From c036c6c8c31e73403155ca467dd92a2046a43e7a Mon Sep 17 00:00:00 2001 From: CacinieP <129482851+CacinieP@users.noreply.github.com> Date: Sat, 14 Nov 2026 18:59:53 +0800 Subject: [PATCH 1/2] docs: complete TeXada release history and sync v0.4.10 guidance --- CHANGELOG.md | 64 +++++++++++++--- README.md | 149 ++++++++++++++++--------------------- docs/README.md | 4 +- docs/linux-installation.md | 8 +- 4 files changed, 130 insertions(+), 95 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index a715b4e..0fd78b4 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,37 +1,64 @@ # Changelog / 变更记录 +This history includes stable releases, prerelease tags and explicitly marked development snapshots. / 本记录覆盖正式版、预发布标签和明确标注的开发快照。 + ## Unreleased ## 0.4.10 - 2026-11-14 -- 首次启动按系统语言选择界面,中文环境默认中文,其余环境默认英文。已保存的语言偏好优先,调整缩放不会固定自动选择的语言。 -- Choose the first-run UI language from the system locale and fall back to English. Preserve saved language preferences independently of zoom. +### 中文 + +- 首次启动按系统语言选择界面。Linux 按 `LC_ALL`、`LC_MESSAGES`、`LANG` 的优先级判断,macOS 和 Windows 使用 WebView 的语言环境。中文环境默认中文,其余环境默认英文。显式保存的语言优先,单独调整缩放会继续沿用自动语言选择。见 [#90](https://github.com/CacinieP/TeXada-the-Math-Agent/pull/90)。 +- 正式版 Ubuntu 22.04 安装检查验证了 deb 和 AppImage 的离线启动、英文首屏和确定性公式渲染。独立 Ubuntu 24.04.5 的英文首屏、CPU 文本转写和 OCR 实测使用的是同一修复分支的 v0.4.9 CI 候选包。见 [发布工作流](https://github.com/CacinieP/TeXada-the-Math-Agent/actions/runs/43581742151)和 [Ubuntu 24.04 实测](docs/validation/2026-11-14-ubuntu24-appimage/README.md)。 +- AppImageHub 复测通过的是正式发布的 `TeXada_0.4.10_amd64.AppImage`,报告记录 glibc 2.35;目录收录仍待维护者审核。见 [上游测试报告](https://github.com/AppImage/appimage.github.io/pull/11396#issuecomment-6875598626)。 -- Verify the English first window from the actual Linux AppImage and deb with a fresh profile under Xvfb. See [Ubuntu 24.04 validation](docs/validation/2026-11-14-ubuntu24-appimage/README.md). +### English + +- Select the first-run UI language from the system locale. On Linux, check `LC_ALL`, `LC_MESSAGES` and `LANG` in that order. On macOS and Windows, use the WebView locale. Chinese locales select Chinese, and other locales select English. An explicit language choice takes precedence, and changing zoom alone preserves automatic selection. See [#90](https://github.com/CacinieP/TeXada-the-Math-Agent/pull/90). +- Verify offline startup, the English first window and deterministic formula rendering for the released Ubuntu 22.04 deb and AppImage. The separate Ubuntu 24.04.5 English-window, CPU transcription and OCR run used a v0.4.9 CI candidate from the locale-fix branch. See the [release workflow](https://github.com/CacinieP/TeXada-the-Math-Agent/actions/runs/43581742151) and [Ubuntu 24.04 hardware validation](docs/validation/2026-11-14-ubuntu24-appimage/README.md). +- AppImageHub passed the public `TeXada_0.4.10_amd64.AppImage` with a glibc 2.35 baseline. Catalog review is pending. See the [upstream test report](https://github.com/AppImage/appimage.github.io/pull/11396#issuecomment-6875598626). + +[GitHub release / 正式发布](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.10) · [Compare / 版本差异](https://github.com/CacinieP/TeXada-the-Math-Agent/compare/v0.4.9...v0.4.10) ## 0.4.9 - 2026-11-13 +### 中文 + - 新增术语名词表 `src/texada/core/glossary.py`(24 条,含格林/高斯/斯托克斯/贝叶斯/全概率/华里士/施密特/拉格朗日/罗尔/柯西/泰勒/洛必达/一阶线性 ODE/二阶常系数/中心极限/样本方差/欧拉/夹逼/路径无关/二项分布/分部积分/N-L/柯西不等式/秩不等式),每条带名称别名、LaTeX、出处与备注。 - 词表沿既有确定性快路径接入:`DeterministicCandidateEngine._canonical_concept` 兜底查表(规则名仍为 `nl_canonical_concept`),`/api/convert` 在同一处先查表,命中返回 `Source.TEMPLATE`、`tokens_used=0`。匹配要求规范化后与别名完全相等,带参数的句子与复合问句仍交给模型;每条 LaTeX 必须通过 KaTeX 校验,失败则回退模型路径。 -- 实测(MiniCPM-V 4.7 1B):真实语料 31/52 → **47/52**,其中 `named-formula` 5/20 → **20/20**;52 条里 21 条零 token 直出,词表命中 p50 ≈ 12 ms,模型路径 ≈ 0.8 s。100 条 golden set 仍为 59/95,未借词表刷分。 +- 实测(MiniCPM-V 4.7 1B):真实语料从 v0.4.7 的 31/52 升至 **47/52**,其中 `named-formula` 从 5/20 升至 **20/20**;52 条中 21 条由模板直出,模板路径 p50 为 12.3 ms。100 条 golden set 的结构通过率仍为 59/95,整套运行的 p50/p95 为 1255/3325 ms。见[真实语料逐条结果](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/real-world-baseline.json)和[黄金集报告](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/reports/2026-11-13-baseline-live.md)。 - 评测 runner 现在区分「确定性直出」与「模型结果」并单列条目,避免把查表得分读成模型能力;README 中英双段与 `docs/architecture.md`、`eval/known-gaps.md`(G-017)同步。 -- Serve bare theorem names from a curated glossary (24 entries) on the existing deterministic-candidate path, so the small model is no longer asked to recall textbook formulas. +### English + +- Add a 24-entry formula glossary with aliases, LaTeX and provenance. A normalized exact match on a bare name uses the existing deterministic candidate path, including `/api/convert`. A failed KaTeX check falls back to the model. Parameterized and compound requests continue through the model path. +- On the 52-case real-world corpus, the MiniCPM-V 4.7 1B configuration went from 31/52 in v0.4.7 to 47/52. Named formulas went from 5/20 to 20/20 via the glossary. The evaluator records template, model and agent paths separately. The 100-case golden set retained 59/95 structural passes. ## 0.4.8 - 2026-11-13 +### 中文 + - 修复计算范围门禁的误拒:`u 除以 v 的求导法则`、`泰勒公式的求导法则`、`基本积分公式表` 这类"要排版的内容"此前因动词 `求导` 命中而被整句拒绝。现在动词扫描前先剥离 `求导/积分/极限 +(的)法则|公式|规则|表` 固定搭配,`计算这个求导公式`、`推导…的求导法则` 仍按计算请求拒绝。 - 基线报告把范围拒绝与失败分开统计。`ComputationOutsideScope` 不再计入 Errors 与分母,改为单列条目(当前为 golden `010`、`073`);两条已登记为有意的产品边界(`eval/known-gaps.md` G-016):`010` 要的是因式分解结果,`073` 无法用正则与真实求导请求区分。 - 复测 100 条 golden set(MiniCPM-V 4.7 1B,内置 llama-server):结构通过 59/95(62.1%)、编译 92/100、渲染 98/100、轨迹 98/100、确定性修复 53/53;延迟 p50/p95 = 1.5s/4.2s。分母由 94 变 95 是因为 `062` 重新进入评测,不代表通过率下降。 -- Treat nominalised `求导法则/公式/表` phrases as content to typeset rather than computation requests, and report scope-gate refusals separately from failures in the baseline. +### English + +- Treat phrases such as `求导法则` (differentiation rule), `基本积分公式表` (table of standard integrals) and equivalent limit phrases as notation requests. Explicit requests to calculate or derive a result still hit the computation gate. +- Report `ComputationOutsideScope` separately and exclude the two scope refusals (`010`, `073`) from the structural denominator. Restoring case `062` gives 59/95 structural passes, 92/100 compile, 98/100 render, 98/100 trace and 53/53 deterministic repairs. The tagged run recorded 1548/4218 ms p50/p95. See the [v0.4.8 report](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.8/eval/reports/2026-11-13-baseline-live.md). ## 0.4.7 - 2026-11-13 +### 中文 + - 修复拼错命令与非法环境名没有任何自动修复路径的问题。