You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
docs(roadmap): tighten v1 acceptance criteria per pillar
Restructure each pillar's Acceptance block into shippable artifacts and
demonstrable signals. Add bare-metal (BCM) coverage and Vera Rubin to the
accelerator axis. Reframe the contribution bar around per-persona
self-service thresholds for user, integrator, and contributor.
Copy file name to clipboardExpand all lines: ROADMAP.md
+58-23Lines changed: 58 additions & 23 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -24,7 +24,12 @@ four axes plus the validation that backs it.
24
24
**Service axis.** Validate end-to-end AICR on managed Kubernetes services — both Cloud Service Provider (CSP) and partner-cloud offerings — alongside vanilla Kubernetes.
25
25
26
26
**Accelerator axis.** H100 is mature; v1 needs L40 and B200 in addition to the
27
-
GB200 work already in flight, with GB300 framed as a near-horizon target.
27
+
GB200 work already in flight, with GB300 and Vera Rubin (VR) framed as
28
+
near-horizon targets.
29
+
30
+
**Service axis (bare metal).** Bare-metal service classes — including BCM and
31
+
equivalent operator-managed Kubernetes distributions — are represented in the
32
+
v1 matrix alongside CSP and vanilla Kubernetes.
28
33
29
34
**Intent axis.** Both training and inference per (service × accelerator) cell once
30
35
that cell is in scope.
@@ -37,14 +42,24 @@ hand-authored criteria.
37
42
advisory on PRs (Tier 1 generic plus diff-aware Tier 2) with the full matrix
38
43
running on merges to `main` and nightly. v1 closes the gap by adding a nightly
39
44
UAT cadence on live managed Kubernetes services and tightening these signals
40
-
into the merge path so coverage stays validated between releases.
45
+
into the merge path so coverage stays validated between releases. Results are
46
+
published through a Testgrid-style dashboard so the matrix has a single,
47
+
durable evidence channel anyone can read.
41
48
42
49
**CNCF AI Conformance.** Certify AICR-on-EKS as conformant once the WG finalizes
43
50
requirements. Treat conformance evidence as a first-class output of the validator.
44
51
45
-
**Acceptance:** every (service × accelerator × intent) cell in the v1 matrix
46
-
validates against a snapshot, generates a bundle, deploys via UAT, and is exercised
47
-
by nightly automation.
52
+
**Acceptance:**
53
+
54
+
*Shippable artifacts:*
55
+
- v1 matrix of (service × accelerator × intent) cells declared in this document, with bare-metal classes (BCM) and forward-looking accelerators (GB200, GB300, VR) represented as in-flight slots.
56
+
- CNCF AI Conformance evidence emitted by the validator and regenerated per release.
57
+
- Nightly UAT cadence on live managed Kubernetes wired into the merge signal.
58
+
- Testgrid (or equivalent dashboard) brought up as the canonical evidence channel for recipe validation.
59
+
60
+
*Demonstrable signals:*
61
+
- Sustained green-rate across the matrix in Testgrid across multiple release cycles.
62
+
- Recipe portfolio carries enough gravity that the dominant service / accelerator / platform needs of most users land inside the supported matrix.
48
63
49
64
### 2. Stability
50
65
@@ -66,9 +81,16 @@ content so bundles are portable.
66
81
**Versioning policy.** Document the major-bump policy for each surface in
67
82
`RELEASING.md` and wire compatibility tests into CI.
68
83
69
-
**Acceptance:**`v1.0.0` ships with a frozen, documented contract for the CLI,
70
-
REST API, and bundle layout; any subsequent breaking change is detected in CI and
71
-
gated behind a major-version bump.
84
+
**Acceptance:**
85
+
86
+
*Shippable artifacts:*
87
+
- Frozen baseline tag for CLI flags, REST endpoints (`api/aicr/v1/server.yaml`), Go SDK exported surface (`pkg/client/v1`), and bundle layout + schemas (recipe, validation, snapshot kinds).
