A LITTLE MORE CLARITY

The little release notes

Cress AI 1.0.120 · English-only local AI public Windows edition.

New in 1.0.120 — source selection and presentation

Fixes the reported “who won fifa 2022” response that refused to select an answer despite an explicit winner sentence in the returned excerpts. A narrow fallback can select a verbatim positive winner assertion matching the public event and year. It does not hardcode Argentina or apply to arbitrary questions, and different explicit winners are unresolved by this fallback.

Other factual questions use constrained sentence-ID selection by the same local model instead of asking it to copy long quote JSON. The app maps the selected ID to exact source text; invalid or truncated choices fail safely. The short source-linked answer appears first. Full excerpts, retrieval time and verification limitations remain available in an expandable Sources section; read-aloud skips that disclosure.

Validation for 1.0.120

461 automated app tests and syntax checks passed. The screenshot regression failed before repair and passed afterward. The exact wording returned a short Argentina answer through real DuckDuckGo lookup and the running local Public app. Two real Qwen3.5 2B CPU checks selected the exact supplied Ottawa and Alexander Fleming evidence sentences. An isolated packaged UI check confirmed the short answer and collapsed source disclosure. These are limited checks, not a general accuracy benchmark or full-page verification.

The model, saved consent and desktop-action boundaries are unchanged. AI and search excerpts can still be wrong, stale or conflicting. No cloud inference, model upgrade or automatic download was added. Old saved replies are not rewritten: ask again to receive a new answer.

Retained from 1.0.119 — Cress AI

A fresh name and a cress-sprout identity for your local workspace. App labels, assistant identity, Windows executable and installer branding now use Cress AI. Existing legacy storage paths and app/installer IDs are deliberately retained for update compatibility, so chats, settings, saved approval choices, credentials and verified local models remain in their original profile. No account or new model download is required just for the rename.

The local model, privacy boundaries and supported actions are unchanged. The 1.0.118 factual lookup and uncertainty improvements remain. A new name is not a factual-accuracy improvement, trademark clearance or a guarantee that search engines recognise the app.

Retained from 1.0.118 — factual lookup and uncertainty

Opted-in web excerpts now cover more explicit public factual questions, including causes, mechanisms, winners and fact checks. Up to ten search candidates are ranked by question-term relevance, with a small institutional/reference-source preference and different-hostname display priority. Up to five linked excerpts are shown. Ranking is not truth verification, source agreement detection or full-page reading.

The local model may select an exact quote, not invent an answer or citation. When no supported quote is selected, Cress AI explicitly says it is unsure rather than guessing. Local chat instructions also discourage invented names, dates, statistics and citations, and ask for correction of false premises. Instructions cannot guarantee compliance. Qwen3.5 2B is unchanged; factual mistakes remain possible.

Validation for 1.0.118

454 automated app tests pass, including common-question routing/grounding fixtures, rejected invented quotes, missing evidence and privacy checks. Real searches and packaged UI follow-ups returned linked excerpts for geography, science and film questions. These checks are not a measured model-accuracy improvement, full-page fact verification or a guarantee of correct answers. The installer was extracted without running it, and all 83 clean-build files matched size and SHA-256.

Web excerpts remain off for new profiles until enabled in Preferences → Tools & voice. Saved consent stays unchanged on update. No cloud AI, new model/runtime/voice, automatic retry or autonomous control. Close Cress AI and install this update on each PC; there is no automatic updater.

Retained from 1.0.117 — YouTube command typo fix

The explicit request play christmas hits on yotube now routes to the supported YouTube action instead of the generic unsupported-controls response. This adds only a fixed service-name alias; the query is not rewritten and arbitrary similar websites are not accepted. Existing Ask first/Automatic policy, validation and safe search-results fallback remain. YouTube may require consent, sign-in or Play; playback is not guaranteed.

435 app tests and syntax checks passed. The exact screenshot request was reproduced before repair, then checked through HTTP and the packaged UI using mocked video lookup/opening. No real music played, volume changes, screenshots of personal content or external sites were needed. Retains 1.0.116 reliability repairs. No new model, download or autonomous control.

Retained from 1.0.116 — reliability fixes

Failed model/runtime downloads now preserve pre-existing partial files. New files still require exact size and SHA-256 verification. Accessibility references expire at the two-minute boundary and are rechecked after native approval; an expired target requires a fresh inspection. Controls and uncertain actions require an explicit positive native approval.

The website build also removes known generated discovery drafts when they are excluded, preventing stale draft artifacts in reused output folders. The public app includes an up-to-date setup README rather than historical release instructions.

Validation for 1.0.116

434 automated app tests and syntax checks passed, including regression cases reproduced before repair. Isolated packaged checks cover chat writing, PC clock routing, Send/Stop recovery, supported window sizes and computer-control UI with mocked operations. These are code and interface checks, not a new real-model accuracy or performance benchmark. No bug-free, security or arbitrary-app compatibility guarantee.

No new model, runtime, voice or autonomous/visual agent. Existing chats, settings, encrypted credentials, models, voices and saved approval/web-excerpt choices remain in their separate profile. Install this update manually on each PC.

Retained from 1.0.115

Automatic routine window management, bounded explicit tasks and exact-label control selection. This is an app-tool update, not a new visual model or unrestricted computer agent.

Routine computer controls — enable separately

In Tools → Computer control, approve enabling controls for this session. They start off and do not inherit your saved desktop Automatic/Ask first preference. Name one already-open, unambiguous same-user app. Try maximize Paint, minimize Notepad, restore Paint, focus Paint or inspect Paint.