KaTeX 成为唯一语法权威后只产出 `katex_parse` 错误,而 fixer 只处理旧的三类错误;现在 `Undefined control sequence: \X` 走截断命令白名单(`\fr`→`\frac`、`\bet`→`\beta` 等,只在完整命令边界替换),`No such environment: X` 走唯一前缀映射(`case`→`cases`,多候选拒绝猜测)。确定性修复改为最多三轮且可组合(改环境名 + 补 `\end`),每轮重新校验,最终仍非法则保留原答案。 - 补齐“只转写不解答”的两段缺口。输入侧 `scope` 增补口语化求值说法(`等于几`、`是几`、`是多少`、`加到 N 的和`),这些真实用例此前会绕过门禁;输出侧新增求值守卫,检出“等号右边是用户从未写过的常数”时带 `transcribe_only` 约束重试一次,重试仍求值则保留首个答案并写 `fix_log=["evaluation-suspected"]`,不静默替换为另一个错答案。 - 新增真实语料评测集 `eval/real_world_set.yaml`(52 条)与判定器 `eval/real_world_lib.py`、在线基线运行器 `eval/run_real_world.py`、`tests/test_real_world_set.py`(225 条离线断言)。语料覆盖术语名、讲义/真题整句、易混清单、只转写不解答与半截 LaTeX 补全,逐条可溯源;`docs/test-prompts-real-world.md` 为同步的人读表格,`eval/known-gaps.md` 增补 G-009…G-014。 -- 基线(`minicpm-v4.7-1b`,llama-server,52 条):通过 31 条;`no-solving` 6/6(由输入侧门禁拒绝)、`completion` 10/11、`lecture-sentence` 6/8、`common-mistake` 4/7、`named-formula` 5/20。本轮不宣称模型质量结论,明细见 `eval/real-world-baseline.json`。 +- 基线(`minicpm-v4.7-1b`,llama-server,52 条):通过 31 条;`no-solving` 6/6(由输入侧门禁拒绝)、`completion` 10/11、`lecture-sentence` 6/8、`common-mistake` 4/7、`named-formula` 5/20。见 [v0.4.7 冻结结果](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.7/eval/real-world-baseline.json)。 - `ConversationMemory` 明确标注为“只记录、不喂给模型”,两处写入点加注释与回归测试指引,避免后续维护者误当上下文丢失的 bug 修回去。 -- Fix silent repair gaps for mistyped commands and illegal environment names, add an output-side evaluation guard, and ship a 52-entry real-world corpus with an offline judge and live baseline runner. +### English + +- Repair allowlisted truncated commands and unambiguous environment names. Combine fixes across up to three revalidated passes, preserving the original invalid result when repair fails. +- Catch more colloquial computation requests at the input gate. On `/api/convert`, retry once when an unsolicited numeric result is detected and record `evaluation-suspected` if the retry also performs evaluation. Conversation memory remains a record of turns and is excluded from model context. +- Add the 52-case real-world corpus, offline assertions and live runner. The first tagged baseline passed 31/52, including 5/20 named formulas and all six scope refusals. ## 0.4.6 - 2026-11-10 @@ -39,7 +66,8 @@ - Reject packaged ELF imports newer than glibc 2.35, including weak imports. Validate deb installation and extracted AppImage startup with external networking disabled and loopback retained. - Linux validation covers desktop/backend startup and deterministic KaTeX rendering; it does not test model inference, every Linux distribution, or AppImageHub acceptance. - Preserve published v0.4.5 assets. Version 0.4.6 requires freshly built installers, matching SHA-256 checksums and source-bound validation reports. -- 修复 Linux 构建基线和离线启动验证;保留 v0.4.5 原始发布文件,新版使用独立版本与校验记录。 +- 桌面、Python 后端和固定版本的 llama-server 统一在 Ubuntu 22.04 / glibc 2.35 基线上构建。打包检查拒绝更高版本的 glibc 符号引用,包括弱符号。deb 和解包后的 AppImage 在关闭外网、保留回环时验证启动与确定性渲染。 +- 保留 v0.4.5 原始发布文件,v0.4.6 使用新构建的安装包、对应的 SHA-256 校验值和源提交绑定的安装报告。 ## 0.4.5 - 2026-11-08 @@ -117,6 +145,20 @@ - Include [PR #45](https://github.com/CacinieP/TeXada-the-Math-Agent/pull/45): patched Python locks, rustls/webpki and offline KaTeX 0.18.2, plus the documented upstream two-line glib 0.18.5 source backport required by GTK3. Lock checks and the optimized Linux iterator regression gate releases; do not describe the backport as an official fixed crate or a dismissed alert. - This formal release is scoped to engineering acceptance. Historical failures and metrics remain unchanged; independent human transcription review, real scanned/handwritten material and Windows/Intel model quality remain follow-up work. Compilation, rendering, commits and installer startup do not establish mathematical accuracy. Linux compatibility follows final native Ubuntu package checks. +## 0.4.2-pre-inline-fix.1 - 2026-11-04 + +### 中文 + +- v0.4.2 正式版之前的打包检查点,包含 [#47](https://github.com/CacinieP/TeXada-the-Math-Agent/pull/47) 的 V8 生命周期和 Linux 安装依赖修复,以及 [#48](https://github.com/CacinieP/TeXada-the-Math-Agent/pull/48) 的安装检查修复。macOS DMG 校验保留挂载输出,Windows 冒烟测试用文件身份核对内置运行时,兼容扩展长度路径前缀。 +- 后续 [#49](https://github.com/CacinieP/TeXada-the-Math-Agent/pull/49) 的行内补全按键隔离与过期提示修复进入 v0.4.2 正式标签。 + +### English + +- Packaging checkpoint before v0.4.2. Includes the V8 lifecycle and Linux runtime dependency fixes from #47 and the installer smoke fixes from #48. Keep DMG mount output visible and compare the Windows bundled runtime by file identity so extended-length paths work. +- The inline-completion key handling and stale-hint fixes from #49 landed in the final v0.4.2 tag. + +[Tag / 标签](https://github.com/CacinieP/TeXada-the-Math-Agent/tree/v0.4.2-pre-inline-fix.1) · [Compare with v0.4.2 / 与正式版对比](https://github.com/CacinieP/TeXada-the-Math-Agent/compare/v0.4.2-pre-inline-fix.1...v0.4.2) + ## 0.4.2-rc.3 - 2026-10-20 ### 中文 @@ -858,6 +900,8 @@ OCR 继续使用独立配置的视觉模型;不安装 OCR 模型也可使用 ## 0.2.5 - 2026-07-09 +Development snapshot recorded in [commit `2a756b4`](https://github.com/CacinieP/TeXada-the-Math-Agent/commit/2a756b4), with no `v0.2.5` tag. + - Fixed the OCR tab so a second recognition can start immediately by dragging, pasting, or selecting another image after a completed run. - Prevented duplicate OCR drag/drop handlers from being registered when reopening the OCR tab. - Changed the desktop shell to open visible by default instead of starting hidden. @@ -866,6 +910,8 @@ OCR 继续使用独立配置的视觉模型;不安装 OCR 模型也可使用 ## 0.2.5 - 2026-07-09 中文 +此版本见[源码提交 `2a756b4`](https://github.com/CacinieP/TeXada-the-Math-Agent/commit/2a756b4),仓库没有 `v0.2.5` 标签。 + - 修复 OCR 页:第一次识别完成后,可以直接拖图、粘贴或选择另一张图片继续第二次识别。 - 避免反复进入 OCR 页时重复注册拖放事件,导致一次拖图触发多次请求。 - 桌面壳打开时默认显示窗口,不再默认隐藏。 