21. Milestones

Six milestones, from M0 at the end of 2026 to M5 at the end of 2028. Each one has a scope, a list of deliverables, exit criteria that aik can check, and a budget. A milestone is done when every exit criterion holds on the release of its month, not when its date passes. If a date passes first, the release ships with what holds, and the gap is listed in the release notes and in the coverage report (18.7).

The plan follows the lesson of algorithm (01.6): the pipeline of algorithm-work could write 3,617 pages in 7 days, and the gates failed first (the deploy, the audits, the stubs). So M0 builds the gates and the gold set before any bulk writing, and every later milestone grows the corpus only as fast as the gates and the reviewers can follow.

21.1 Overview

Milestone	Date	Theme	Nodes	Cards	Core at	Release
M0	2026-12	the toolchain, the gates and the gold set	about 30	about 10,000	(core not yet set)	`2026.12`
M1	2027-03	harvest everything; three pilot areas; the core list approved	about 1,500	about 100,000	C1 (all carded)	`2027.03`
M2	2027-07	the core written	about 8,000	about 220,000	C2	`2027.07`
M3	2027-12	the core grounded and evidenced; all 52 top areas started	about 16,000	about 330,000	C3; C4 for core methods, architectures and metrics	`2027.12`
M4	2028-06	breadth; the results register complete	about 25,000	about 420,000	C5 for half of the core	`2028.06`
M5	2028-12	complete	about 32,000	about 480,000	C5; C6 for the gold set and 10% of the core	`2028.12`

The node and card counts are Estimates (18.8), and M1 measures them on three pilot areas. The growth from M1 to M5 is about 1,500 nodes and 18,000 cards a month.

21.2 M0: toolchain, gates and gold set (2026-11-13 to 2026-12-31)

Scope. Everything that a writer needs before it writes the first bulk page.

Deliverables:

Repositories. tamnd/ai and tamnd/ai-agent with the trees of 20.1 and 20.2, the licences, CODEOWNERS, and the workflows check.yml, tags.yml, lean.yml, build.yml, deploy.yml.
aik, copied and adapted. The packages of 03.4 that come from algorithm-agent and programming (node, card, schema, tag, graph, taxonomy, markup, katex, lean, framework, runner, testkit, viz, status, audit, site, search, prov, release), working on the new kinds. The new packages in a first version: facts, claims, cite, exp for cpu-small, results, harvest, universe, coverage, queue.
work/, copied. All scripts of 14.1 with their tests on recorded replies, calling aik. writer.py runs the loop of 14.6 end to end on one node. ship.sh and deploy.sh ship one PR to the live site.
Schemas. node.v1, card.v1, facts.v1, claims.v1, experiment.v1, expect.v1, levels.v1, targets.v1.
Registers. topics/areas.yaml with the 9 groups, 52 top areas and about 615 sub-areas of 05.2; registers/canon.yaml with the 19 canon works of 5.10; registers/models.yaml (14.4); registers/frameworks.yaml with the tier-1 images (08.2); registers/notation.yaml (06.4); registers/scope.yaml (18.5); registers/watch.yaml.
Harvest of five sources into ai-raw and the inventory: Epoch AI, the Papers with Code dump, the AI Incident Database, arXiv metadata of the categories of crosswalk/arxiv.yaml, and Wikidata. aik universe merge and aik card sync produce the first cards.
Bridge cards. aik bridge sync al, m2 and pg for the items of 5.9.
The gold set (06.12), each page to W5 and to the top of each other ladder that applies to its kind, reviewed line by line by a person:

Kind	Gold page	Kind	Gold page
method	`dropout`	dataset and benchmark	`imagenet`, `imagenet-ilsvrc-2012`
architecture	`transformer`	task	`image-classification`
metric	`bleu`	policy	`eu-ai-act`
concept and theorem	`pac-learning`, `pac-bound-finite-hypothesis-class`	system	`alphago`
finding	`neural-scaling-laws`	software and hardware	`pytorch`, `nvidia-h100`
argument	`chinese-room`	work	`vaswani-2017-attention`
open problem	`does-scaling-lead-to-general-intelligence`	person, organization, event	`alan-turing`, `deepmind`, `dartmouth-workshop`
risk	`reward-hacking`	model	`llama` (family)

Plus gold cards, one for each card-only kind and for the card forms of the main kinds: a definition card (logit), a model card (Llama 3.1 405B), a result card (a Llama 3.1 405B score on MMLU with its setting), a benchmark card (MMLU), a dataset card (FineWeb), a work card, an incident card from the AI Incident Database, a provision card (EU AI Act Article 50), an event card, and an exercise card. And q-learning as the first page written by the pipeline (not by hand), as the test of the whole loop.

