Data Injection for LLMs: Anatomy of an Auditable Signal

What a semantically pure signal is

A semantically pure signal is a record with explicit state: you know when it was detected, whether a human reviewed it, and whether a model generated an analysis on it. It isn’t a free-text paragraph —it’s a row with verifiable fields. That’s what makes it injectable into an LLM context without risking a hallucination: the model doesn’t have to infer the data’s provenance, the provenance is already in the record.

Anatomy of a signal

The real audit fields of a signal —not a simplified example— are these:

$ curl revsign.os/signal/8f2a1c
→ {
  company_domain:   "acme.com",
  signal_type:      "tech_stack_change",
  intent_score:     82,
  status:           "pending",
  approval_status:  null,
  date_detected:    "2026-07-14T09:00:00Z",
  reviewed_by:      null,
  reviewed_at:      null,
  insight_status:   "pending",
  version:          1
}
# Note: there are no "ingested_at"/"validated_at"/"logged_at" fields —
# the real chain is date_detected → approval_status/reviewed_at → insight_status.

Chain of custody: detection → review → insight

date_detected marks the ingestion moment. approval_status and reviewed_by/reviewed_at are the explicit human-approval checkpoint (HITL: human-in-the-loop) —a signal doesn’t move to action without that field being set. insight_status, along with insight_attempts and insight_model, records whether and how a model generated further analysis on the already-approved signal. Every stage lives in the same row, not in a separate log that can go missing.

How the context fed to the model gets assembled

1. Signals are filtered by status != 'suppressed' and by the intent band relevant to the query.
2. date_detected and approval_status are attached to every signal included in the context, so the model can cite the date and review state, not just the content.
3. Signals without reviewed_at are excluded from any context meant to produce an actionable recommendation —only what a human has already reviewed gets used for decisions, not for exploration.
4. The full record is handed over, not a summary: the model receives the fields, not a paraphrase of the fields.

Generic RAG vs. validated signal injection

Known limits

This doesn’t solve everything. Intent-score quality depends on tech-stack detection coverage —an unmonitored tool category can’t generate a delta. The conflict filter operates at the company-domain level, not at the individual-contact level. And the HITL checkpoint means no signal reaches a production LLM without a human having reviewed it at some point —that’s a deliberate design decision, not a limitation to be solved.

$ curl revsign.os/docs/data-injection-for-llms
→ {
  title:    "Data Injection for LLMs: Anatomy of an Auditable Signal",
  answers:  "How do I feed an LLM auditable data about my own company?",
  schema:   [TechArticle, HowTo]
}