diff --git a/README.md b/README.md index 73eb7cd..feb98e3 100644 --- a/README.md +++ b/README.md @@ -32,12 +32,11 @@ ## TeXada -**v0.4.4** adds seven parameterized matrix presets, revision-bound matrix -checks, source-faithful finite sum/integral descriptions and macOS insertion -focus handoff diagnostics. Presets support type search, history restoration -and backup replay. Final package checks and pending native insertion acceptance -are recorded in the release documentation. -[Release scope and installer checks](docs/release-v0.4.4.md) · [Changelog](CHANGELOG.md#044---2026-11-06) +**v0.4.10** follows the system language on first launch. Chinese locales start +in Chinese, other locales start in English, and a saved choice takes precedence. +It includes the 24-entry formula glossary, seven parameterized matrix presets and +the Ubuntu 22.04 / glibc 2.35 Linux build baseline. +[Download v0.4.10](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.10) · [Changelog](CHANGELOG.md#0410---2026-11-14) TeXada helps transcribe formulas into LaTeX. Describe notation in natural language, paste a clear typeset formula screenshot, or give it an @@ -90,13 +89,13 @@ do not establish transcription accuracy. Compare the source and candidate; The built-in and named MiniCPM-V 4.7 1B profiles default to candidate generation before tool review. Other checkpoints keep direct planning; explicit experimental -settings remain respected. See [release scope and validation](docs/release-v0.4.4.md). +settings remain respected. See [tool routing and validation](docs/architecture.md). ### Highlights | Capability | Description | |------------|-------------| -| Natural language to LaTeX | Describe a formula and get copyable LaTeX or Markdown | +| Natural language to LaTeX | Describe a formula and get copyable LaTeX or Markdown. Exact matches in the formula glossary use a deterministic template. | | Formula OCR | The vision role proposes a formula; the shared Planner reviews it through TeX tools | | Completion | Rules or the configured model propose a completion; the shared Agent validates it | | Validation and repair | Check, fix, render and highlight LaTeX | @@ -170,30 +169,29 @@ The Ollama port is configurable. The default is `http://localhost:11434`, but Se ### Download -Open the [current public release](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/latest) -and choose the asset for your machine: +Open the [latest public release](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/latest) +or choose a v0.4.10 installer below. | Platform | Installer | |----------|-----------| -| macOS Apple Silicon | `TeXada__aarch64.dmg` | -| macOS Intel | `TeXada__x64.dmg` | -| Windows x64 | `TeXada__x64-setup.exe` | -| Linux x64, Debian/Ubuntu | `TeXada__amd64.deb` | -| Linux x64, AppImage | `TeXada__amd64.AppImage` | +| macOS Apple Silicon | [TeXada_0.4.10_aarch64.dmg](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_aarch64.dmg) | +| macOS Intel | [TeXada_0.4.10_x64.dmg](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_x64.dmg) | +| Windows x64 | [TeXada_0.4.10_x64-setup.exe](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_x64-setup.exe) | +| Linux x64, Debian/Ubuntu | [TeXada_0.4.10_amd64.deb](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_amd64.deb) | +| Linux x64, AppImage | [TeXada_0.4.10_amd64.AppImage](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_amd64.AppImage) | macOS requires 13.3 or later. Model weights download separately in Settings. -Windows installers are not code-signed. Linux compatibility is scoped to the -final Ubuntu native build and installed-package checks; it is not a claim of -support for every distribution. Check the release assets and attached validation -records for the final package results; older beta/RC checks are historical. See the -[Linux installation guide](docs/linux-installation.md) for verified environments, locale behavior and the local model. +Both macOS DMGs are signed and notarized. The Windows installer is unsigned. +The release manifest identifies the five installers, their SHA-256 checksums and +per-package installation reports. Ubuntu 22.04 release checks cover offline startup +and the English first window for both Linux installers. The public x64 AppImage +also passed the AppImageHub test. See the [release manifest](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/release-manifest.json), [Linux installation guide](docs/linux-installation.md) and [catalog test](https://github.com/AppImage/appimage.github.io/pull/11396#issuecomment-6875598626). Generated formulas still need comparison with the source. Historical beta runs include mathematically incorrect committed results and have no independent -human adjudication; this release does not convert those records into accuracy -claims. See the [release limitations](docs/release-v0.4.4.md#known-limits). +human adjudication. See the [evaluation limitations](eval/known-gaps.md). -v0.4.4 uses `AGPL-3.0-or-later`. Immutable older tags retain their own licenses: +TeXada v0.4.10 uses `AGPL-3.0-or-later`. Older tags retain their own licenses. `v0.3.6` and earlier artifacts use `GPL-3.0-or-later`; `v0.3.7` and later use `AGPL-3.0-or-later`. Historical installers remain on the [release list](https://github.com/CacinieP/TeXada-the-Math-Agent/releases). @@ -201,8 +199,8 @@ v0.4.4 uses `AGPL-3.0-or-later`. Immutable older tags retain their own licenses: Closing the main window hides TeXada; restore it from the tray or the existing shortcut. To stop the app, choose **Quit** in its menu. If a model download is interrupted, retry in Settings: received data can resume, and completion appears -only after the whole batch succeeds. Native close/restore/Quit acceptance for -this patch remains tracked in [#50](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/50). +only after the whole batch succeeds. The v0.4.3 close/restore/Quit and cold +download acceptance record is in [#50](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/50). ### From A Fresh Machine To First Formula @@ -259,7 +257,7 @@ Save after changing the backend or its runtime paths before using service contro ### Quick Start With Built-in llama-server -1. Install and open **v0.4.4** from the release link above. +1. Install and open TeXada from the release link above. 2. Open **Settings → Backend**, select **`Built-in llama-server`**, and click **Save**. Keep the default address `http://127.0.0.1:8080`. Leave **llama-server executable** and **Model directory** blank to use the packaged runtime and `~/.texada/models`. 