Evidence before action.
Open stock research · Watch the walkthrough · Evidence and limitations
THESIS brings official announcements for 19 companies, source-linked AI research drafts and timestamped Bitget stock-token observations into one local workspace. The researcher makes the final decision. A separate paper simulator uses simulated funds.
Open https://thesis-research-bay.vercel.app/ . Select a stock, search its official announcements, read source passages, refresh a public Bitget stock-token quote, and save research in your browser. No account, API key, installation, model download or provider credential is required for this reader.
Publication dates stay visible: retrieving a feed now does not mean its articles were published recently. Article extraction can fail; the UI falls back to the feed excerpt with an explicit label. Quotes expire after fifteen seconds.
This is a source reader, not hosted AI. It copies issuer evidence and makes zero model calls. It does not generate business conclusions or execute trades. Optional local-model AI and the existing Groq workflow remain in the local app. Public hosted AI would require a separately configured operator-owned model service; no such shared credential is currently configured.
The full sidebar keeps Overview, Live inputs, Announcement review, saved AI
reviews, Portfolio & journal, Demo lab, Model setup, the video walkthrough and
browser-saved research directly accessible. demo.html#research, #journal,
#demo, #setup and #walkthrough open the corresponding existing feature;
Back and Forward retain the selected section.
The existing workspace contains saved AI reviews, scripted paper-execution examples, the video and original evidence downloads. Saved records keep their original dates and values. These navigation changes do not enable fresh hosted AI generation or a hosted paper runner. Current source reading and public quotes for 19 stocks remain available without a key. Video and original evidence files stay pinned to their existing immutable deployment.
The reader covers 19 companies: Apple, NVIDIA, Microsoft, Amazon, Meta, AMD, Intel, Broadcom, Alphabet/Google, Netflix, Coinbase, Salesforce, Qualcomm, PayPal, Uber, Nike, Coca-Cola, Visa and Mastercard. Each has an allowlisted issuer feed and a verified Bitget rToken quote endpoint. The last eleven were checked on 9 October 2026; upstream availability can change.
Company marks are bundled with the app, so visitors do not contact a third-party
logo service. public/logos/sources.json records the original website/icon source
and attribution. Logos identify companies; they do not imply endorsement.
public/ contains the browser UI and original recorded demo. The stateless,
read-only api/research.py Vercel Function uses hosted_reader.py. It accepts
catalogue symbols and feed event IDs, never visitor URLs. Issuer host/path
allowlists and bounded responses constrain collection. Successful sources are
CDN cached for five minutes, quotes for ten seconds. Errors are not cached.
Deploy the repository to Vercel using vercel.json. The hosted reader never
starts the local paper runner, writes SQLite files, reads credentials or calls an
AI provider. Browser saves stay on that device; clearing browser data removes
them. No model or Bitget trading key is embedded in public assets.
-
Install Python 3.10 or newer if needed. Clone this repository or download and extract its source ZIP into a NEW folder. Keep your existing THESIS installation and records in their current folder.
-
In Terminal, change to the repository folder containing
launch.py, then run:python3 launch.py
On Windows, use
py launch.py. The launcher opens http://127.0.0.1:8765. Keep the Terminal window open. Stop with Control+C in that window. -
Open Announcement review and click Refresh news.
-
Search any of the 19 companies, choose an announcement and click Read source. No key or model is required. The exact source passages are saved and clearly marked as zero-model-call source briefs.
-
For optional AI explanations, connect a local Ollama model under Model setup with the API key field blank. The advanced Groq connection also remains available if you choose to supply your own provider credential.
-
Inspect cited text and limitations. Portfolio & journal → Export complete run downloads local research and paper records with a hash manifest.
Install Ollama and qwen3:8b if you want the separate local paper decision model.
The launcher supplies its local endpoint defaults; Model setup has its
connection test. Research uses the separately selected Groq connection.
Refresh inputs, try One cycle, inspect the journal, then optionally Start
paper run. Pause stops monitoring. The app/computer must remain running.
Old announcements are not backdated into the one-hour entry window. A closed or
unverified session, stale quote or failed rule may prevent a simulated trade.
No exchange trading key is required, and the app never sends exchange orders.
The linked browser demo uses an actual user-exported run with 42 collected announcements, 23 research records across development versions and four separate simulated accounts. Those accounts contain zero recorded decisions and zero fills. Records include failed attempts and older model errors. A completed status means the app's mechanical checks completed; it does not establish semantic accuracy. One recent saved Groq source brief took 2.56 seconds of model time; its completed assessment took 3.64 seconds. These are individual observations, not a latency benchmark.
The draft's proposed future earnings check must be read conditionally: check whether relevant metrics are disclosed. The source does not guarantee that a future filing will contain Siri AI metrics. This editorial correction is shown separately in the browser demo; the original saved answer is not rewritten.
The source-copy and citation checks do not establish that an interpretation is true. No trading advantage, profitability, general model accuracy or independent user study has been demonstrated. The Demo lab uses scripted fictional decisions.
Python standard library server, SQLite, plain JavaScript/CSS/HTML. No pip install
is needed. Research uses Groq's openai/gpt-oss-120b; optional local paper decisions
use Ollama/Qwen. Official issuer feeds/articles and public Bitget APIs supply data.
The server binds to loopback with host/origin and per-session request checks.
Do not expose this local server directly as a shared multi-user hosted service.
The launcher uses an app-only Bitget DNS fallback with HTTPS verification intact.
It does not edit system DNS settings.
- Stop an earlier THESIS process if port 8765 is already in use.
- Reload the tab after a restart if request/session verification fails.
- Use the visible Groq cooldown; do not repeatedly click while it is waiting.
- Public news or market services can fail independently. Inspect saved records and try a fresh refresh later; never present saved quotes as live.
- A fresh install starts with empty working records. The demo evidence is separate.
- If
launch.pycannot resolve Bitget, the app can still open for saved inspection.
The Python checks require Python 3.10+. The separate JavaScript checks require Node.js 18+. Running the app itself does not require Node.js.
python3 -m unittest discover -s tests
node --test tests/test_research_ui.cjs tests/test_reader_ui.cjsThese checks use fixtures and test doubles. They do not verify live model quality. Historical development notes are preserved in DEVELOPMENT-HISTORY.md and the other included notes; this README is the current setup guide.
| File or folder | Purpose |
|---|---|
launch.py, app.py |
Local launcher and HTTP server |
pipeline.py, issuer_article.py |
Public news and market collection |
research.py, evidence_brief.py, review_steps.py |
Saved research workflow and source briefs |
research_groq.py, research_reasoning.py, research_grounding.py |
Research model adapter, prompt and evidence checks |
agent.py, runner.py |
Paper decisions, execution checks and monitoring |
interface.html, research.js, research.css, journal.js, journal.css |
User interface |
run_export.py |
Evidence export and manifest |
tests/ |
Automated checks and development fixtures |
evidence/ |
Bundled observations and explicitly fictional demo inputs |
The app creates run-data/ locally when started. That folder is excluded from
Git. Share a deliberately reviewed export when publishing research evidence.
The Vercel site combines current public-source research with the original saved
AI and paper workspace. This repository also runs fresh research locally, with
an optional model connection.