About Seismogramm

Seismogramm is Seismo's greenfield intelligence selection and synthesis pipeline. It uses a decoupled service architecture to run a multi-pass Gemini model flow, optimizing context length, reducing token cost, and preventing hallucinated signals.

1. Execution Lifecycle

To maximize summary density and avoid prompt dilution, Seismogramm divides its workflow into logical execution passes:

• Query Expansion (Pre-pass, Research mode only): Gemini expands a single target query into a list of 3–10 high-fidelity, non-redundant synonyms, translations, and associated entities to ensure comprehensive database retrieval.
• Gather → cluster → cap: PHP loads eligible rows, collapses near-duplicates, then keeps at most 100. Magnitu decides who is in the cap, not editorial rank.
• Mode floor (Briefing and Headlines): one short Flash call labels each capped row news / enumerated junk (quiz, gadget, lifestyle, celebrity, fun-news — not press releases) / unsure, and groups same-fight rewrites. PHP drops junk from the pool and keeps one original per dossier before the XML is sealed, so Pass 1 never sees a Mag-90 quiz sitting on top of real news. Guard rails refuse to starve a briefing. Off: Advanced → Disable pool floor, or seismogramm:pool_floor=0.
• Pass 1: Selection Phase (Triage): The model evaluates the sealed XML against the mode’s selection instructions and returns a strict JSON schema containing used_entry_keys (entry_type:entry_id) and required selection_reasoning. Returning fewer than the item ceiling is required when the rest is filler. Conversational responses are prohibited in this phase.
• Pass 2: Synthesis Phase (Formatting): Resolves the selected keys and generates a clean XML context containing only the target entries. The model then formats the final Markdown briefing according to the selected tone preset and layout instructions. The first Pass-2 call already states PHASE 1 is done (Markdown + (entry_type:entry_id) citations, no JSON object); if Flash still echoes Pass-1 JSON, synthesis retries once. Headlines Pass 2 is a local list; Events uses extract/rank/writeup.

2. Presets Parameters Matrix

Source pills and Only items with tag sit above Advanced (they apply without opening the panel). EU Digest / Security Digest auto-tag eu / security; Foreign Lex auto-tags Government / Parl and defaults Included Sources to Feeds + Non-CH Lex + Leg. Briefing Regulatory defaults to Lex + Leg. Save-preset knobs restore an explicit tag or modules[] instead of letting Persona overwrite them. Feeds cycles Feeds → Substack → off (same pattern as Lex → CH Lex → Non-CH Lex → off). Turning every Headlines pill off stays off (Generate sends an explicit empty selection; CLI/no-JS still get Media + Newsletter).

3. Preset Mode Specifications

Briefing Mode

Designed to scan the entire incoming landscape to compile a daily or weekly intelligence digest for Swiss decision-makers. Tone and layout are determined dynamically by the selected Tone Preset:

• Executive (Default): Usually 2 sentences; a third only with a date, article/§, amount, or named body from the source; never a fourth.
• Digest: Same two-sentence default and typed third (no required “why this matters at work”).
• Regulatory: Same length rule; third sentence is a sourced date/article/amount, with jurisdiction. Default sources: Lex + Leg only (pills stay editable).
• Newsroom: Focuses on public news values, controversies, and story-hooks.
• Quotes: Captures direct statements, stances, and opinions.
• Contrarian: Two sentences for the break; no “why bemerkenswert” closer.
• EU Digest: EU story first; same two-sentence default (no required Swiss coda).
• Security Digest: Security story first; same length rule.
• Foreign Lex: Foreign gazette / parliament → Swiss transmission (or honest «kein direkter CH-Effekt»). Auto-tag Government / Parl; default sources Feeds + Non-CH Lex + Leg. Pass 1 prefers evidenced foreign instruments (incl. SEE/origin / winning comma windows); CH Leg is sequel, not the lead.

Kern-Logik: Uses Magnitu relevance scores as a prior, but the AI persona/goal can override ties if the context fit is stronger. A mode floor (quiz, gadget, lifestyle, celebrity, fun-news) runs before XML so Pass 1 never sees that junk; corporate PR is still a tone job. Mail that would Expand to the full newsletter (isolation looks truncated) is sent to Gemini as that fuller plain text. Context Cap: Up to 100 entries.

Lex in the pool: Briefing and Blindspot pass operative corpus (or the recipe’s winning omnibus unit) into Pass 2 — not the gazette blurb alone. XML also carries a deterministic <lex_extract> (instrument / scope / pflicht / frist). Missing the word «Schweiz» is not a discard signal: preferential EU∪EWR gates can still be Swiss-relevant by omission.

Blindspot Mode

Focuses on identifying "hidden signals"—high-impact regulatory and legislative moves that have not yet been picked up by general media outlets or newsletters (information asymmetry). Supports curated tones (Regulatory, Executive, Newsroom, SRF Monitor, Security Monitor). Non-CH Lex with Swiss-nexus / territorial-gate cues (including deep mid-body EWR/SEE language) is sorted ahead of bare foreign Lex so it survives the context cap. Context Cap: Up to 100 entries.

