Generate assistant capability report from evidence

README.md

@@ -378,6 +378,11 @@ KEY=VALUE`, and `/forget` in the shell. `SPRITE=` and `ICONS=` fields are
 reserved for Clippy-style
 artwork; the current renderer is a text-mode bubble/action UI so it works
 without VGA.
+Regenerate the evidence-backed capability/functionality report with:
+
+```sh
+python3 scripts/build_assistant_capability_report.py
+```
 
 Run the non-greedy sampling matrix with:
 

docs/releases/v0.1.0-preview.md

@@ -176,6 +176,7 @@ python3 scripts/evaluate_assistant_kdb_binary.py
 python3 scripts/evaluate_assistant_kdb_term_index.py
 python3 scripts/import_assistant_notes.py --self-test
 python3 scripts/evaluate_assistant_consistency.py
+python3 scripts/build_assistant_capability_report.py
 QEMU_TIMEOUT_SECONDS=240 bash qemu/run_assistant_stress_486.sh
 python3 scripts/stress_assistant_behavior.py --log qemu/evidence/assistant_stress_486.log
 python3 scripts/verify_workspace_tracking.py

qemu/evidence/assistant_capability_functionality_report.md

@@ -3,22 +3,23 @@
 Date: 2026-05-21
 Status: `PASS`
 
+This report is generated from repository evidence files by `scripts/build_assistant_capability_report.py`.
+
 ## Runtime Capability
 
-- Runs under FreeDOS/QEMU 486 with five assistant packs: `CHAT`, `DOSHELP`, `OFFICE`, `DEV`, and `PORTABLE`.
+- Runs under FreeDOS/QEMU 486 with 5 assistant packs: `CHAT`, `DOSHELP`, `OFFICE`, `DEV`, `PORTABLE`.
 - Supports hot pack switching through `PACKS.TXT` and each pack's `PACK.INI`.
 - Supports pack-local model paths, pack-local art assets, pack-local golden rows, pack-local help/knowledge rows, and editable `USER.TXT` notes.
-- Uses retrieval-first answering before model synthesis: golden rows, compiled knowledge recall, session memory, and fallback checks are all explicit in `ASSIST_REPLY`.
+- Uses retrieval-first answering before model synthesis: golden rows, compiled knowledge recall, session memory, and fallback checks are explicit in `ASSIST_REPLY`.
 - Reports structured provenance and timing for every reply: `source`, `recall`, `recall_score`, `t_retrieve_ms`, `t_golden_ms`, `t_memory_ms`, `t_model_ms`, and `t_total_ms`.
-- Interactive shell exposes `/capabilities`, `/limits`, `/sources`, `/status`, `/about`, and `/pack`.
-- The answer display now includes a compact source line such as `Source: golden / kb2_term ( 60 ms)`.
+- Interactive shell exposes `/capabilities`, `/limits`, `/sources`, `/status`, `/about`, `/pack`, `/memory`, `/remember KEY=VALUE`, and `/forget`.
 
 ## Recall And Storage
 
 - Text KDB remains the readable source/fallback format: `KDB.TXT`, `KDBIDX.TXT`, and `KDB?.TXT`.
-- New compiled KB2 recall is shipped for each pack: `KB2ALL.BIN`, `KB2IDX.TXT`, `KB2?.BIN`, and `KB2TERM.TXT`.
+- Compiled KB2 recall ships for each pack: `KB2ALL.BIN`, `KB2IDX.TXT`, `KB2?.BIN`, and `KB2TERM.TXT`.
 - KB2 files use fixed-width records for 486-friendly sequential reads and avoid reparsing large text rows during recall.
-- `KB2TERM.TXT` is a compact per-pack inverted term index. The DOS runtime uses it to score likely row IDs first, then falls back to binary buckets and finally text KDB recall.
+- `KB2TERM.TXT` is a compact per-pack inverted term index. The DOS runtime scores likely row IDs first, then falls back to binary buckets and finally text KDB recall.
 - Current compiled KB2 payload sizes:
   - `CHAT`: 78 rows, 23 buckets, 159616 binary bytes, 4280 term-index bytes.
   - `DOSHELP`: 26 rows, 21 buckets, 55488 binary bytes, 2193 term-index bytes.
@@ -27,7 +28,7 @@ Status: `PASS`
   - `PORTABLE`: 11 rows, 16 buckets, 23968 binary bytes, 1292 term-index bytes.
 - Binary recall evaluation: `PASS 42/42`.
 - Binary candidate row scan ratio: `0.531`.
-- Binary candidate byte ratio: `0.689`.
+- Binary candidate byte ratio: `0.688`.
 - Term-index recall evaluation: `PASS 42/42`.
 - Term-index candidate row scan ratio: `0.145`.
 - Term-index candidate byte ratio: `0.315`.
@@ -43,19 +44,9 @@ Status: `PASS`
 - KDB binary gate: `PASS 42/42`.
 - KDB term-index gate: `PASS 42/42`.
 