88
+
- CI diff-gate that fails on unintended breakage of any of the four surfaces.
89
+
- Deprecation channel documented in `RELEASING.md`.
90
+
91
+
*Demonstrable signals:*
92
+
- Sustained external use of the Go SDK, REST API, and schemas by consumers outside the AICR team, exercised against the frozen surface across release cycles without breaking-change requests.
93
+
- Deprecation channel exercised at least once in practice before tagging `v1.0.0`.
72
94
73
95
### 3. UX
74
96
@@ -88,9 +110,21 @@ contributors aren't copy-pasting drift into the tree.
88
110
the review pipeline fast and predictable; a local registry for nvkind keeps the
89
111
inner-loop tight for contributors without cloud access.
90
112
91
-
**Acceptance:** an external contributor adds a new validated leaf recipe using
92
-
only public docs and the CLI, and the review pipeline merges community PRs at the
93
-
cadence the inbound flow demands.
113
+
**Acceptance:** the bar is *self-service without filing a support issue or
114
+
reading source code* for three personas — **user**, **integrator**, and
115
+
**contributor**.
116
+
117
+
*Per-persona self-service thresholds:*
118
+
-**User** — install through bundle generation, deployment validation, and evidence collection completable using only `docs/user/`. The same path eventually submits that evidence to the Testgrid pipeline described in §1.
119
+
-**Integrator** — Go SDK or REST API embeddable into a downstream pipeline using only godoc and `docs/user/api-reference.md`.
120
+
-**Contributor** — new or modified leaf recipes authorable and mergeable using only `docs/contributor/` plus CLI tooling, and every recipe change carries validation evidence that the CI gate verifies before merge.
121
+
122
+
*Shippable artifacts behind the thresholds:*`aicr recipe validate` verb, scaffolding generator, reusable overlay template, KWOK e2e auto-generation for new leaves, mixin-based composition discipline, merge queue + unified merge-gate, recipe-change CI gate that requires verifiable validation evidence.
123
+
124
+
*Demonstrable signals:*
125
+
- A community-authored leaf recipe merged end-to-end without maintainer hand-holding.
126
+
- An independent user deploys a bundle and submits evidence through the documented pipeline.
127
+
- An external integrator reference exists (in-tree example, downstream repo link, or recorded user story).
94
128
95
129
### 4. Security
96
130
@@ -116,17 +150,18 @@ as the binaries that read it.
116
150
**End-user verification.** A top-level guide that walks a consumer from a
117
151
downloaded bundle to a fully-verified deploy, covering all signing modes above.
118
152
119
-
**Acceptance:** every artifact (binary, image, recipe data, bundle) has retrievable
120
-
provenance and a documented verification path that works in air-gapped, private,
121
-
and public-trust deployments.
153
+
**Acceptance:**
154
+
155
+
*Shippable artifacts:*
156
+
- SLSA attestation on every artifact (binary, image, recipe data, bundle) with cosign signing.
157
+
-`aicr bundle verify` honoring trust-level evaluation across signing modes.
158
+
- Air-gapped signing flow, private Sigstore deployment, and KMS-backed key custody documented and tested.
159
+
- Verify-on-deploy path documented and tested.
160
+
- Signed validator catalog and component registry.
161
+
- End-user verification guide published in `docs/user/`.
122
162
123
-
## Revision History
163
+
*Demonstrable signals:*
164
+
- Consecutive releases pass the full provenance chain in CI.
165
+
- An independent party verifies a release end-to-end across the supported signing modes.
124
166
125
-
| Date | Change |
126
-
|------|--------|
127
-
| 2026-04-28 | Restructured around v1 objectives |
128
-
| 2026-02-17 | Expanded Recipe Creation Tooling with validation framework details, scaffolding, and workflow integration |
129
-
| 2026-02-14 | Moved implemented items to Completed: EKS H100 recipes, snapshot-to-recipe, monitoring, Nodewright Ubuntu |
0 commit comments