3. In **Local inference service (llama-server)**, click **Download model**. Wait until downloading finishes for both the model weights and the vision projector. They are not included in the installer. @@ -588,11 +586,10 @@ recorded outputs with `--mode recorded`. Details and measured boundaries live in [eval/known-gaps.md](eval/known-gaps.md) and [eval/reports/2026-09-19-baseline-live.md](eval/reports/2026-09-19-baseline-live.md). -#### Re-measured on 2026-11-13 (v0.4.7) +#### Golden set reruns on 2026-11-13 (v0.4.8 and v0.4.9) -Same runner, current default model **MiniCPM-V 4.7 (1B)** via the bundled -llama-server on Apple silicon. Different model and pipeline from the 2026-09-19 -row above, so the two are **not** a like-for-like improvement claim: +Both tagged runs used **MiniCPM-V 4.7 (1B)** via the bundled llama-server on +Apple silicon. The September baseline used MiniCPM5-2B and a different pipeline. | Golden set (100 entries) | Result | |--------|--------| @@ -601,10 +598,11 @@ row above, so the two are **not** a like-for-like improvement claim: | Render layer | 98/100 | | Trace layer | 98/100 | | Deterministic repair pass rate | 53/53 (100%) | -| Latency p50 / p95 | 1.5s / 4.2s | +| Latency p50 / p95, v0.4.8 | 1548 ms / 4218 ms | +| Latency p50 / p95, v0.4.9 | 1255 ms / 3325 ms | -Seven of the 95 anchor-checkable entries are answered by deterministic rules -(glossary / preset / transcription) with zero tokens; the rest are model runs. +The v0.4.9 report records seven zero-token deterministic results across the +golden set. See the [v0.4.8 report](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.8/eval/reports/2026-11-13-baseline-live.md) and [v0.4.9 report](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/reports/2026-11-13-baseline-live.md). Two golden entries (`010`, `073`) are refused by the computation-scope gate instead of transcribed; they are listed separately in the baseline report and @@ -615,9 +613,10 @@ two are documented as a deliberate boundary in *result*, and `073` cannot be separated from a real differentiation request by pattern alone. -The new real-world corpus (`eval/real_world_set.yaml`, 52 entries from real -study notes, exam explanations and run logs) measures a different axis — -inputs the 100-entry set does not contain: +#### Real-world corpus (v0.4.7 and v0.4.9) + +The 52-case corpus draws on study notes, exam explanations and run logs. +v0.4.7 passed 31/52. The table shows the v0.4.9 rerun. | Real-world corpus | Pass | |--------|------| @@ -628,22 +627,17 @@ inputs the 100-entry set does not contain: | `no-solving` (must not be evaluated) | 6/6 (refused by the scope gate) | | `completion` (half-finished / mistyped LaTeX) | 10/11 | -`named-formula` no longer measures the model. A bare theorem name is answered by -the curated glossary (`src/texada/core/glossary.py`, 24 entries) under the -existing `nl_canonical_concept` rule, in ~12 ms with zero tokens; 21 of the 52 -entries are served that way. Before the glossary the same 1B model scored 5/20 -on this slice — that gap is knowledge recall, not transcription. The remaining -failures are genuine model results. +The glossary raised `named-formula` from 5/20 to 20/20 by serving exact matches +on bare theorem names. The v0.4.9 run recorded 21 `template` results across all +52 cases, with a median of 12.3 ms. Its five failures were reported on the model +or agent paths, including one completion case. Reproduce with `python eval/run_real_world.py --base-url ... --output ...`; -per-case verdicts and reasons are in [eval/real-world-baseline.json](eval/real-world-baseline.json), -and the raw run is in [eval/reports/2026-11-13-baseline-live.md](eval/reports/2026-11-13-baseline-live.md). -Named theorems are the weakest slice: the 1B model often answers a different -theorem or just echoes `\text{...}`. +per-case verdicts are available for [v0.4.7](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.7/eval/real-world-baseline.json) and [v0.4.9](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/real-world-baseline.json). ### Quality Gate -The current `main` branch has no known blocking issue after the latest audit pass. That means tests and static checks pass; it is not a mathematical promise that no software bug can exist. Maintainer validation details live in [Contributing](.github/CONTRIBUTING.md), [Source audit](docs/audit.md), and GitHub Actions. +The [Audit workflow](https://github.com/CacinieP/TeXada-the-Math-Agent/actions/workflows/audit.yml) tracks test and static-check results by commit. See [Contributing](.github/CONTRIBUTING.md) and [Source audit](docs/audit.md) for the checks and their scope. ### Documentation @@ -668,13 +662,13 @@ releases retain the license recorded in their immutable tag. See [LICENSE](LICEN ## TeXada -**v0.4.4 正式版** 提供七种参数化矩阵预设、修订绑定结构检查、有限求和与区域积分保真,以及 macOS 插入焦点交接和阶段诊断。预设支持类型搜索、历史恢复和备份重放。最终安装包检查及待完成的原生插入验收记录见发布说明。[正式版说明](docs/release-v0.4.4.md) · [更新日志](CHANGELOG.md#044---2026-11-06) +**v0.4.10 正式版** 首次启动时跟随系统语言,中文环境默认中文,其他环境默认英文,已保存的语言选择优先。当前版本包含 24 条公式术语词表、七种参数化矩阵预设和 Ubuntu 22.04 / glibc 2.35 Linux 构建基线。 [下载 v0.4.10](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.10) · [更新日志](CHANGELOG.md#0410---2026-11-14) TeXada 是一个用于数学公式识别、转写与排版的本地工具。你可以输入自然语言、 粘贴清晰的排版公式截图,也可以交给它一段不完整的 TeX;TeXada 会把输入转换为 结构化公式,用确定性工具检查,再渲染为可核对、可复制的结果。手写与真实扫描页面尚未完成独立验收,转写仍需逐项对照原件。 -**支持范围:**只识别、转写、编辑和排版公式;不提供求值、化简、因式分解、解方程、求导、积分计算或证明。导数与积分的记号仍支持输入。`7+5` 保留为表达式,不补成 `12`。自然语言和 OCR 候选现在直接提交,请对照原文与候选结果;编译通过不代表转写准确。详见 [正式版范围与验证](docs/release-v0.4.4.md)。 +**支持范围:**识别、转写、编辑和排版公式。求值、化简、因式分解、解方程、求导、积分计算和证明超出支持范围。导数与积分的记号仍支持输入。`7+5` 保留为表达式。自然语言和 OCR 候选会直接提交,请对照原文检查转写结果。详见 [工具路由和校验](docs/architecture.md)。 它不是聊天界面套壳,而是基于端侧 Agent 的结构化数学编辑器。产品明确只有两个 模型保留规划/文本与视觉/OCR 两个职责。MiniCPM-V 4.7 profile 用同一个 @@ -704,7 +698,7 @@ Level 1,但只在明确存在 before/after 公式时启用,不会把自然 | 能力 | 说明 | |------|------| -| 自然语言转 LaTeX | 输入公式描述,生成可复制的 LaTeX 或 Markdown | +| 自然语言转 LaTeX | 输入公式描述,生成可复制的 LaTeX 或 Markdown。精确命中术语词表时使用确定性模板。 | | 公式 OCR | 视觉职责产出候选,共享 Planner 再通过 TeX 工具审查 | | 公式补全 | 规则或配置的模型产出候选,再由统一 Agent 校验 | | 校验与修复 | 检查、修复、渲染并高亮 LaTeX | @@ -769,23 +763,23 @@ Ollama 端口不是写死的。默认地址是 `http://localhost:11434`,但可 ### 下载 -前往[当前正式发布页](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/latest),选择适合自己机器的安装包: +前往[最新正式发布页](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/latest),或从下表下载适合自己机器的 v0.4.10 安装包。 | 平台 | 安装包 | |------|--------| -| macOS Apple Silicon | `TeXada__aarch64.dmg` | -| macOS Intel | `TeXada__x64.dmg` | -| Windows x64 | `TeXada__x64-setup.exe` | -| Linux x64,Debian/Ubuntu | `TeXada__amd64.deb` | -| Linux x64,AppImage | `TeXada__amd64.AppImage` | +| macOS Apple Silicon | [TeXada_0.4.10_aarch64.dmg](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_aarch64.dmg) | +| macOS Intel | [TeXada_0.4.10_x64.dmg](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_x64.dmg) | +| Windows x64 | [TeXada_0.4.10_x64-setup.exe](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_x64-setup.exe) | +| Linux x64,Debian/Ubuntu | [TeXada_0.4.10_amd64.deb](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_amd64.deb) | +| Linux x64,AppImage | [TeXada_0.4.10_amd64.AppImage](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/TeXada_0.4.10_amd64.AppImage) | -macOS 需要 13.3 或更新版本;模型权重安装后单独下载。