FinTick

The market whispers before the wire catches up.

Designed by Claude Opus · Built with Qwen 3.8 and GPT 5.6 Sol via Hermes.

FinTick watches one exceptionally fast public financial stream and measures a useful gap:

When did the stream say it, and when did independent news confirm it?

Every post gets a durable outcome. Repeated posts collapse into one event. Concrete claims become structured facts. Independent news either corroborates the event, disputes it, or stays silent.

That silence matters.

Important

BREAKING · no corroboration yet means the stream may be ahead of the wire. FinTick treats that as a result, not a pipeline failure.

The stream is the signal. External stories with real URLs are sources. FinTick never inflates confidence by counting four rewrites of the same post as four independent confirmations.

The Edge Board

The browser UI is an auto-refreshing event board rather than a raw feed reader:

• uncorroborated events rise to the top;
• validation status is the visual focal point;
• repetition appears as via the stream · seen N×;
• confirmed events show how far the news lagged the stream;
• pipeline health exposes pending, retrying, and terminal outcomes.

The local board runs at http://127.0.0.1:8137.

Try it in 60 seconds

FinTick requires Python 3.11 or newer. The fixture demo is offline and needs no account, token, model server, or package install.

git clone https://github.com/msitarzewski/fintick.git
cd fintick
./run-demo.sh

The demo ingests four differently worded NVDA posts, assigns all four to one event with a fixture model response, and starts the Edge Board on port 8137. Open http://127.0.0.1:8137 to inspect the result.

Run the health check and full test suite separately:

python3 -m fintick doctor
python3 -m unittest discover -s tests -v

Run one pipeline cycle

python3 -m fintick ingest --fixture reference/nvda_repost_cluster.json
python3 -m fintick aggregate
python3 -m fintick validate
python3 -m fintick serve --port 8137

Run continuously

Run each worker in a separate terminal, or install the user services described below:

python3 -m fintick ingest --watch
python3 -m fintick aggregate --watch
python3 -m fintick validate --watch
python3 -m fintick serve --port 8137

Provider boundaries

Stage	Production default	Offline / local path
Stream ingest	Public Bluesky AppView API, no authentication	Checked-in fixtures
Event aggregation	Any OpenAI-compatible endpoint (FINTICK_LLM_*) — a free local ollama server by default, or a cloud model like GPT-5.6 Luna	Injected fixture responses or the same endpoint
News validation	common.vision Partner API	Direct Google News RSS or stubbed candidates
Storage and dashboard	SQLite + Python standard library	The same code path

Inference is a single forced-JSON call to an OpenAI-compatible chat endpoint — no agent, no shell, no OAuth, no credentials on disk beyond the API key you place in .env. Untrusted stream text is treated strictly as content; it never drives tools. Point FINTICK_LLM_BASE_URL/_API_KEY/_MODEL at local ollama (free, the default) or any hosted provider.

When common.vision is configured, publisher identity comes from article.metadata.source; article.feed is retained separately as ingestion provenance. The feed carrying a story is not counted as another publisher.

Architecture

flowchart LR
    A["Public financial stream<br/>signal"] --> B["Durable URI ingest"]
    B --> C["Oldest-pending<br/>accounting ledger"]
    C --> D["Short-ID aggregation<br/>OpenAI-compatible endpoint"]
    D --> E["Canonical events<br/>facts + instruments"]
    E --> F["Independent news hunt<br/>common.vision / RSS"]
    F --> G{"Validation"}
    G -->|no match| H["BREAKING"]
    G -->|corroborated| I["CONFIRMED<br/>+ lead time"]
    G -->|partial or disputed| J["DEVELOPING /<br/>CONTRADICTED"]
    H --> K["JSON API + Edge Board"]
    I --> K
    J --> K

Loading

The accountability contract

LLM output is useful only when the system can prove what happened to every input. FinTick wraps model judgment in a strict ledger:

Invariant	What FinTick enforces
Every selected post gets one outcome	assigned to one event, intentionally ignored with a reason, retryably errored, or terminally errored after bounded attempts
One post URI belongs to at most one event	ambiguous ownership fails closed
Retries preserve context	a failed subset retries as its original group rather than contaminating fresh work
Terminal decisions stay terminal	restarts and repeated runs are idempotent
Time means UTC instant	mixed timezone offsets cannot reorder pending work, health timestamps, or event spans
Social repetition is evidence, not corroboration	only claim-matched external stories can confirm an event
Validation owns validation state	aggregation cannot overwrite status or lead time

The ledger turns model decisions into an audit trail.

