Roadmap

Roadmap

The live execution plan is docs/PLAN.md — that file is the record of truth and moves faster than this page. Founding intent: docs/PLANNING.md (§§1–80). Binding overrides: docs/DECISIONS.md.

Related pages: Releases (what shipped) · Models (the training program's next levers) · Playground (site work).

• Phase status
• Open work
• How phases work

Phase status

Status as recorded in docs/PLAN.md (phase table, current at the v0.6.1 release):

Phase	Scope	Status
0	Workspace scaffold, governance, CI configs	done
1	opencodifier-core canonical IR	done
2	opencodifier-schema adapters + fixtures	done
3–5	opencodifier-engine: graphs, rules, caches, narrowing, deterministic model + calibration seam	done
6	opencodifier-runtime traits + honest ONNX feasibility gate	done — measured gate deferred onnx (D2; later reinstated for the Kai/Julia rungs, D33/D34)
7	opencodifier-model serving contract + embedding classifier	done
8	opencodifier-cli	done — D13 exit codes
9	opencodifier-http	done — axum 0.8, loopback gate
10	opencodifier-mcp	done — rmcp 2.2 stdio, six tools
11	E2E recipes + docs book	done
12	Release engineering	done — v0.1.0; pipeline of record green through GitForge
13	Decision-model benchmark	done — D16 tier scheme
14	Confidence truthfulness (D15 calibration + §19 gates)	done
15	Relational solver (default zero-ML stack)	done
16	Focused-question extraction + reverse escalation	done — D18
17	External anchor: JevBench public split	done — docs/BENCHMARKS.md is the page of record
18	Hardening pass (coverage floor, e2e, D19–D28)	done
19	Adoption + integration documentation	done — docs/INTEGRATIONS.md, D30/D31
20	Coverage restoration (escalation core)	done — 98.39 % in-image DA; ratchet retired at the reachable ceiling; floor rests at 97.5 fixed
23	v0.5.0 release cut	done — 9-leg matrix, single-tree attestations
21	Site WCAG 2.1 AAA + progressive enhancement	opened — audit + fixes landed; deploy + X/Discord presence pending
22	Docs freshness sweep	in progress — 2026-10-15 sweep landed (SECURITY matrix, D11 axum pin, §36 route amendment, itemgen in ARCHITECTURE, site counts); wiki cross-checks ride the same phase
24	RESEARCH §15.6 enhancement program	in progress — items 1–4, 8–10 + §15.8 fold-ins done; items 5/6 sized, deferred per the §15.6 note

(Phase 22 exists because docs lag code — the phase table above will drift from docs/PLAN.md; the in-repo table wins.)

Open work

From the plan:

• Phase 21 — site accessibility follow-through: the remaining deploy + ecosystem-presence items (verified links before handles ship — the site's no-guessed-links rule).
• Phase 22 — the docs freshness sweep this wiki accompanies.
• Phase 24 — the §15.6 enhancement program: remaining sized items are NLI-first (24j/24k); the model-rung options and the verifier path stage behind feature gates until their parity legs close.

The vision rung (designed, not started). A requirement, not a wish: competing decision rungs expose image input, and OpenCodifier needs it as an optional lever that routes only when needed — comprehensive over the image-bearing decision surface (screenshot/chart/document classification, table extraction routing) and fast on CPU. The shape follows the binding rules: an ImageBackend seam beside InferenceBackend/EmbeddingBackend in opencodifier-runtime (ONNX behind the feature gate, per D11 — no Python at runtime, §73); the rung sits above lexical/embedding in the escalation ladder and fires only on image-bearing requests; outputs are candidate-conditioned typed decisions, never free-form generation; confidence passes the mandatory calibration gate (raw softmax is never exposed); cache keys carry the vision model + calibration versions like every other cache key; models are explicit opt-in pulls, never auto-downloaded (the release policy). "Comprehensive and fast" is sized honestly: small quantized vision encoders (SigLIP-class) through the same ONNX CPU path, with the verifier contract (agree → accept, disagree → abstain) mirrored from the text ladder. Sizing and the first parity fixtures are the open work.

From the training program (docs/TRAINING.md, evidence-ranked next levers): consensus-flip patch / strands consensus pass (quality-first), the T3.4 4B contingency, and the #150 eval-lane-v2 work (the CPU fork lane's determinism-replay gate needs a load-pinned window or a tolerance definition — recorded, not hand-waved).

Deliberately not scheduled: registry publication to crates.io/Homebrew/winget (explicit operator go — D24), a default-on escalation ladder (per-domain calibration precedes it — D25/F24), and any browser model path (needs its own integrity + supply-chain record — D23).

How phases work

Every phase carries acceptance criteria in docs/PLAN.md, and every landed phase passes the same gates (Architecture, Contributing). The plan records honest status — including measured failures and falsified levers — because a plan that cannot say "this did not work" cannot be trusted to say what did.