Windows 安装包未做代码签名。Linux 兼容范围以最终 Ubuntu 原生构建和安装包检查记录为准,不代表所有发行版已经通过验收。最终产物及检查结果以 Release 资产和附带记录为准,旧 Beta/RC 的实测仅属于历史版本。验证环境、首屏语言和本地模型说明见 [Linux 安装指南](docs/linux-installation.md)。 +macOS 需要 13.3 或更新版本,两个 DMG 均已签名和公证。Windows 安装包未签名。模型权重安装后单独下载。发布清单列出五个安装包、对应的 SHA-256 校验值和安装验证报告。两个 Linux 安装包均在 Ubuntu 22.04 上通过离线启动与英文首屏检查,正式发布的 x64 AppImage 也通过了 AppImageHub 复测。 [发布清单](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/download/v0.4.10/release-manifest.json) · [Linux 安装指南](docs/linux-installation.md) · [目录复测](https://github.com/AppImage/appimage.github.io/pull/11396#issuecomment-6875598626) -生成结果仍需对照原文。历史 Beta 记录中存在数学错误公式成功提交,且没有独立人工复核;本次正式发布不把这些记录改写为准确率结论。详见[已知限制](docs/release-v0.4.4.md#known-limits)。 +请对照原文检查生成结果。历史 Beta 记录中存在数学错误公式成功提交的情况,且缺少独立人工复核。详见 [评测限制](eval/known-gaps.md)。 -v0.4.4 使用 `AGPL-3.0-or-later`。旧标签及安装包继续适用其标签内许可证:`v0.3.6` 及更早产物为 `GPL-3.0-or-later`,`v0.3.7` 起为 `AGPL-3.0-or-later`。历史安装包仍保留在 [Release 列表](https://github.com/CacinieP/TeXada-the-Math-Agent/releases)。 +v0.4.10 使用 `AGPL-3.0-or-later`。`v0.3.6` 及更早产物为 `GPL-3.0-or-later`,`v0.3.7` 起为 `AGPL-3.0-or-later`。历史安装包见 [Release 列表](https://github.com/CacinieP/TeXada-the-Math-Agent/releases)。 -关闭主窗口会隐藏 TeXada,可通过托盘或现有快捷键恢复;需要停止应用时,请在菜单中选择 **Quit / 退出**。模型下载中断后可在设置中重试,已接收的数据支持续传,整批文件成功后才显示完成。本修版的原生关闭、恢复及退出复验继续在 [#50](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/50) 跟踪。 +关闭主窗口会隐藏 TeXada,可通过托盘或现有快捷键恢复。需要停止应用时,请在菜单中选择 **Quit / 退出**。模型下载中断后可在设置中重试,已接收的数据支持续传,整批文件成功后才显示完成。v0.4.3 的原生关闭、恢复、退出及冷下载验收见 [#50](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/50)。 ### 从裸机到第一次使用 @@ -841,7 +835,7 @@ TeXada 的 release 安装包面向普通用户。你不需要安装 Python、Nod ### 内置 llama-server 快速启动 -1. 从上方 Release 链接安装并打开 **v0.4.4**。 +1. 从上方 Release 链接安装并打开 TeXada。 2. 打开**设置 → 后端连接**,选择 **`Built-in llama-server`**,点击**保存**。 地址保留 `http://127.0.0.1:8080`;**llama-server 路径**和**模型目录**留空,使用应用内置运行时及 `~/.texada/models`。 3. 在**本地推理服务(llama-server)**中点击**下载模型**,等待模型权重和视觉投影文件全部下载完成。安装包不包含这些文件。 @@ -1155,10 +1149,9 @@ Linux 尚未实现系统光标自动键入;点击公式后需手动粘贴。Wa [eval/known-gaps.md](eval/known-gaps.md) 与 [eval/reports/2026-09-19-baseline-live.md](eval/reports/2026-09-19-baseline-live.md)。 -#### 2026-11-13 复测(v0.4.7) +#### 2026-11-13 黄金集复测(v0.4.8 与 v0.4.9) -同一套 runner,当前默认模型 **MiniCPM-V 4.7(1B)**,跑在 Apple 芯片上的内置 -llama-server。模型与管线都和上面 2026-09-19 那次不同,**不能当作同口径提升**: +两次标签对应的实测均使用 Apple 芯片上的内置 llama-server 和 **MiniCPM-V 4.7(1B)**。9 月的基线使用 MiniCPM5-2B,运行管线也不同。 | 100 条 goldenset | 结果 | |------|------| @@ -1167,9 +1160,10 @@ llama-server。模型与管线都和上面 2026-09-19 那次不同,**不能当 | 渲染层 | 98/100 | | 轨迹层 | 98/100 | | 确定性修复通过率 | 53/53(100%) | -| 延迟 p50 / p95 | 1.5s / 4.2s | +| v0.4.8 延迟 p50 / p95 | 1548 ms / 4218 ms | +| v0.4.9 延迟 p50 / p95 | 1255 ms / 3325 ms | -95 条可判条目里有 7 条由确定性规则(词表/预设/转写)零 token 直出,其余是模型跑出来的。 +v0.4.9 的黄金集记录了 7 条零 token 的确定性结果。 [v0.4.8 报告](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.8/eval/reports/2026-11-13-baseline-live.md) · [v0.4.9 报告](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/reports/2026-11-13-baseline-live.md) 其中 `010`、`073` 两条被计算范围门禁拒绝而不是转写,基线报告里单列、不计入 分母。`062`(`u 除以 v 的求导法则`)此前也被误拒,v0.4.8 已修——`求导法则/ @@ -1177,8 +1171,9 @@ llama-server。模型与管线都和上面 2026-09-19 那次不同,**不能当 [eval/known-gaps.md](eval/known-gaps.md) G-016:`010` 要的是因式分解**结果**, `073` 无法用正则与真实的求导请求区分。 -新增的真实语料(`eval/real_world_set.yaml`,52 条,取自本机讲义、真题详解与 -运行日志)测的是 100 条集里没有的那一档输入: +#### 真实语料复测(v0.4.7 与 v0.4.9) + +52 条语料取自讲义、真题详解和运行日志。v0.4.7 通过 31/52,下表为 v0.4.9 的结果。 | 真实语料 | 通过 | |------|------| @@ -1189,20 +1184,14 @@ llama-server。模型与管线都和上面 2026-09-19 那次不同,**不能当 | `no-solving`(只转写不解答) | 6/6(由范围门禁拒绝) | | `completion`(半截/拼错 LaTeX) | 10/11 | -`named-formula` 这一档**不再测模型**:只给术语名时由受控词表 -(`src/texada/core/glossary.py`,24 条)沿既有 `nl_canonical_concept` 规则直出, -约 12 ms、零 token;52 条里有 21 条走这条路。加词表前同一个 1B 模型在这档是 -5/20——差的是"记不记得住定理",不是"会不会转写"。其余失败项仍是模型真实结果。 +`named-formula` 由 5/20 升至 20/20,提升来自只给术语名时的词表精确匹配。v0.4.9 的 52 条结果中有 21 条标为 `template`,这一组的延迟中位数为 12.3 ms。剩余 5 条失败发生在 `model` 或 `agent` 路径,包括 1 条补全。 -复现:`python eval/run_real_world.py --base-url … --output …`;逐条判定与理由见 -[eval/real-world-baseline.json](eval/real-world-baseline.json),原始报告见 -[eval/reports/2026-11-13-baseline-live.md](eval/reports/2026-11-13-baseline-live.md)。 -术语名是最弱的一档:1B 模型经常答成另一个定理,或只吐一句 `\text{...}`。 +复现命令为 `python eval/run_real_world.py --base-url … --output …`。逐条判定见 [v0.4.7 结果](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.7/eval/real-world-baseline.json)与 [v0.4.9 结果](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/real-world-baseline.json)。 ### 质量门禁 -当前 `main` 分支在最新审计后没有已知阻塞问题。这里的含义是测试和静态检查通过,不是声称软件绝对不可能有 bug。维护者验证细节见 [贡献指南](.github/CONTRIBUTING.md)、[源码审计](docs/audit.md) 和 GitHub Actions。 +[Audit 工作流](https://github.com/CacinieP/TeXada-the-Math-Agent/actions/workflows/audit.yml)按提交记录测试和静态检查结果。检查项及覆盖范围见 [贡献指南](.github/CONTRIBUTING.md)和[源码审计](docs/audit.md)。 ### 文档 @@ -1218,12 +1207,6 @@ llama-server。模型与管线都和上面 2026-09-19 那次不同,**不能当 -## v0.4.4 release scope - -The built-in backend uses one MiniCPM-V-4.7-1B Q4_K_M checkpoint for text and OCR. Download the model in Settings and start llama-server. Existing Ollama and OpenAI-compatible settings remain selectable. Model files are not included in installers. - -Generated formulas require mathematical review: syntax validation and successful submission do not prove correctness. See [the v0.4.4 release notes](docs/release-v0.4.4.md) for current behavior and package verification; [beta validation](docs/beta-release-validation.md) remains historical evidence. - ## Author and citation · 作者与引用 Created and maintained by [CacinieP](https://github.com/CacinieP). 由 CacinieP 创建与维护。 