Verified on representative data

The current release candidate has been exercised against a copied-live 1,006-post database:

Check	Result
Post accounting	1,006 / 1,006 durable outcomes
Assigned to events	283
Intentionally ignored	45
Historical / out of scope	678
Pending, retrying, or terminal errors	0
Posts owned by more than one event	0
Offline test suite	143 / 143 passing

SQLite model

The database lives at data/fintick.db:

• posts retains every stream post by immutable URI;
• post_aggregation_decisions records the complete accounting trail;
• events stores canonical headlines, facts, instruments, status, and lead time;
• event_signals maps repeated posts to their single event;
• event_validations stores claim-matched external stories with publisher and feed provenance kept separate.

Aggregation uses deterministic event keys and never overwrites validation-owned fields. Validation is safe to rerun: an event can start as breaking, become confirmed when the wire catches up, or lose stale evidence that no longer meets the claim-overlap standard.

Unattended operation

The installer writes four rootless systemd --user units: ingest, aggregate, validate, and dashboard.

./setup-fintick-services.sh --dry-run
./setup-fintick-services.sh
systemctl --user status 'fintick-*'

Worker	Default cadence
Ingest	poll every 15 minutes and paginate to the last durable URI
Aggregate	process the oldest 50 pending posts, then check again every minute
Validate	search and revalidate eligible events every 5 minutes
Dashboard	serve on loopback port 8137; browser refresh every 20 seconds

A handoff that can say "no"

The installer treats service replacement as a transaction rather than a pile of shell commands:

1. Preflight rejects live worker processes, port collisions, dormant Supervisor definitions, unsafe credential files, and uncertain system state.
2. SQLite's online backup API captures committed WAL data before anything changes.
3. Replacement units must remain active through bounded liveness checks.
4. The dashboard must prove it serves the exact operational database by device and inode identity.
5. Post accounting must conserve the corpus, and a pre-existing backlog must move.
6. On failure, replacements are proven stopped before the database family and prior unit state are restored.

Rollback is fail-closed. Prior writers restart only after database recovery, unit-artifact restoration, and daemon-reload all succeed. Persistent and runtime masks are preserved exactly. Uncertain or transitional unit states produce an incomplete rollback with private evidence retained; they are never interpreted as success.

Note

Only the validation unit reads %h/.config/fintick/environment. If you enable common.vision, create that file with a secure editor, set mode 0600, and add FINTICK_COMMON_VISION_TOKEN. The installer verifies path, type, owner, and mode without reading the token value.

Model benchmark

The fixed 28-post corpus exercises the same short-ID parser and accounting contract used in production:

# Uses FINTICK_LLM_* from the environment (local ollama by default):
python3 -m fintick.benchmark
# Or point it at a specific endpoint/model:
python3 -m fintick.benchmark --base-url https://api.openai.com/v1 --model gpt-5.6-luna

The benchmark prints aggregate metrics only. Raw model responses are not written to the repository.

Repository map

Path	Purpose
PRD.md	product contract and acceptance criteria
STATUS.md	verified release state and operational evidence
AGENTS.md	engineering, integrity, inference, and service rules
reference/nvda_repost_cluster.json	canonical four-post aggregation fixture
reference/common_vision_partner_api.md	credential-free Partner API contract used by validation
tests/	offline acceptance, accounting, validation, dashboard, and handoff regressions

Related work: Agency Agents

Agency Agents is Michael's open library of specialized AI agent personalities, processes, and deliverables.

FinTick is a focused operating system rather than an agent catalog, but the projects share a bias: specialized intelligence should produce work you can inspect, test, and use. Agency Agents is related work, not a FinTick runtime dependency.

Support the work

If FinTick or Agency Agents helps your work, you can support Michael's open-source projects through GitHub Sponsors.

The collaboration

Claude Opus designed the product and acceptance contract. Qwen 3.8 and GPT 5.6 Sol built and hardened FinTick through Hermes. Michael set the taste, reframed the stream as signal rather than "sources," and kept the work honest: runnable systems over AI theater.

That provenance is part of the project, not decorative copy. FinTick was built by making model output answer to fixtures, ledgers, tests, copied-live data, and independent review.

License

MIT © 2026 Michael Sitarzewski.

FinTick is financial event-intelligence research software. It does not place trades and is not investment advice.