From bab0dd3645190949eb1c2e43cb15ebffc7618824 Mon Sep 17 00:00:00 2001 From: Timothy Wayne Gregg Date: Sat, 3 Oct 2026 09:58:41 -0400 Subject: [PATCH] Reconcile roadmap and source-linked delivery evidence --- docs/BUILD_REPORT.md | 63 ++++++++++++++++++++++++++++++++++++++++++++ docs/ROADMAP.md | 26 +++++++++--------- 2 files changed, 77 insertions(+), 12 deletions(-) diff --git a/docs/BUILD_REPORT.md b/docs/BUILD_REPORT.md index be1a050..fd93d68 100644 --- a/docs/BUILD_REPORT.md +++ b/docs/BUILD_REPORT.md @@ -140,3 +140,66 @@ Not executed: any live TypeSafe or teacher request, real GEPA optimization, DSPy patch releases above 3.4.0, `dspy[typesafe]`/`ReAnchor`/native decision types, Python 3.11-3.13 and non-Windows runs, remote CI, `build`/wheel/demo under 3.4.0, and a re-read of primary documentation (the inspection date is unchanged). + +## October 3, 2026 organization delivery checkpoint + +These Linux Python 3.14.7 receipts extend the historical records above; they do +not replace Windows, live-provider, or independent quality qualification. The +exact optional profiles used DSPy 3.3.1 or 3.4.0, GEPA 0.1.4, TypeSafe SDK 0.7.0, +LiteLLM 1.103.2 and Pydantic 2.13.5. No paid inference or private sharing was +authorized. + +- [PR #99](https://github.com/Jev-Engineering/TypeWright/pull/99), triage + regressions: head `a7ab0042957f67fc857981ba9ae62775a9604355`, merge + `5165fb6d8d05307e7b1467f5fd2bbb34cdf8e5d4`; isolated full suite **618 passed**, + zero skipped; Ruff and independent review passed. +- [PR #100](https://github.com/Jev-Engineering/TypeWright/pull/100), owner-approved + hierarchy tracking split: merge `0563956cbac62bd899d5fddad734f84a4ae608e2`. + Software acceptance for #20/#21 is complete; their original live criteria remain + explicitly open in #74 and #6. Independent scope/state review passed. +- [PR #101](https://github.com/Jev-Engineering/TypeWright/pull/101), opt-in native + calibration fitting: head `c316283c8fc612990f74246958edaab9aaa4ec08`, merge + `aa448275822e69a6356ee32ac36ba3ffa73ab950`; isolated full suite **640 passed**, + zero skipped. Calibration isolation, unchanged results and independently + committed legacy artifact/checksum compatibility were verified. +- [PR #102](https://github.com/Jev-Engineering/TypeWright/pull/102), teacher raw + completion metering: head `f9080cba8416a88642bf8ed14d91bdbd505c523d`, merge + `67a53ee047dd9a14a70abdf13641974087d42457`; isolated full suites **628 passed**, + zero skipped, on each DSPy profile. An early inherited legacy-completion path + accidentally attempted an uncredentialed synthetic provider request and received + HTTP 401. That failure is retained; composition and loopback/socket guards + replaced the bypass before the successful final suites. No paid inference or + private examples were sent. +- [PR #103](https://github.com/Jev-Engineering/TypeWright/pull/103), compile-only + native adapter: head `de7c82dd3fe5ba3b81ceceb49cab07db206c4bae`, merge + `821b29d627b602b0b042b474fd5a2248febd5c74`; isolated full suite **685 passed** on + DSPy 3.4.0; **667 passed, 18 skipped** on 3.3.1 because native qualification + requires exactly 3.4.0. The historical dependency assertion initially failed + after `all` selected the exact native profile; the revised test preserves the + separate teacher range and both full reruns passed. Independent source, wiring + and dependency review passed. Ruff, build, 51 wheel source members, all four + entrypoints, fresh no-key demos and the synthetic study journey passed. Runtime + and adapter-module imports loaded no DSPy, GEPA or SDK. + +Other retained attempts: #99 initially failed two subprocess imports without the +checkout's explicit `PYTHONPATH`; a mutable-worktree attempt during edits also +failed. The stable isolated rerun passed. #101 had an interrupted mutable-root +run and a nested replay identity failure while a concurrent build created +package metadata. A stable post-build causal check and full suite passed. Early +installed-wheel study attempts combined incompatible