Living Evidence Atlas — mini
One estimand cell, 18 experiments represented by 19 effect-size records, six machine-checkable claims and three gaps the page computes about itself — as a map you and your agent explore together.
What this is. A living document answers questions about its own literature. An Atlas is the index above many such documents: it holds the derived layer — the graph of cells, claims, records and edges, the syntheses, and the gaps that fall out of them. This page is that direction rendered at demo scale, over the one literature this project ships: teacher expectancy and pupil IQ.
Everything on the map is computed in your browser from the same 19 effect-size records from 18 experiments the document carries, and the six claims are imported from the same module the document boots from — the map’s rules cannot drift from the document’s (one shared module), and the test suite holds the displayed sentences to the same text. Design rationale and the full ontology: DESIGN.md (§7 for the layer this belongs to). Siblings: the exemplar document and the workspace that builds one.
This is the Atlas direction at demo scale: one literature, computed live, read-only. No dossier scores, no numeric power, record verification rungs unassigned.
Nothing here mutates the graph or the evidence base — there is no propose_edge, no
persistence and no backend. The weeks moderator is labelled candidate because it is a
study-level, observational comparison across randomized experiments, not a randomized contrast. Gaps are computed
from this collection frame: “unknown / not-searched” is a different state from “measured and absent”.
The Pellegrini & Hicks (1972) experiment contributes separate aware- and blind-tester records. The reference fit reproduces the historical 19-row analysis and does not model their within-experiment covariance.
If you are a human
Click any node — the estimand cell in the middle, a claim, one of the 19 effect-size record dots, or a dashed gap card on the right. Its detail opens in the panel. Every node is keyboard-reachable: tab to it and press Enter. You can run the agent’s tools yourself from the Tool console.
If you are an agent
Call atlas_overview first. get_cell, list_claims and
get_gaps are pure reads; evaluate_claim, synthesize,
get_study_brief and focus_node render into the map the human is looking at and land in
the ledger. Cite numbers from the tools, never from the picture.
The map
One estimand cell (centre), the two constructs it relates, the document that asserts claims about it, the six claims (upper arc), the 19 evidence records from 18 experiments (lower fan — dot size grows with precision (1/vi), floored for legibility; colour by weeks of prior teacher–pupil contact, ordered left to right by that same field), the candidate moderator, and three computed gaps (dashed, right).
Detail
Nothing selected yet — pick a node on the map, or let your agent drive
(focus_node). Whatever either of you opens, both of you see.
Audit ledger
Append-only record of everything done to this map in your session — by agents and by you. Pure reads are not ledgered; anything that changes what is on screen is.
Tool console
Drive the map’s tools by hand (no agent required)
The same ten tools an agent sees through WebMCP — same schemas, same effects on the map, same ledger. It exists so that nothing an agent can do here is hidden from you.
Notes on what is computed and what is not
The synthesis is a random-effects model (REML by default; DerSimonian–Laird and fixed-effect through
synthesize). For this fixture, selected numerical outputs reproduce R metafor reference values
(the REML fit and the capped-weeks meta-regression) to the tested precision. The coverage gap is the
largest interior gap between observed values of prior contact, and the zero-crossing quoted on its card comes from a
live meta-regression fit on min(weeks, 3): neither that band nor that crossing point is typed in — change the records
and both move. Two things on this page are authored rather than computed, and are labelled as such: the
legend’s week bands (≤ 1 / 2–7 / ≥ 17) are a display choice, and the cap at 3 weeks is a modelling choice about
functional form, not a fitted parameter.
The study brief deliberately stops short of a sample size: the design inputs that would justify one (an equivalence
margin, a design effect, an allocation ratio, attrition, α and power) do not exist in this corpus, and a pooled τ² is
not an outcome variance. Subgroup and moderator statements are observational comparisons across randomized
experiments and can be confounded by other study features.