Capabilities

Pipeline Tools

The tool surface available to each pipeline stage. Tools are deterministic, audited, and rate-limited — the model proposes, the tool decides.

Why tools, not free-form prompts?

The free-form-LLM problem

A free-form LLM can confidently invent prices, hallucinate filings, or claim a stop was hit when it wasn’t. Telumin solves this by exposing a small, well-typed tool surface to each pipeline stage. Models propose; tools execute against real data; results are logged. If a tool says no, the pipeline stops.

The Tool Surface

Each pipeline stage has access to a different subset of tools. The SkepticValidator can read everything but cannot place orders; the SignalPublisher can write to the Evidence Vault and trigger workflows but cannot edit history.

01

DataForager tools

Patchright scraper pool, RSS / REST / WebSocket fetchers, EDGAR pull, earnings transcript fetcher, social filter (Reddit / X / StockTwits). Rate-limited per source. Outputs canonical claim documents.

02

Normalizer tools

Ticker resolver (figi / cusip / exchange), event-type classifier, source-tier registry, conflict-clusterer. Output: structured normalized claim ready for quant context.

03

QuantAnalyzer tools

Market data (last, OHLCV, options chains), implied / realized vol calculator, earnings calendar, options-flow snapshot, short-interest snapshot, beta / drawdown calculator. Pure-functional — no side effects.

04

SkepticValidator tools

Source-count enforcer (the 2-source rule), tier check, thesis-condition retriever (from the Investment Diary), historical accuracy lookup, and the verdict writer. Cannot place trades. Verdict is immutable once written.

05

SignalPublisher tools

Evidence Vault writer (append-only), Signal API push, webhook fan-out, Temporal workflow trigger (Thesis Guard / Earnings Distillation). Cannot edit historical signals.

07

Investment Diary

Read access for every stage; write access only from the operator. The Skeptic uses the Diary’s Day-1 conditions; the Thesis Guard workflow watches for contradictions. Live orders against undocumented positions are blocked at the workflow layer.

Skeptic Pass Requirements

A claim cannot become a published signal unless every condition is met:

Independent sources≥ 2
Source tier≥1 from Tier 1 (filings/exchanges) or Tier 2 (wires/calls)
Quant sanityTrade structure plausible vs IV regime & earnings proximity
Diary alignmentSignal direction reconciles with thesis or is flagged Drift
Confidence floor≥ 65 (0–100 scale)

Confidence Scoring

Every published signal carries a confidence score (0–100). Four weighted factors:

Source quality — 35%

Tier-1 filings + Tier-2 wires score higher than two Tier-3 social posts. Single-source social claims are vetoed before this score is even computed.

Quant context fit — 30%

Does the trade structure make sense given vol regime, earnings proximity, and options flow? Improbable structures get demoted.

Historical accuracy — 25%

Track record of the signal pattern on prior data. Novel patterns score conservatively until they accumulate evidence.

Recency — 10%

Inputs from today’s tape and today’s filings score higher than reasoning extrapolated from week-old data.

Pipeline Performance

Skeptic agreement (test set)94% precision on labelled holdout
Filing → Signal latency<3 min (8-K / 10-Q / 13F)
Single-source social vetoesAlways rejected
ReproducibilityVersioned models per signal
Audit trailTool-by-tool decision log per signal

See the tools in the Evidence Vault

Every signal in the Evidence Vault links to the exact tool calls that produced it — what was queried, what was returned, what the Skeptic decided, and why.