fixtures and were correctly +rejected; the final complete fixture passed. These attempts are separate from +the successful final acceptance receipts above. + +Each PR above completed **10 successful exact-head hosted checks** and +**five successful checks observed on its actual merge commit**. For #103 the +[merge workflow](https://github.com/Jev-Engineering/TypeWright/actions/runs/37127560430) +ran on `821b29d627b602b0b042b474fd5a2248febd5c74`; this is executed remote +evidence, unlike the earlier shipped-configuration checkpoint. Local main was +fast-forwarded to that merge without changing original user files or deleting +retained packet/worktree evidence. + +NOT_RUN: live native provider qualification, representative independent quality +studies, hierarchy paid selection/teacher sharing/held-out dispatch, production +activation, administrative CI/protection changes and a fully resolved dependency +lock. The gateway destination patch in PR #98 still awaits the owner's endpoint +policy decision. Synthetic checks do not establish gains or production approval. diff --git a/docs/ROADMAP.md b/docs/ROADMAP.md index 44fd721..c433ade 100644 --- a/docs/ROADMAP.md +++ b/docs/ROADMAP.md @@ -8,8 +8,10 @@ rather than capabilities claimed by the current release. 1. **Verify and lock a real environment.** Install the optional SDKs, run their contracts and authorized inference smoke tests, generate a real dependency lock, and add sanitized regression fixtures from observed response schemas. This - includes DSPy 3.4, which is untested here and changes LM internals and the - Pydantic minimum ([#90](https://github.com/Jev-Engineering/TypeWright/issues/90)). + includes future SDK/DSPy releases beyond the verified profiles. DSPy 3.3.1 + and 3.4.0 teacher contracts are verified offline; native compilation requires + the exact DSPy 3.4.0 / SDK 0.7.0 profile. Live provider qualification and a + resolved transitive lock remain outstanding ([#90](https://github.com/Jev-Engineering/TypeWright/issues/90)). 2. **Establish useful benchmarks.** The BANKING77 and multi-benchmark protocols and the hierarchy study runner exist, but their live phases have not been executed and the hierarchy study still needs independently reviewed inputs. @@ -33,15 +35,15 @@ rather than capabilities claimed by the current release. 5. **Beginner interface.** A local wizard for labels and rubrics, improved failure cluster visualization, provider-budget preview, and editor support from schemas. -6. **Align with DSPy's Jev support.** DSPy 3.4 ships an experimental `TypeSafe` - client, decision types, and the `ReAnchor` optimizer. Evaluate moving - compile-time evaluation onto DSPy's adapter while keeping this project's - budget, model-identity, and consent guarantees, and decide deliberately - whether the frozen runtime keeps its own backend so it never imports DSPy - ([#91](https://github.com/Jev-Engineering/TypeWright/issues/91), after #90). - A generative-LM baseline run through DSPy's decision types is not planned at - present. +6. **Maintain native integration qualification.** Compile-time native evaluation + now uses the verified DSPy 3.4.0 `TypeSafe` wrapper; frozen execution and + held-out evaluation retain the direct SDK and never import DSPy + ([#91](https://github.com/Jev-Engineering/TypeWright/issues/91)). Recheck full + payload/envelope preservation, identity, consent, zero retries and budgets + against upstream changes before expanding its exact version guard. This + offline software qualification does not establish live quality or justify + replacing the approved native policy fitting with `ReAnchor`. A generative-LM + baseline through DSPy's decision types remains outside the current plan. -Do not add a user interface ahead of verifying the native adapter and measuring a -real task. Do not make “improved prompts” a promise independent of data quality and +Do not add a user interface ahead of measuring a real task. Do not make “improved prompts” a promise independent of data quality and held-out results.