For explicit routine steps: computer task: maximize Paint; focus Paint; inspect Paint. Up to 8 semicolon-separated routine steps, 60 seconds overall; inspect, focus, minimize, maximize, restore and numbered scrolling only. Syntax and app names validate before execution. A later runtime failure stops remaining steps; earlier requests are not undone. Windows restores a minimized window to its previous normal/maximized state.

After inspection, use click control 3 in Paint, click "Exact button label" in Paint or type "hello" into control 2 in Paint. Exact labels must match uniquely; control references expire after two minutes and are consumed after an action request. Every click and field replacement still requires separate native approval and cannot be batched. Innocuous labels can hide sending, deleting, overwriting, purchases, installations or account/security changes. Typing replaces the field and may trigger app side effects: review the approval.

Use the persistent Stop computer control button to disable controls and cancel pending work. Already requested actions cannot be undone. This is not unrestricted or unattended control: no visual screen understanding, no arbitrary automatic shell commands, no forced app closing or automatic save/security-prompt dismissal. System/security windows, terminals, Cress AI and OpenCode are protected. Some apps have no usable accessibility controls.

Retained from 1.0.114

Local conversation instructions now focus on direct writing, history-aware follow-up edits and complete code examples. Reply ceilings adapt to explicit brevity (160 tokens), normal chat (256) and detailed or code requests (512). The existing Qwen3.5 2B model is unchanged: this is not a model intelligence or speed upgrade, and longer replies may be slower. Real checks still found factual and instruction-following mistakes.

Includes the 1.0.113 repairs: ordinary writing no longer misroutes into desktop commands; current time/date questions use your PC clock; retired phone promotions are removed from Settings. Existing chats, encrypted credentials, settings, voice IDs, model files and saved Automatic/web-excerpt preferences remain. No cloud fallback or unapproved model downloads.

Retained local model and actions

Qwen3.5 2B Q4_K_M replaces the active Qwen2.5 model, using the configuration chosen after a hands-on test: temperature .3, top_k 20, top_p .8, repeat penalty 1.15, seed 42, adaptive 160/256/512-token reply ceilings, 4096 context, CPU only and no context shifting. No cloud inference or silent model fallback. Approve about 1.30 GB of model/runtime downloads if the exact verified Qwen3.5 2B setup is absent. Existing verified 2B files in the active profile setup are reused. Old Qwen2.5/4B folders remain inactive for rollback, never automatic fallback. Voice downloads are separate.

The small diagnostic comparison was slower than Qwen2.5 and still made factual/logic errors. This is a user-selected upgrade, not a guarantee of better answers or speed on every PC. Check important answers.

Normal app closing

“Close Paint” or “Close Spotify” requests a normal Windows close of matching visible, same-user windows. Save/confirmation prompts remain for you; Cress AI never force-quits or dismisses them. System/security windows and Cress AI are protected. Some apps may need manual closing. A requested close does not prove the app exited. Saved Automatic runs supported requests immediately; Ask first previews them.

Follow-up requests

After one named app, try “Close it” or “Open it again” (reopening uses the launch allowlist). For a built-in house, rocket, smiley, tree, flower or sun, try “Make it red and bigger”, “Change the background to black”, “Add text Hello” or “Make it a house instead”. Nine English colors, bounded sizes, backgrounds, short captions and supported subjects are available. Each edit creates a new PNG; it does not edit arbitrary files, photos or Paint contents. Existing files are never overwritten. Unclear or unsupported requests need clarification, not guessed actions. This is not unrestricted PC control or a coding agent.

Preserved safeguards and data

Chats, settings, encrypted credentials, old models/voices and saved Automatic/Ask first/web-excerpt preferences are preserved. New profiles default to Ask first and excerpts off. Ten validated launches, Google/YouTube, Windows output volume and optional bounded web excerpts remain. No arbitrary commands, deletion or automatic shell/code execution. Supported accessibility clicks and typing require separate native confirmation. Internet requests remain explicit or consented as described in Privacy. English-only speech setup downloads only English components.

1.0.115 validation

419 current automated app tests and syntax checks passed. A disposable real Windows fixture verified maximize/minimize/restore through observed state, routine tasks, denied/approved clicks and text replacement. Packaged UI checks used mocked app operations to verify routing, escaped labels and Stop preventing later steps. Fresh installed-archive checks used isolated storage and a model fixture. The earlier 1.0.115 installer matched its tested Public installation. These checks do not prove arbitrary-app compatibility, security, accuracy or successful app workflows.

Earlier validation and limits

Earlier automated desktop tests and packaged native checks covered real local CPU replies and cancellation/recovery, drawing follow-ups/no-overwrite, close/reopen UI, approval dismissal, local speech language migration, web-excerpt consent and mocked media/volume actions. A disposable Windows fixture verified a normal close and an app declining closing without being killed. No user apps were closed in tests. No broad app-compatibility, security or accuracy guarantee. Tests do not prove YouTube playback.

Updating Windows

Install this release on each PC; older installations do not update automatically. Close Cress AI before updating. Windows 10/11 x64, 8 GB RAM recommended and at least 2 GB free disk for local AI plus app/voice storage. No profiles, credentials or model caches are bundled.

Download integrity

Cress-AI-Setup-1.0.120-x64.exe · 120559638 bytes

SHA-256: 7caec0dc5543b59a1b49e508f7b29e2b3e0c7cb243ec32f2544b159829e5bb8e

Unsigned installer. Verify the source/checksum and keep antivirus/SmartScreen enabled. Developed by Jvo. Shared green orb retained.