diff --git a/docs/README.md b/docs/README.md index c13d588..40051d7 100644 --- a/docs/README.md +++ b/docs/README.md @@ -54,5 +54,7 @@ This directory keeps technical and maintenance documentation. Community health f | [Security](../.github/SECURITY.md) | Supported versions and private vulnerability reporting. | | [Support](../.github/SUPPORT.md) | Bug-report checklist and model/Ollama troubleshooting entry point. | | [Code of Conduct](../.github/CODE_OF_CONDUCT.md) | Community behavior expectations. | -| [v0.4.4 release](release-v0.4.4.md) | Matrix and transcription changes, package identity, installation gates and pending native acceptance. | +| [v0.4.10 release](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.10) · [Changelog entry](../CHANGELOG.md#0410---2026-11-14) | Installers, first-run locale selection, and Linux release and catalog evidence. | +| [v0.4.9](release-v0.4.9.md), [v0.4.8](release-v0.4.8.md), [v0.4.7](release-v0.4.7.md) | Glossary, computation-scope fixes, deterministic repair and tagged evaluation baselines. | +| [v0.4.4](release-v0.4.4.md), [v0.4.3](release-v0.4.3.md), [v0.4.2](release-v0.4.2.md) | Matrix and transcription changes, plus historical release and installer notes. | | [Changelog](../CHANGELOG.md) | Release history. | diff --git a/docs/linux-installation.md b/docs/linux-installation.md index b1a5a18..ff0dc3d 100644 --- a/docs/linux-installation.md +++ b/docs/linux-installation.md @@ -19,7 +19,9 @@ chmod +x ./TeXada__amd64.AppImage 安装包内含 `llama-server`。使用内置后端时,可在设置里下载 MiniCPM-V 4.7 1B Q4_K_M 权重和 F16 视觉投影,下载后可在本地执行公式转写与图片 OCR。预留约 2 GiB 磁盘空间存放模型文件。 -Ubuntu 22.04 CI 在 glibc 2.35 基线上构建并检查两个包,用 Xvfb 在关闭外部网络但保留本机回环的环境中验证启动、英文窗口和确定性公式渲染。独立 Ubuntu 24.04.5 x86_64 机器实测使用 Intel Core i5-7200U(2 核 4 线程),系统识别内存约 11 GiB。该机在 Xvfb 验证了英文首屏,并通过打包后的本地 API 验证了 CPU 文本转写和图片 OCR。[Ubuntu 24.04 实测与截图](validation/2026-11-14-ubuntu24-appimage/README.md) +v0.4.10 的 Ubuntu 22.04 发布 CI 在 glibc 2.35 基线上构建两个 Linux 安装包,并使用 Xvfb 在关闭外网、保留本机回环时验证启动、英文首屏和确定性公式渲染。正式发布的 x64 AppImage 随后通过 AppImageHub 测试。 [发布工作流](https://github.com/CacinieP/TeXada-the-Math-Agent/actions/runs/43581742151) · [AppImageHub 复测](https://github.com/AppImage/appimage.github.io/pull/11396#issuecomment-6875598626) + +独立 Ubuntu 24.04.5 x86_64 实测使用语言修复分支的 v0.4.9 CI 候选 AppImage,硬件为 Intel Core i5-7200U(2 核 4 线程),系统识别内存约 11 GiB。该机通过 Xvfb 验证英文首屏,通过打包后的本地 API 验证 CPU 文本转写和图片 OCR。 [Ubuntu 24.04 实测与截图](validation/2026-11-14-ubuntu24-appimage/README.md) Linux 点击公式后可手动粘贴。图形桌面的双击启动、Wayland 全局快捷键、托盘和手动粘贴属于独立验收项。 @@ -31,6 +33,8 @@ Starting with v0.4.10, a fresh profile follows the system locale. Chinese locale The installer includes `llama-server`. In Settings, download the MiniCPM-V 4.7 1B Q4_K_M weights and F16 visual projector for local formula transcription and image OCR. Allow about 2 GiB of disk space for the model files. -CI builds on Ubuntu 22.04 with a glibc 2.35 baseline and checks both packages under Xvfb with external networking disabled and loopback available. A separate Ubuntu 24.04.5 x86_64 machine with a 2-core, 4-thread Intel Core i5-7200U and about 11 GiB of visible RAM verified the English window under Xvfb and CPU text and image inference through the packaged local API. [Hardware test and screenshot](validation/2026-11-14-ubuntu24-appimage/README.md) +The v0.4.10 release workflow built both Linux installers on Ubuntu 22.04 with a glibc 2.35 baseline. Under Xvfb, it checked startup, the English first window and deterministic rendering with external networking disabled and loopback available. AppImageHub passed the public x64 AppImage. [Release workflow](https://github.com/CacinieP/TeXada-the-Math-Agent/actions/runs/43581742151) · [AppImageHub test](https://github.com/AppImage/appimage.github.io/pull/11396#issuecomment-6875598626) + +A separate Ubuntu 24.04.5 x86_64 machine ran the v0.4.9 CI candidate from the locale-fix branch. The tester used a 2-core, 4-thread Intel Core i5-7200U and about 11 GiB of visible RAM to verify the English window under Xvfb and CPU text and image inference through the packaged local API. [Hardware test and screenshot](validation/2026-11-14-ubuntu24-appimage/README.md) Clicking a formula copies it for manual paste on Linux. Interactive desktop launch, Wayland shortcuts, tray integration and manual clipboard use have separate acceptance requirements. From 2ff241aa61d5a9adf08a7b7679915bafdbacd4df Mon Sep 17 00:00:00 2001 From: CacinieP <129482851+CacinieP@users.noreply.github.com> Date: Sun, 15 Nov 2026 05:17:33 +0800 Subject: [PATCH 2/2] docs: clarify TeXada release and validation evidence --- CHANGELOG.md | 27 ++++++++++++++++++++++----- README.md | 18 ++++++++---------- docs/README.md | 2 +- docs/release-v0.4.9.md | 4 ++-- 4 files changed, 33 insertions(+), 18 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 0fd78b4..6e73435 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -2,6 +2,8 @@ This history includes stable releases, prerelease tags and explicitly marked development snapshots. / 本记录覆盖正式版、预发布标签和明确标注的开发快照。 +Dates in headings preserve the historical changelog dates. Where the source milestone and GitHub publication differ, both dates are labeled below in Asia/Shanghai time. / 标题保留历史日志日期。源码里程碑与 GitHub 公开日期不同时,下文按北京时间分别标注。 + ## Unreleased ## 0.4.10 - 2026-11-14 @@ -26,24 +28,24 @@ This history includes stable releases, prerelease tags and explicitly marked dev - 新增术语名词表 `src/texada/core/glossary.py`(24 条,含格林/高斯/斯托克斯/贝叶斯/全概率/华里士/施密特/拉格朗日/罗尔/柯西/泰勒/洛必达/一阶线性 ODE/二阶常系数/中心极限/样本方差/欧拉/夹逼/路径无关/二项分布/分部积分/N-L/柯西不等式/秩不等式),每条带名称别名、LaTeX、出处与备注。 - 词表沿既有确定性快路径接入:`DeterministicCandidateEngine._canonical_concept` 兜底查表(规则名仍为 `nl_canonical_concept`),`/api/convert` 在同一处先查表,命中返回 `Source.TEMPLATE`、`tokens_used=0`。匹配要求规范化后与别名完全相等,带参数的句子与复合问句仍交给模型;每条 LaTeX 必须通过 KaTeX 校验,失败则回退模型路径。 -- 实测(MiniCPM-V 4.7 1B):真实语料从 v0.4.7 的 31/52 升至 **47/52**,其中 `named-formula` 从 5/20 升至 **20/20**;52 条中 21 条由模板直出,模板路径 p50 为 12.3 ms。100 条 golden set 的结构通过率仍为 59/95,整套运行的 p50/p95 为 1255/3325 ms。见[真实语料逐条结果](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/real-world-baseline.json)和[黄金集报告](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/reports/2026-11-13-baseline-live.md)。 +- 实测(MiniCPM-V 4.7 1B):真实语料从 v0.4.7 的 31/52 升至 **47/52**,其中 `named-formula` 从 5/20 升至 **20/20**;52 条中 21 条由模板直出,模板路径 p50 为 12.3 ms。100 条黄金集的结构层仍通过 59/95,有延迟记录的样本 p50/p95 为 1255/3325 ms。见[真实语料逐条结果](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/real-world-baseline.json)和[黄金集报告](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/reports/2026-11-13-baseline-live.md)。 - 评测 runner 现在区分「确定性直出」与「模型结果」并单列条目,避免把查表得分读成模型能力;README 中英双段与 `docs/architecture.md`、`eval/known-gaps.md`(G-017)同步。 ### English - Add a 24-entry formula glossary with aliases, LaTeX and provenance. A normalized exact match on a bare name uses the existing deterministic candidate path, including `/api/convert`. A failed KaTeX check falls back to the model. Parameterized and compound requests continue through the model path. -- On the 52-case real-world corpus, the MiniCPM-V 4.7 1B configuration went from 31/52 in v0.4.7 to 47/52. Named formulas went from 5/20 to 20/20 via the glossary. The evaluator records template, model and agent paths separately. The 100-case golden set retained 59/95 structural passes. +- On the 52-case real-world corpus, the MiniCPM-V 4.7 1B configuration went from 31/52 in v0.4.7 to 47/52. Named formulas went from 5/20 to 20/20 via the glossary. Templates served 21 of the 52 cases with a median latency of 12.3 ms. The evaluator records template, model and agent paths separately. The 100-case golden set retained 59/95 structural passes. Among samples with a recorded latency, p50/p95 were 1255/3325 ms. ## 0.4.8 - 2026-11-13 ### 中文 - 修复计算范围门禁的误拒:`u 除以 v 的求导法则`、`泰勒公式的求导法则`、`基本积分公式表` 这类"要排版的内容"此前因动词 `求导` 命中而被整句拒绝。现在动词扫描前先剥离 `求导/积分/极限 +(的)法则|公式|规则|表` 固定搭配,`计算这个求导公式`、`推导…的求导法则` 仍按计算请求拒绝。 -- 基线报告把范围拒绝与失败分开统计。`ComputationOutsideScope` 不再计入 Errors 与分母,改为单列条目(当前为 golden `010`、`073`);两条已登记为有意的产品边界(`eval/known-gaps.md` G-016):`010` 要的是因式分解结果,`073` 无法用正则与真实求导请求区分。 -- 复测 100 条 golden set(MiniCPM-V 4.7 1B,内置 llama-server):结构通过 59/95(62.1%)、编译 92/100、渲染 98/100、轨迹 98/100、确定性修复 53/53;延迟 p50/p95 = 1.5s/4.2s。分母由 94 变 95 是因为 `062` 重新进入评测,不代表通过率下降。 +- 基线报告把范围拒绝与失败分开统计。`ComputationOutsideScope`(黄金集 `010`、`073`)单列,排除在 Errors 和结构层分母之外。编译、渲染和轨迹层仍按 100 条统计。两条已登记为有意的产品边界(`eval/known-gaps.md` G-016):`010` 要的是因式分解结果,`073` 无法用正则与真实求导请求区分。 +- 复测 100 条 golden set(MiniCPM-V 4.7 1B,内置 llama-server):结构通过 59/95(62.1%)、编译 92/100、渲染 98/100、轨迹 98/100、确定性修复 53/53;有延迟记录的样本 p50/p95 为 1548/4218 ms。`062` 恢复计入结构层,分母由 94 增至 95,通过数仍为 59。 ### English - Treat phrases such as `求导法则` (differentiation rule), `基本积分公式表` (table of standard integrals) and equivalent limit phrases as notation requests. Explicit requests to calculate or derive a result still hit the computation gate. -- Report `ComputationOutsideScope` separately and exclude the two scope refusals (`010`, `073`) from the structural denominator. Restoring case `062` gives 59/95 structural passes, 92/100 compile, 98/100 render, 98/100 trace and 53/53 deterministic repairs. The tagged run recorded 1548/4218 ms p50/p95. See the [v0.4.8 report](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.8/eval/reports/2026-11-13-baseline-live.md). +- Report `ComputationOutsideScope` separately and exclude the two scope refusals (`010`, `073`) from the structural denominator. Restoring case `062` increases the structural denominator from 94 to 95, with 59 passes in both runs. The run recorded 92/100 compile, 98/100 render, 98/100 trace and 53/53 deterministic repairs. Among samples with a recorded latency, p50/p95 were 1548/4218 ms. See the [v0.4.8 report](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.8/eval/reports/2026-11-13-baseline-live.md). ## 0.4.7 - 2026-11-13 @@ -62,6 +64,8 @@ This history includes stable releases, prerelease tags and explicitly marked dev ## 0.4.6 - 2026-11-10 +Source milestone / 源码里程碑 2026-11-10 · [GitHub publication / 公开发布](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.6) 2026-11-13. + - Fix Linux compatibility by building the desktop, Python backend and pinned llama-server on Ubuntu 22.04 / glibc 2.35. - Reject packaged ELF imports newer than glibc 2.35, including weak imports. Validate deb installation and extracted AppImage startup with external networking disabled and loopback retained. - Linux validation covers desktop/backend startup and deterministic KaTeX rendering; it does not test model inference, every Linux distribution, or AppImageHub acceptance. @@ -71,13 +75,18 @@ This history includes stable releases, prerelease tags and explicitly marked dev ## 0.4.5 - 2026-11-08 +Source milestone / 源码里程碑 2026-11-08 · [GitHub publication / 公开发布](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.5) 2026-11-09. + - Clear stale history previews and their action closures on close, refresh/filter and successful history deletion without stealing search-field focus. / 关闭、刷新/筛选及成功清空历史时同步清理旧预览和操作引用,避免残留已移除内容或抢走搜索输入焦点。 - 修复历史矩阵/预设卡片渲染:KaTeX 每次使用独立可变宏表,避免冻结宏表在矩阵环境中抛错;保留宏隔离及可见的源码回退。历史「查看」改为页内只读数学预览,不再自动跳到预设或覆盖草稿;只有「复用输入」才恢复参数。大矩阵首尾列均可滚动查看。见 [#70](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/70)。 - Fix matrix/preset rendering with per-render mutable macro copies. View historical formulas in a read-only History preview without navigating or changing drafts; only explicit reuse restores editor parameters. Preserve full source, visible render failures, and scrollable edges for wide matrices. [Validation](docs/issue70-history-preview.md). +- 带补充上下文的矩阵和转写请求仍须满足原始公式约束,候选与约束冲突时阻止提交。 / Matrix and transcription requests with supplemental context retain their source formula constraints. A conflicting candidate is blocked from commit. [#69](https://github.com/CacinieP/TeXada-the-Math-Agent/pull/69) ## 0.4.4 - 2026-11-06 +Source milestone / 源码里程碑 2026-11-06 · [GitHub publication / 公开发布](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.4) 2026-11-08. + ### 中文 - macOS 插入前先隐藏窗口和应用,隐藏失败时返回错误并停止发送粘贴命令。托盘、快捷键和显示命令恢复应用后显示并聚焦窗口。新增无公式正文的插入阶段诊断,目标应用实际收文与恢复流程继续由 #64 原生验收。 @@ -161,6 +170,8 @@ This history includes stable releases, prerelease tags and explicitly marked dev ## 0.4.2-rc.3 - 2026-10-20 +Source milestone / 源码里程碑 2026-10-21 · [GitHub publication / 公开发布](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.2-rc.3) 2026-10-21. + ### 中文 - **行为变更:自然语言、OCR 与补全候选不再要求人工审阅,直接提交**。审阅面板不再出现,复制/插入立即可用,结果直接进入历史;compile/render/intent/CAS 证据仍分别报告,编译通过不代表转写准确,confidence 不是语义保证。公式改动范围守卫(MutationScopeGuard)保留:越界仍阻止提交并返回 `mutation_scope_review_required`。Studio 移交端点(`/api/handoff/formula`,`requires_review`)保留为唯一仍产生待审阅候选的路径。移除 `nl_require_review` 配置项。 @@ -400,6 +411,8 @@ OCR 继续使用独立配置的视觉模型;不安装 OCR 模型也可使用 ## 0.4.0 - 2026-08-25 +Source milestone / 源码里程碑 2026-08-26 · [GitHub publication / 公开发布](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.0) 2026-08-26. + ### English - Added an append-only `FormulaState` runtime with monotonic revisions, @@ -882,6 +895,8 @@ OCR 继续使用独立配置的视觉模型;不安装 OCR 模型也可使用 ## 0.2.6 - 2026-07-20 +Source milestone / 源码里程碑 2026-07-21 · [GitHub publication / 公开发布](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.2.6) 2026-07-21. + - Added Settings -> Data controls for full backup export/import, history export/import, and confirmed history clearing. - Added JSON backup APIs for full data export/import and history-only export/import. - Added merge-mode history import with duplicate skipping and replace-mode support in the backend. @@ -984,6 +999,8 @@ Development snapshot recorded in [commit `2a756b4`](https://github.com/CacinieP/ ## 0.1.0 - 2026-07-07 +Source milestone / 源码里程碑 2026-07-08 · [GitHub publication / 公开发布](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.1.0) 2026-07-08. + - Released the Tauri desktop app surface for macOS DMG and Windows NSIS installers. - Added local-first Ollama MiniCPM defaults with configurable text and vision models. - Added OpenAI-compatible cloud mode with separate text and vision model settings. diff --git a/README.md b/README.md index feb98e3..eccddda 100644 --- a/README.md +++ b/README.md @@ -199,8 +199,7 @@ TeXada v0.4.10 uses `AGPL-3.0-or-later`. Older tags retain their own licenses. Closing the main window hides TeXada; restore it from the tray or the existing shortcut. To stop the app, choose **Quit** in its menu. If a model download is interrupted, retry in Settings: received data can resume, and completion appears -only after the whole batch succeeds. The v0.4.3 close/restore/Quit and cold -download acceptance record is in [#50](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/50). +only after the whole batch succeeds. Window hiding, exit logging and download resumption fixes shipped in v0.4.3. [#50](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/50) tracks the pending macOS GUI retest of continuous cold download, window restoration and Quit. ### From A Fresh Machine To First Formula @@ -598,15 +597,15 @@ Apple silicon. The September baseline used MiniCPM5-2B and a different pipeline. | Render layer | 98/100 | | Trace layer | 98/100 | | Deterministic repair pass rate | 53/53 (100%) | -| Latency p50 / p95, v0.4.8 | 1548 ms / 4218 ms | -| Latency p50 / p95, v0.4.9 | 1255 ms / 3325 ms | +| Recorded latency p50 / p95, v0.4.8 | 1548 ms / 4218 ms | +| Recorded latency p50 / p95, v0.4.9 | 1255 ms / 3325 ms | The v0.4.9 report records seven zero-token deterministic results across the golden set. See the [v0.4.8 report](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.8/eval/reports/2026-11-13-baseline-live.md) and [v0.4.9 report](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/reports/2026-11-13-baseline-live.md). Two golden entries (`010`, `073`) are refused by the computation-scope gate instead of transcribed; they are listed separately in the baseline report and -kept out of the denominators. `062` (`u 除以 v 的求导法则`) used to be refused too +excluded from the structural denominator. The compile, render and trace layers still use all 100 entries. `062` (`u 除以 v 的求导法则`) used to be refused too and is fixed in v0.4.8 by treating `求导法则/公式/规则/表` as nouns. The remaining two are documented as a deliberate boundary in [eval/known-gaps.md](eval/known-gaps.md) (G-016): `010` asks for a factorisation @@ -779,7 +778,7 @@ macOS 需要 13.3 或更新版本,两个 DMG 均已签名和公证。Windows v0.4.10 使用 `AGPL-3.0-or-later`。`v0.3.6` 及更早产物为 `GPL-3.0-or-later`,`v0.3.7` 起为 `AGPL-3.0-or-later`。历史安装包见 [Release 列表](https://github.com/CacinieP/TeXada-the-Math-Agent/releases)。 -关闭主窗口会隐藏 TeXada,可通过托盘或现有快捷键恢复。需要停止应用时,请在菜单中选择 **Quit / 退出**。模型下载中断后可在设置中重试,已接收的数据支持续传,整批文件成功后才显示完成。v0.4.3 的原生关闭、恢复、退出及冷下载验收见 [#50](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/50)。 +关闭主窗口会隐藏 TeXada,可通过托盘或现有快捷键恢复。需要停止应用时,请在菜单中选择 **Quit / 退出**。模型下载中断后可在设置中重试,已接收的数据支持续传,整批文件成功后才显示完成。关闭隐藏、退出记录和下载续传修复已随 v0.4.3 发布。[#50](https://github.com/CacinieP/TeXada-the-Math-Agent/issues/50) 跟踪 macOS 正常桌面会话中仍待完成的连续冷下载、窗口恢复和退出复验。 ### 从裸机到第一次使用 @@ -1160,13 +1159,12 @@ Linux 尚未实现系统光标自动键入;点击公式后需手动粘贴。Wa | 渲染层 | 98/100 | | 轨迹层 | 98/100 | | 确定性修复通过率 | 53/53(100%) | -| v0.4.8 延迟 p50 / p95 | 1548 ms / 4218 ms | -| v0.4.9 延迟 p50 / p95 | 1255 ms / 3325 ms | +| v0.4.8 有记录的样本延迟 p50 / p95 | 1548 ms / 4218 ms | +| v0.4.9 有记录的样本延迟 p50 / p95 | 1255 ms / 3325 ms | v0.4.9 的黄金集记录了 7 条零 token 的确定性结果。 [v0.4.8 报告](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.8/eval/reports/2026-11-13-baseline-live.md) · [v0.4.9 报告](https://github.com/CacinieP/TeXada-the-Math-Agent/blob/v0.4.9/eval/reports/2026-11-13-baseline-live.md) -其中 `010`、`073` 两条被计算范围门禁拒绝而不是转写,基线报告里单列、不计入 -分母。`062`(`u 除以 v 的求导法则`)此前也被误拒,v0.4.8 已修——`求导法则/ +其中 `010`、`073` 两条被计算范围门禁拒绝而不是转写,基线报告将其单列,并从结构层分母排除。编译、渲染和轨迹层仍按 100 条统计。`062`(`u 除以 v 的求导法则`)此前也被误拒,v0.4.8 已修——`求导法则/ 公式/规则/表` 现在按名词处理。剩下两条是**有意的产品边界**,登记在 [eval/known-gaps.md](eval/known-gaps.md) G-016:`010` 要的是因式分解**结果**, `073` 无法用正则与真实的求导请求区分。 diff --git a/docs/README.md b/docs/README.md index 40051d7..7a77338 100644 --- a/docs/README.md +++ b/docs/README.md @@ -55,6 +55,6 @@ This directory keeps technical and maintenance documentation. Community health f | [Support](../.github/SUPPORT.md) | Bug-report checklist and model/Ollama troubleshooting entry point. | | [Code of Conduct](../.github/CODE_OF_CONDUCT.md) | Community behavior expectations. | | [v0.4.10 release](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.10) · [Changelog entry](../CHANGELOG.md#0410---2026-11-14) | Installers, first-run locale selection, and Linux release and catalog evidence. | -| [v0.4.9](release-v0.4.9.md), [v0.4.8](release-v0.4.8.md), [v0.4.7](release-v0.4.7.md) | Glossary, computation-scope fixes, deterministic repair and tagged evaluation baselines. | +| [v0.4.9 source tag / 源码标签](release-v0.4.9.md), [v0.4.8](release-v0.4.8.md), [v0.4.7](release-v0.4.7.md) | Glossary, computation-scope fixes, deterministic repair and tagged evaluation baselines. | | [v0.4.4](release-v0.4.4.md), [v0.4.3](release-v0.4.3.md), [v0.4.2](release-v0.4.2.md) | Matrix and transcription changes, plus historical release and installer notes. | | [Changelog](../CHANGELOG.md) | Release history. | diff --git a/docs/release-v0.4.9.md b/docs/release-v0.4.9.md index 20d4f54..14a63c8 100644 --- a/docs/release-v0.4.9.md +++ b/docs/release-v0.4.9.md @@ -1,11 +1,11 @@ # TeXada v0.4.9 -发布日期 / Release date 2026-11-13(Asia/Shanghai)。 +源码标签日期 / Source tag date 2026-11-13(Asia/Shanghai)。 v0.4.9 解决"只报定理名就答不出/答错"这一档:术语名改用受控词表直出, 把 1B 模型从"回忆教科书公式"里解放出来。 -[下载 Release](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.9) · [完整 diff](https://github.com/CacinieP/TeXada-the-Math-Agent/compare/v0.4.8...v0.4.9) · [CHANGELOG](../CHANGELOG.md#049---2026-11-13) · [使用指南](../README.md) +[查看源码标签 / View source tag](https://github.com/CacinieP/TeXada-the-Math-Agent/tree/v0.4.9) · [v0.4.10 安装包 / Installers](https://github.com/CacinieP/TeXada-the-Math-Agent/releases/tag/v0.4.10) · [完整 diff](https://github.com/CacinieP/TeXada-the-Math-Agent/compare/v0.4.8...v0.4.9) · [CHANGELOG](../CHANGELOG.md#049---2026-11-13) · [使用指南](../README.md) ## 本版变化