-Covered categories include:
-
-- General chat, identity, local inference, local limits, offline/no-web behavior, prompt quality, repeated-answer recovery, confidence framing, simple explanation, and lightweight planning.
-- Troubleshooting, debugging, release checks, DPMI/CWSDPMI, CONFIG.SYS, AUTOEXEC.BAT, FAT image limits, QEMU logs, and real-hardware copy preparation.
-- Rewriting, summarizing, shortening, release notes, status updates, handoff notes, bug reports, meeting notes, risk registers, project plans, customer replies, and user docs.
-- Developer-pack guidance for retrieval-first design, authoring packs, fast recall storage, release checks, failure records, and modern 486 assistant architecture.
-- Portable-intelligence guidance for BASIC teaching, C/assembly/Eshkol ports,
-  hot-swappable weights, compact recall, and old-hardware proof.
+Covered categories include general chat, identity, local inference, offline limits, prompt repair, repeated-answer recovery, troubleshooting, DOS setup, office writing, developer pack authoring, and portable-intelligence concepts.
 
-Usefulness workflows currently cover operator prompts, trust/offline limits, DOS
-setup and repair, hardware transfer and emulator evidence, office handoffs,
-planning and risk, developer pack authoring, fast local recall architecture,
-and portable intelligence.
+Usefulness workflows currently cover operator prompts, trust/offline limits, DOS setup and repair, hardware transfer and emulator evidence, office handoffs, planning and risk, developer pack authoring, fast local recall architecture, and portable intelligence.
 
 ## DOS/QEMU Stress Result
 
@@ -65,27 +56,33 @@ and portable intelligence.
 - Stress source mix: `golden=26 retrieval=16 model=0 fallback=0 memory=8`.
 - Average total reply time in the stress report: `134 ms`.
 - Average retrieval time in the stress report: `80 ms`.
-- Recall modes in the stress report: `kb2_term=46 kb2_bucket=3 none=1`.
+- Recall modes in the stress report: `kb2_bucket=3 kb2_term=46 none=1`.
 - Visible-answer validation: `PASS`.
 
+## Hardware-Capture Rehearsal
+
+- QEMU rehearses the physical `C:\GPT2\HWVALID.BAT` path before real transfer.
+- Hardware-capture rehearsal: `PASS`.
+- Hardware-capture assistant stress replies: `50`.
+- Hardware-capture stress source mix: `golden=26 retrieval=16 model=0 fallback=0 memory=8`.
+- Hardware-capture average total reply time: `28 ms`.
+- Hardware-capture average retrieval time: `24 ms`.
+- Physical machine capture status: PENDING: no staged physical `hardware_<machine>_manifest.md` capture is present yet.
+
 ## Authoring And Import
 
 - `scripts/import_assistant_notes.py` can import ASCII notes into `USER.TXT` or `KNOW.TXT`.
-- Import is dry-run by default.
 - `--target user` writes machine-local notes without changing bundled pack knowledge.
 - `--target know --rebuild-kdb` updates bundled pack knowledge and regenerates KDB/KB2 artifacts.
-- `scripts/create_assistant_pack.py` can create a complete lightweight pack
-  from a folder of ASCII notes, sharing `PACKS\CHAT\MODEL` by default.
-- The pack generator writes `PACK.INI`, authoring files, `USER.TXT`,
-  `USAGE.TXT`, generated KDB buckets, compiled KB2 pages, and `KB2TERM.TXT`.
+- `scripts/create_assistant_pack.py` can create a complete lightweight pack from a folder of ASCII notes, sharing `PACKS\CHAT\MODEL` by default.
+- The pack generator writes `PACK.INI`, authoring files, `USER.TXT`, `USAGE.TXT`, generated KDB buckets, compiled KB2 pages, and `KB2TERM.TXT`.
 - Authoring validator checks required pack files, source rows, generated text KDB, generated binary KDB, and model references.
 
 ## Release Payload
 
+- Preview package manifest: `included`.
 - Preview release tracked-input gate: `PASS`.
 - Preview artifact verifier: `PASS`.
-- DOSBox zip unzip test: `PASS`.
-- Launch-kit zip unzip test: `PASS`.
 - Release sidecar hashes: `PASS`.
 - Runtime bundles exclude host-only `TRAIN.TXT` and `TOKBASE.TXT`.
 
@@ -96,6 +93,7 @@ and portable intelligence.
 - Long, ambiguous, or out-of-domain prompts should be shortened or moved into an appropriate pack.
 - No live web, news, package registry, or network lookup is available inside DOS.
 - Current 486 stress replies did not require raw model generation; that is intentional for reliability and speed on this hardware class.
+- Physical 486-class board evidence is still pending until real hardware returns the `HWVALID.LOG`, `QUAL.LOG`, `PERF.LOG`, `ASSIST.LOG`, `ASTRESS.LOG`, `ASSISTC.LOG`, and `HWNOTES.TXT` set.
 
 ## Next Production Targets
 

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: `581 files, 119,878,941 bytes`
+Package status: `583 files, 119,897,176 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.