A grade you can trust requires a tool willing to say no.
Child Evidence sits in the gap between legal-research databases and science search engines. It does one thing neither does: it certifies the evidentiary weight of a developmental-science claim for legal use. Here is exactly how — across 485 verified studies and 44 bridged cases.
The five gates
- 01
Sourced
Every claim is traced to verified, peer-reviewed studies. No citation is ever generated from a model's memory — each one is a foreign key to a real, human-verified source row. The system is physically unable to render a citation that isn't a real study.
- 02
Weighed
A human reviewer grades evidence strength on replication, study design, and effect size — High, Moderate, Low, Very low. The grade is stored with its criteria and rationale, never as an opaque score.
- 03
Calibrated
A second, separate axis records scientific consensus — Established, Emerging, Contested, Disputed — so a strong finding that the field still debates is never mistaken for settled doctrine.
- 04
Bridged
Each claim links to the case law that already relied on the same science. This is the slow, hard-to-copy work that turns a scientific finding into a cited, judicially-accepted proposition.
Stress-tested
Limits and counter-considerations are surfaced deliberately — what opposing counsel will say, where the science is contested. When a proposition can't clear the bar, it is rejected and never certified.
The grading scale
Two independent axes, graded by a human reviewer. These rows render from the same definitions the product uses, so this page cannot drift from the instrument.
Evidence strength
- High
- Moderate
- Low
- Very low
Scientific consensus
- Established
- Emerging
- Contested
- Disputed
What “Certified for citation” means
A claim wears the seal only when both axes clear the bar: evidence strength of High or Moderate and consensus of Established or Emerging. A claim with Low or Very low evidence, or Contested or Disputed consensus, is still published and graded — but never certified. That scarcity is what keeps the seal defensible in front of a court.
Where AI is used — and where it is barred
Nothing you read on a published claim page is generated by a model at request time. AI appears in exactly three places, each with a defined job and a fail-closed posture:
Editorial drafting & screening (offline)
Google Gemini models help editors draft claim candidates and screen whether a candidate is actually grounded in its source paper. The screen is advisory only, runs at temperature 0 against a strict schema, and fails closed: an API error or malformed answer is recorded as “not grounded,” never as a pass. The model can never set an evidence or consensus rating, and nothing it drafts publishes without human sign-off.
Retrieval certification (offline)
A version-pinned Anthropic Claude Haiku model (claude-haiku-4-5-20251001) grades our search quality against a fixed evaluation set — temperature 0, answers constrained to the exact records supplied, verbatim-quote grounding enforced in code, and every failure mode resolved as a rejection.
Search ranking (the one request-time use)
Your search query is converted to a numerical embedding (OpenAI text-embedding-3-small) to rank published claims by relevance. That is the only moment user input reaches a model, it is ranking only — no generation — and inputs are never used for training. If the embedding service is unavailable, search degrades to keyword matching rather than failing.
The limits, stated honestly: ranking can miss a relevant claim or surface a marginal one, editorial screens can err before a human catches it, and science moves after a review date. That is why every claim shows its last-reviewed date and full citations — verify every study yourself before you cite it.
The non-negotiables
- A human performs every act of judgment — the canonical phrasing, both ratings, the guardrails, and sign-off. AI drafts; it never publishes.
- The public product only ever shows published, human-reviewed claims — enforced at the database, not just the interface.
- The platform never needs privileged case facts. Claims are about the science, not your client — privacy by design.