Negative-Space Protocol:

1. Echo Footprint: Titles-only index (global_fingerprint / GLOBAL POOL INDEX) from echo sources over a wider window (default 7 days) than the candidate lookback (default 3 days). Soft-capped. Default echo modules: Media + Newsletter. 600-char teasers stay on the gather for PHP ID matching, not Pass 1.
2. Honest vacuum: Pass 2 may claim silence only inside that echo window when echo_status=vacuum (never “never covered”, «Radar der Öffentlichkeit», and never when Flash returns inconclusive). PHP hard-coverage is ID-only; soft headline shortlists need Flash same-dossier confirm after Pass 1. When the window is empty for that item, a fixed 60-day media+newsletter index may quote same-dossier titles after a Flash confirm — not thematic near-misses. prior_match_count=0 means no 60-day sentence.
3. Primary Triage: The 100-row Gemini cap is Lex / Leg / Documents only. Echo modules stay on the title index and do not compete for those slots.
4. Negative-Space Check: PHP verdicts own coverage. Pass 1 ranks remaining primaries by impact; it does not guess echo from the fingerprint.
5. Output Selection: Prioritizes high-impact regulatory items with zero overlap in the echo footprint. Runs with an empty echo corpus are rejected server-side.

Research Mode

Allows targeted ad-hoc investigations. It takes a mandatory user query and filters the entire source database for matches. Bypasses Magnitu relevance scores. Extracts keyword snippets of exactly 600 characters to keep context inputs focused. Context Cap: Up to 300 entries.

Monitor Mode

Tracks specific organizations, companies, or individuals from a watchlist (either custom or built-in Swissmem) across all active channels. Requires verbatim name verification; loose semantic groupings are rejected. If the gathered pool contains zero verified matches, the pipeline bypasses Gemini execution entirely and returns the standard empty template. Context Cap: Up to 100 entries.

Risiko Mode

Risk flagger (code preset Risk): cards for Risks / Shifts / Threats / Impact — not a Lage essay. Soft press is alerts-only; Lex/Leg/docs/email and Substack Voice use include-more. Motions seat soft; primary ceiling high. Mode floor + thin seats; Risk-only Pass 1. Empty → Keine Risiko-Flags. Context Cap: Up to 100 entries.

Events Mode

Builds an attendable agenda from Agenda inventory plus scavenged mail/feeds. Lookback is fixed at 7 days (hidden); the visible Include events through horizon (default 14 days) filters by event start date. Agenda feeds need a start in the horizon (event_meta.starts_at or published_date proxy) before Gemini; scavenged sources are extracted then horizon-capped. Pool priority stays event/agenda-first even when Advanced is open. Final lists reserve non-Agenda slots among ranked on-desk events so calendar firehoses do not drown invitation gems. Featured count is an upper bound (default 10) — thin desks stay short. Context Cap: Up to 200 entries.

Headlines Mode

Skim list of the day’s most important media + newsletter stories. Same Persona dropdown + role card as Briefing, curated to Newsroom (default), Executive, SRF Monitor, EU Digest, and Security Digest; Headlines-specific Pass‑1 criteria prefer national / international / CH–EU over Kanton/Gemeinde routine and soft filler. Default lookback is Today (Europe/Zurich calendar day); Magnitu is a soft prior (no hard score gate). Pool order demotes regional/`punkt4`-class digests before Mag sort (forced even when Advanced is open). Source pills default to media + newsletter and are honored without opening Advanced. Story clustering collapses near-duplicates and prefers the media article URL over a newsletter wrap. The same mode floor as Briefing drops quizzes/gadgets before Pass 1. Pass 1 selects up to 20 IDs and labels themes; Pass 2 is a local numbered Markdown list grouped by theme (CH–EU → Security → Energy → … → Weitere) when multiple themes appear, not Gemini prose. Context Cap: Up to 100 entries.

4. Pipeline & Selection Engines

• Standard Selection Engine: Sends the entire XML context pool in a single payload to Gemini. Recommended for smaller pools.
• Tournament Selection Engine: Splits large context pools into parallel batches (size ~35). The winners of each batch are consolidated and sent to a final selection pass. Mandatory for Research; Auto for Briefing / Monitor / Headlines when the capped pool exceeds 80 entries (still Auto when Advanced is open unless you pin Standard). In-flight cap: Settings → General Parallel Gemini requests (default 5).

5. Resilience & Observability

• Tournament Batch Recovery: If a parallel batch fails, the system retries it sequentially once before continuing, appending error metrics to the metadata.
• User-Initiated 429 Retry: HTTP 429 shows Gemini’s own error text. Retry with smaller pool appears only for request-rate / token-rate throttling. Prepaid/billing failures (empty credits) do not offer a smaller cap — the pool size is not the cause. Research still fails fast with a shorter-lookback hint on a true rate throttle.
• Token & Cost Observability: Real-time token usage is tracked by phase (selection, summary, context caching; Events: extract, rank, local render) and used to calculate execution costs shown in the UI footer. Wall-clock per phase is archived as phase_durations_ms.
• Context Caching: For pools where the title fingerprint exceeds 50,000 characters, Seismo uploads the fingerprint once using Gemini Context Caching and references it across parallel tournament requests to save token costs.

6. Pipeline Diagram

mermaid
flowchart LR
    UserQuery[Search Query] -->|Research Mode| QE[Query Expansion]
    QE --> A[Gather & Filter]
    OtherPresets[Other Presets] --> A
    A --> B[XML Context Pool]
    B --> C[Pass 1: Selection Phase]
    C --> D{Selection Engine}
    D -->|Standard| E[Full Pool Triage]
    D -->|Tournament| F[Parallel Batches]
    E & F --> G[Filter XML to Selected Keys]
    G --> H[Pass 2: Synthesis Phase]
    H --> I[Markdown Briefing]