
Cumulative Layout Shift	0.00082719	at most 0.02	0.00082719

The v1.2.1 Stable release was fully verified locally on Windows at implementation commit 98d0c4a7bda01cc15303bd0a4939bc0829b98181: 28/28 browser tests, HTML 4,591,487 bytes, gzip 664,842 bytes, artifact SHA-256 a8365ca06bf53d74457c540e906b436e0b38e2a0d3386eff2fd2866713e90691, and Lighthouse medians of score 53, FCP 23,188.613 ms, LCP 23,361.613 ms, TBT 157 ms, and CLS 0. Release-prep PR #14 squash commit 59d4d90ebcb8bb8b230d9cea2286214a3d2011a5 is the tagged and deployed release commit. Publication remains 339 nodes, 711 relationships, and 339 node URLs with zero semantic changes; the semantic digest remains 865174514ba64e20d6f2a90471a6766b6d5fa18f5b0e62c85d9601de077a50f2. The full exact merged-tree hosted browser/Lighthouse RC rerun was not performed; live deployment and byte reconciliation passed.

The original Windows-only score and TBT limits proved too narrow on the configured Ubuntu runner. Two independent hosted attempts against the exact same application bytes both produced score medians of 47, with TBT medians of 362.5 and 362 ms; the six raw samples ranged from score 44 to 48 and TBT 325 to 440.5 ms, with no audit warnings. The cross-platform score floor is therefore five points below the hosted median, while the TBT ceiling rounds to about 25% above the hosted maximum. The paint ceilings and CLS ceiling are unchanged. The normal hosted gate then passed all five revised limits on ubuntu-24.04 in Actions run 32489666292, completing the v0.2.2 source checkpoint. These measurements do not represent live GitHub Pages delivery or real-user field performance.

The release manifest records the target package version, dataset edition, publication state, exact full commit, observed Node and npm versions, and every payload file's media type, byte count, and SHA-256. In clean-source mode, staging also proves that the configuration, metadata, individual artifacts, and complete directory inputs are regular committed blobs from the advertised HEAD; symlinks, gitlinks, Git LFS pointers, replacement objects, index concealment flags, dirty submodules, and generated or Git-administration input paths fail closed. The manifest cannot contain its own digest without a cryptographic self-reference, so it explicitly excludes itself. Local dirty-tree staging remains available for pre-commit review but is labeled non-clean and cannot be deployed.

The static contract uses a pinned browser-compatible HTML attribute decoder, rejects nested iframe[srcdoc] browsing contexts, validates live-document fragments without treating inert <template> contents as targets, and checks JSON Schema reference closure and application-state ID uniqueness. Runtime-created fragments are verified by the Chromium smoke gate. Nested JSON Schema $id scopes are intentionally prohibited until the validator implements their full base-URI semantics.

The stable Opportunity endpoints are ./data/opportunities/diffusion-models.alpha.json and ./data/opportunities/opportunity-map.schema.json. The former ./src/data/opportunities/... endpoints remain available for compatibility: the data is an exact second publication of the maintained JSON, while the old schema URL is a small schema with its own truthful $id that delegates to the stable canonical schema. The public ./ai-research-tech-tree.html alias likewise redirects to ./ and preserves query and hash state when JavaScript is available; its no-JavaScript fallback redirects to the root application.

GitHub Actions runs the fast, non-browser integrity tier for pull requests and does not upload a preview from that path. A maintainer can manually dispatch the release-candidate mode for full browser/Lighthouse validation and a downloadable preview artifact. There is no automatic post-merge main rerun. Under RD-006, RD-007, and RD-016, Pages is one manual protected-main job that checks out the exact annotated v1.2.1 tag, builds, stages, checks release identity, and uploads and deploys the artifact without rerunning the full suite. The v1.2.1 public deployment passed; the full exact merged-tree hosted browser/Lighthouse suite remains an explicitly unperformed additional check.

Contributing

The easiest contribution is a structured issue:

• correct a node, date, description, or source;
• propose a development or open research direction; or
• correct or add a relationship.

Opportunity-map contributions must also name the local status scope, evidence grade, and supporting source IDs. Candidate applications and open opportunities remain explicitly hypothesis-grade until their stated evidence threshold is met; adding a plausible idea is not the same as documenting a demonstrated capability.

For a pull request, identify the affected stable IDs, state the exact proposed claim, provide a source with a durable locator, distinguish evidence from interpretation, disclose material AI assistance, run the build and tests, and include the generated artifact diff. Full requirements are in CONTRIBUTING.md.

GitHub issues and pull requests are the project's track-changes system. The maintainer retains final editorial responsibility for what is included and how uncertain claims are represented.

Citation

Use the repository's Cite this repository control or CITATION.cff. The file remains synchronized to the tagged v1.0.0 dataset release and edition 2026-08-21-stable-1; the public v1.2.1 UI release does not change it. A future archive may add a DOI without changing stable atlas IDs.

Licensing

• Software and build tooling: MIT
• Original atlas prose, annotations, and graph data: CC BY-SA 4.0
• Linked or quoted third-party works are not relicensed: see THIRD_PARTY_NOTICES.md

Created and maintained by @neb6dav. Contributions are reviewed through the project's public editorial process.