The site. The Worker ai-site with R2 ai-site, serving the gold set, the area pages, the collection pages of the M0 cards, search, the .md and .json twins, and /about/coverage/. The live check of deploy.sh passes.

Exit criteria.

aik check and aik build pass on main; aik site check passes; the live site shows the commit.
Every gold page has W5, G3, the ladders that apply at their top (X3 in the three tier-1 frameworks for dropout, transformer and bleu; R3 for dropout; F2 for pac-bound-finite-hypothesis-class; V2 for transformer; D3 for the facts of llama, imagenet-ilsvrc-2012 and nvidia-h100).
q-learning went through writer.py with the cross-family audit and reached W3 without a hand edit, and its run record is in prov/runs/.
Cost per node measured on the gold set and on q-learning; the M1 budget lines reset from it (14.8).
Release 2026.12.

21.3 M1: harvest everything; three pilot areas (2027-01 to 2027-03)

Scope. The universe complete for every source of 18.4; the pipeline proven on three top areas; the core list approved.

Deliverables:

All adapters of 18.4 in production, with weekly harvests; the universe register for all 52 top areas; every row with an outcome or counted as pending.
The daily job of 14.9 live; the one-day delay measured.
The catalog phase (14.3) on all 52 top areas, so that the core list (5.10) can be computed; registers/core.yaml approved by the editors.
The write phase (14.6) on three pilot top areas, chosen to test different kinds (Default): one area of methods and architectures (deep-learning-training), one of evidence and results (evaluation), and one of argument and policy (ai-safety-and-alignment or ai-governance-and-law).
The results register (10.6) loaded from the Papers with Code dump with the citation check (Q-35) and from the leaderboards of registers/leaderboards.yaml.
Licence checks of the sources marked "checked at M1" (Q-23).
The person review process (14.11): at least one reviewer for each of the 9 groups.

Exit criteria.

Universe: no source of 18.4 is unharvested; pending rows under 5% of rows with an outcome.
About 1,500 nodes, all in the pilot areas or the core, every published node at W2 or more; about 100,000 cards, all at D1.
The core list approved and every core item at C1.
Gate pass rate on the first try and audit pass rate measured by kind; the item caps and prompt versions reset.
The scale of 18.8 re-estimated from the pilot areas, and this document updated.
Release 2027.03.

21.4 M2: the core written (2027-04 to 2027-07)

Scope. Every core item has a page; the evidence machinery runs at scale.

Deliverables:

The write phase on every top area that holds core items, in reading order (5.7).
Implementations at X2 in NumPy, PyTorch and JAX for every core method, architecture and metric with code (08).
Experiments at cpu-small for every core finding and method with a small-scale form (09.2); the GPU pool (09.4) in use at the M2 cap of 09.10.
The Lean package with the first learning-theory, probability and RL modules (11).
The comparison tables and results tables of the core benchmarks computed (10.7); /about/queue/ public.
Courses (5.7) published for the first three paths.

Exit criteria.

Every core item at C2 or more (about 4,000).
About 8,000 nodes and about 220,000 cards.
No claim on a core page without a type and a source (G1); the citation check run on all core pages.
Release 2027.07.

21.5 M3: the core grounded and evidenced (2027-08 to 2027-12)

Scope. Depth on the core; breadth started in every top area.

Deliverables:

Every core item at C3: W3, G2, labels computed on every claim, D2 on every fact field.
Every core method, architecture and metric at C4: X2 in the tier-1 frameworks, R1 where a small-scale form exists, G3.
Tier-2 frameworks for the core architectures and metrics (Q-36).
gpu-medium experiments at the cap of 09.10; the GPU provider revisited (Q-32).
The write phase started in all 52 top areas.
About 300 Lean theorems (11).

Exit criteria.

Core at C3, and the core methods, architectures and metrics at C4.
About 16,000 nodes and 330,000 cards.
A core item below C3 after this date is a failed promise (18.7), and the build reports it.
Release 2027.12.

21.6 M4: breadth and the results register (2028-01 to 2028-06)

Scope. Every area deep; results complete.

