Sia Partners · BID Framework

How BID works — three pillars, ten agents, one auditable pipeline.

Every value in every deliverable traces back to a public-filing line or a Sia proprietary input. The flow below is what runs end-to-end on every BID job.
← Back to deliverables

01 · The flow

A typed JobRequest enters on the left. Each pillar transforms the data and hands it to the next via a Handoff envelope — no agent calls another directly. The orchestrator manages persistence, retries, and audit.

InputJob request
Pillar 1Baseline
Pillar 2Intelligence
Pillar 3Decision
OutputDeliverable
+ audit

02 · The ten agents

Each pillar owns three or four agents. Every agent has a single, narrow responsibility and emits a typed handoff. The same envelope shape flows pillar to pillar — so a downstream agent never reconstructs upstream context.

Pillar 1 Baseline
Pulls raw inputs, canonicalizes naming, resolves entities to a single source of truth. Output: a clean, comparable dataset.
Source / Extraction
SEC EDGAR XBRL + R-file HTML retrieval. Pulls income-statement lines for each cohort firm × period.
Normalization
Canonicalizes entity / metric / period naming, normalizes units, collapses synonymous concepts.
Resolution
Dedups, reconciles cross-source conflicts, resolves entities to canonical IDs.
Pillar 2 Intelligence
Builds the analytical table, derives computed metrics, generates the cohort and per-firm narrative. Output: insights + comparisons.
Analytical Table
Builds the per-entity × per-metric × per-period table — the central data structure every downstream agent reads.
Performance Metrics
Derives ratios, growth rates, and computed metrics (e.g., applied AI = Σ of three sub-track ratios × public denominators).
Comparisons Synthesis
Peer-vs-peer narrative with driver attribution. What's different, why, by how much.
Insight Synthesis
Cohort-level headlines, outlier surfacing, posture framing.
Pillar 3 Decision
Packages the analytical output into a client-ready deliverable, runs the validation gates, ships through the right channel. Output: rendered HTML / PDF + audit sidecar.
Output Ingestion
Packages the Intelligence handoff into the deliverable shape — section registry, audience visibility tags.
Visualization
Chart specs, section dispatch, brand styling, editable layer + chart controls.
Delivery / Distribution
Audience-mode filtering, anonymization check, pre-emission linter, final emission.

03 · What goes in, what comes out

Data in

BID pulls from a small, deliberately tight set of sources. Every value rendered carries the source it came from.

  • SEC EDGAR XBRL — companyfacts API · structured income-statement lines · us-gaap:* concepts per filer
  • SEC EDGAR R-files — HTML income-statement scrape for firms using firm-extension XBRL tags
  • Survey ingest — CSV / Excel respondent-level data (e.g., the inaugural B2B Enterprise AI survey, n=200)
  • Sia proprietary data — measured cost decomposition from in-engagement firms · drives the AI intensity ratios. Specific values held in audit JSON, never client HTML

Audit trail out

Every deliverable ships with a sidecar that records the full pipeline: every agent's input + output, every connector call, every ratio, every reconciliation step.

  • Audit JSON — machine-readable, complete provenance per value · the analyst's source of truth
  • Audit report HTML — human-readable trace · pillar-by-pillar walkthrough · linked from each deliverable's footer
  • Validation gates — anonymization check, engagement-data leak gate, pre-emission linter (quarantine-phrase pass)
  • Lineage on every cell — XBRL tag · accession number · filed date · taxonomy version
The pattern in one line: public-filing denominators × Sia's proprietary calibration ratios, with every value traceable to its source — and a closed-enum section registry + audience-mode gate that decides what surfaces in the client view vs the analyst audit.

04 · See it in action

Pick a live deliverable — every prose block is editable, every chart has controls, and the audit report links from each deliverable's footer.