Track non-hardware release roadmap

.github/ISSUE_TEMPLATE/software_improvement.yml

@@ -0,0 +1,36 @@
+name: Software improvement
+description: Track non-hardware improvements to GPT2-BASIC runtime, packs, storage, or evidence.
+title: "[Software]: "
+labels: ["enhancement", "release-follow-up"]
+body:
+  - type: textarea
+    id: goal
+    attributes:
+      label: Goal
+      description: What software behavior, capability, or evidence should improve?
+    validations:
+      required: true
+  - type: textarea
+    id: scope
+    attributes:
+      label: Scope
+      description: Files, packs, scripts, or release artifacts likely affected.
+    validations:
+      required: true
+  - type: textarea
+    id: evidence
+    attributes:
+      label: Evidence gate
+      description: Commands, reports, or QEMU logs that should prove the improvement.
+      render: shell
+    validations:
+      required: true
+  - type: checkboxes
+    id: boundaries
+    attributes:
+      label: Boundaries
+      options:
+        - label: This does not claim new physical hardware validation.
+          required: true
+        - label: Public docs and release notes stay clear about QEMU versus physical evidence.
+          required: true

docs/assistant-intelligence-roadmap.md

@@ -25,6 +25,21 @@ See [`substrate-portability.md`](substrate-portability.md) for the minimum
 runtime-primitives argument behind the BASIC implementation and future C,
 assembly, and Eshkol ports.
 
+## Current Evidence Baseline
+
+- CHAT pack quality: `PASS 160/160`, average `0.999`.
+- Raw direct prompt gate: `PASS 83/83`.
+- Generalist conversational prompt gate: `PASS 24/24`.
+- Consistency gate: `PASS 498/498` variants across `83/83` groups.
+- Pack retrieval, KDB index, KDB binary, and KB2 term-index gates: `PASS 42/42`.
+- Usefulness workflow gate: `PASS 37/37` tasks across `9/9` workflows.
+- QEMU assistant stress gate: `PASS 50/50` replies across five packs.
+- Runtime bundle verification: preview, DOSBox, hardware-transfer, and launch-kit
+  zips rebuild with sidecar checksums and without host-only training corpora.
+
+Physical 486-class validation remains outside this roadmap pass and is tracked
+separately in public issues #1 and #2.
+
 ## Design Direction
 
 The target experience is a cartridge-like language system:
@@ -64,9 +79,23 @@ significant query words before falling back to the full KDB.
 
 ## Next Milestones
 
-- Add more domain packs for hardware repair, programming, and offline reference
-  manuals using the same generated KDB/KB2 contract.
-- Compare recall-probe timing on QEMU and physical hardware, then decide
-  whether the next storage step should be topic shards or offset tables.
-- Add persistent memory slots beyond name, goal, style, and problem.
-- Add a measured recall benchmark in QEMU and on physical 486 hardware.
+1. Build a denser binary term index for local recall:
+   https://github.com/tsotchke/gpt2-basic/issues/37
+2. Add latency and recall-budget gates to the assistant harness:
+   https://github.com/tsotchke/gpt2-basic/issues/41
+3. Add an inspectable persistent conversation database:
+   https://github.com/tsotchke/gpt2-basic/issues/38
+4. Add a pack intent router for assistant queries:
+   https://github.com/tsotchke/gpt2-basic/issues/39
+5. Add larger programming and offline-reference domain packs:
+   https://github.com/tsotchke/gpt2-basic/issues/40
+6. Define Windows 95 and OS/2 assistant shell parity:
+   https://github.com/tsotchke/gpt2-basic/issues/42
+7. Repair and promote smaller model profiles under strict gates:
+   https://github.com/tsotchke/gpt2-basic/issues/43
+8. Prototype a C or assembly fixed-point kernel compatibility path:
+   https://github.com/tsotchke/gpt2-basic/issues/44
+
+The ordering is deliberate: recall density and latency budgets should come
+before larger packs, and persistent memory plus routing should land before
+native shell ports so every shell can target the same behavior.

docs/public-launch.md

@@ -138,3 +138,15 @@ scale. That honesty makes the demo stronger.
    https://github.com/tsotchke/gpt2-basic/issues/2
 3. Improve real-hardware interaction ergonomics: quiet prefill, compact prompts,
    optional continue mode, and smaller output-shortlist sweeps.
+
+Non-hardware follow-up is tracked separately so software work can continue
+without waiting for physical captures:
+
+- Dense recall storage: https://github.com/tsotchke/gpt2-basic/issues/37
+- Persistent conversation database: https://github.com/tsotchke/gpt2-basic/issues/38
+- Pack intent router: https://github.com/tsotchke/gpt2-basic/issues/39
+- Larger domain packs: https://github.com/tsotchke/gpt2-basic/issues/40
+- Latency and recall budgets: https://github.com/tsotchke/gpt2-basic/issues/41
+- Windows 95 and OS/2 shell parity: https://github.com/tsotchke/gpt2-basic/issues/42
+- Smaller model-profile repair queue: https://github.com/tsotchke/gpt2-basic/issues/43
+- C or assembly fixed-point kernel compatibility: https://github.com/tsotchke/gpt2-basic/issues/44

docs/substrate-portability.md

@@ -123,6 +123,11 @@ Eshkol belongs above that layer: orchestration, richer local workflows,
 tooling, and interfaces. GPT2-BASIC supplies the minimum-substrate witness.
 C and assembly supply the production portability path.
 
+The first scoped compatibility task is public issue #44:
+https://github.com/tsotchke/gpt2-basic/issues/44. It should start with one
+fixed-point kernel and preserve the existing vector/evidence contract before
+claiming any broader port.
+
 ## Proof Obligations
 
 To claim a new substrate, the project should provide evidence for:

qemu/evidence/preview_release_manifest.md

@@ -5,7 +5,7 @@ Generated: `2026-05-12`
 Package tree: `gpt2-basic-preview`
 Package zip: `gpt2-basic-preview.zip`
 Package checksums: `SHA256SUMS.txt`; zip sidecar: `gpt2-basic-preview.zip.sha256`
-Package status: `705 files, 120,011,903 bytes`
+Package status: `706 files, 120,014,180 bytes`
 
 This is an iterative preview payload. It ships only strict-quality release models and assistant packs; rejected repair attempts and old candidates remain repo evidence only.
 

scripts/build_preview_release.py

@@ -77,6 +77,7 @@ class ReleaseModel:
     ".github/ISSUE_TEMPLATE/bug_report.yml",
     ".github/ISSUE_TEMPLATE/config.yml",
     ".github/ISSUE_TEMPLATE/hardware_validation.yml",
+    ".github/ISSUE_TEMPLATE/software_improvement.yml",
     ".github/workflows/preview-release.yml",
 )
 TREE_DIRS = ("src", "scripts", "tests", "data/domain_curriculum", "data/assistant_pack_notes", "hardware")