Deliverables:

Every sub-area with at least the items that its canon and frontier sources name at C2.
The results register with every result of S-R1 (18.5), the computed tables of every benchmark node, and the contamination and saturation notes (10.8).
The "I need to…" search (17) and the comparison pages (16.5) on all core pairs.
Half of the core at C5 (the targets of schema/targets.v1.yaml).
About 450 Lean theorems.

Exit criteria.

About 25,000 nodes and 420,000 cards.
Half of the core at C5; no pending row older than 90 days.
Release 2028.06.

21.7 M5: complete (2028-07 to 2028-12)

Scope. The claim of 18.1 holds.

Deliverables:

Every row with outcome node at the target of its kind and level; every row with outcome card valid.
About 9,000 implementations, about 2,500 experiments and about 600 Lean theorems (02.4).
The gold set and at least 10% of the core reviewed by people at their current hashes (C6).
The datasets of 16.10 at each release, on the site and on the Hub.
The decision on languages other than English (Q-17).

Exit criteria.

About 32,000 nodes and 480,000 cards (18.8), with the counts of the last measurement.
The core at C5; the coverage report has no failed promise.
Release 2028.12.

21.8 Cost Estimates

All costs are Estimates; M0 measures the cost per node, and each milestone resets its lines from the measured cost of the previous one (14.8).

Line	M0	M1	M2	M3	M4	M5	Total
months	1.5	3	4	5	6	6	25.5
model APIs (writers, auditors, verifiers, coders, experimenters, catalog)	$500	$2,000	$8,000	$12,000	$12,000	$10,000	$44,500
GPU pool (09.10)	$150	$300	$1,000	$2,000	$2,400	$2,400	$8,250
hosting: Workers Paid and R2 `ai-site` (03.7)	$10	$20	$40	$60	$90	$120	$340
R2 `ai-raw`, `ai-data`, `ai-runs` (03.7)	$15	$60	$100	$150	$200	$240	$765
CI and the build server	the server exists						$0
Total	$675	$2,380	$9,140	$14,210	$14,690	$12,760	$53,855

How the model line is estimated: 32,000 nodes at about $1 to $1.50 each for the write, gate and audit loop (algorithm's metered pages cost less than $0.10 each, but AI pages have claims, facts and code, and we pay for the cross-family audit and the verifier on every node); about 9,000 implementations at about $0.50; about 2,500 experiments at about $1 for the experimenter; and the catalog phase. The line is high in M2 to M4, when most nodes are written, and lower in M5, when the work is mainly fixes and reviews.

The time of people (editors and reviewers) is not in this table. It is the scarce resource, and 14.11 plans for it.

21.9 Risks

Risk	Effect	What we do
Models invent facts, scores, dates and sources	false pages that look right	facts only from fact records (04.9, D family); the citation check (07.5); the cross-family audit and verifier (14.6); invented sources are blockers
The gates fail at scale, as algorithm's deploy did	a stalled pipeline or a broken site	Workers and R2 from day one (03.7); the live check of `deploy.sh`; M1 tests the pipeline on three areas before M2 writes the core
Model costs pass the lines	the milestone slips	metered runs, item caps and batch caps (14.8); budgets reset from measured cost
The other model family is rate-limited or unavailable	audits wait	items wait instead of breaking the cross-family rule (Q-08); a third family as a fallback
Too few reviewers	C6 stays low; contested pages unbalanced	the reviewer-assistant (14.2); review targets only for the gold set and a sample of the core (C6 at M5); review by subject in `registers/reviewers.yaml`
A source changes its licence or closes, as Papers with Code did	lost data or a takedown	the raw store keeps snapshots (13.6); licence classes per field (19.2); facts and names only from class D sources
Bias on contested topics	the site takes a side	positions as data (06.9), rule N01, the reviewer rules of 15.5 and Q-19
Hazardous content in security and safety pages	harm	`hazard: review` and person review before publishing (Q-24)
The field moves faster than the pipeline	stale pages	the daily job (14.9); facts with `as_of`; no "state of the art" in prose (K04)
The content repository grows too large for git	slow clones and CI	card sharding and Parquet in R2 for `row-only` rows (20.3); the size alarm
Framework releases break tests	X levels fall	pinned images; `frameworks.yml` opens a PR only when tests pass (20.5)
Privacy of living people	harm and takedowns	public professional facts only (P family, 19); person nodes need a reviewer (Q-14)

Source of this page