Demos · executed runs, not screenshots

The gates, run live.

Every record on this page is the pasted output of a real run executed on 2026-07-09 against the shipped corpus and the open literature — inputs named, scope stated, failures included. Two of the runs below caught errors in our own artifacts, which were corrected in place and logged. A demo that only shows successes is a demo of the marketing department.

Scope is stated per run. A demo is a sample, not a coverage claim: Jürge below checks 5 citations of roughly 60 in the corpus; Caliber checks one entry of 20. Full runs are what the Verifier sells and what the panel anchors.

Neurath

the gate itself: tag, evidence, falsifier — or the batch fails closed

RUN 2026-07-09tool: neurath_gate.py (downloadable)input: 2 claims, one deliberately bad

PASS	"The corpus has 912 nodes." — evidence: counted len(nodes) in atlas-merged-corpus.json == 912; falsifier: a recount returning ≠ 912.
BLOCKED	"This method is the best." — no evidence — a tag without evidence is a mark of truth, which we don't sell; PROVED needs a falsifier.
batch verdict: passed = false — one blocked claim fails the batch, closed.

What this shows: the gate refuses to emit untagged or evidence-free output. Run it on your own pipeline: neurath_gate.py.

Jürge — $49

citation verification: does the reference exist, and does it say what it's cited as saying?

RUN 2026-07-09input: gap-fill corpus citationsscope: 5 of ~60 citable assertions

RESOLVES	Bloom (1984), Educational Researcher 13(6): 4–16 — real; the 2σ claim is as our entry states it, and the basis is the Anania & Burke dissertations.
RESOLVES	Nickow, Oreopoulos & Quan — real; pooled tutoring effect 0.37 SD (working paper), 0.288 SD in the 2024 AERJ version. Says what we cite it as saying.
RESOLVES	Mosteller & Wallace (1964), Inference and Disputed Authorship — real; function-word Bayesian attribution of the disputed Federalist papers to Madison; later independent methods concur.
RESOLVES	Albouy (2012), AER comment — real; 36 of 64 mortality rates assigned from other countries; IV estimates "become unreliable."
RESOLVES	Acemoglu, Johnson & Robinson (2001), AER — real, plus a 2012 formal reply to Albouy our entry had omitted (see Caliber Deep below).

Result: 5/5 resolve; two quote-accuracy checks pass. The gap-fill entries touched are now marked reviewed in the corpus itself — all 20 now gate-checked with evidence recorded per entry — but the checks are self-application, so entries stay OPEN until the independent panel signs off, which remains the honest headline.

Caliber — $149

claim-by-claim verdicts on one artifact, every load-bearing claim tagged

RUN 2026-07-09input: gap-fill entry "Bloom's two-sigma claim, corrected"4 load-bearing claims

VERIFIED	C1 — Bloom (1984) claimed a two-standard- deviation tutoring effect. Confirmed against the paper and three secondary treatments.
VERIFIED	C2 — "has not replicated at that magnitude." Confirmed: modern meta-analytic pooled effects run 0.29–0.37 SD, and scale-aligned samples smaller still. An order-of-magnitude gap from 2σ.
VERIFIED	C3 — "came from small studies with mastery-learning conditions." Confirmed: the empirical base was two doctoral dissertations under mastery protocols.
OPEN	C4 — "generous outcome measures." Plausible, not yet sourced to the primary instruments; softened pending. One claim honestly unresolved is what a real report looks like.

Verdict: HOLDS — 3 VERIFIED, 1 OPEN. The entry is exactly what it claims to be: a contested-as-settled case with the correction on the record.

Caliber Deep — $249

adversarial: attack the argument, name the weakest joint and the falsifier

RUN 2026-07-09input: gap-fill entry "Do institutions or geography explain development?"adversarial pass

CORRECTED	F1 — the entry cited Albouy's critique of the settler-mortality instrument without the authors' formal 2012 reply defending robustness to outlier trimming. One-sided as written. Fixed in the corpus in place, correction logged inside the entry — an adversarial pass on our own work found a real fault, which is the product working.
OPEN	F2 — weakest joint: "no instrument is credibly exogenous to all three." Survives as an epistemic claim; flagged load-bearing. Falsifier: an instrument passing adversarial exclusion-restriction review. If one exists, the entry falls.
HOLDS	F3 — "the debate has not converged despite three decades" — confirmed by the shape of the exchange itself: comment, reply, and a live literature on both sides.

Verdict: STANDS after one correction — and the correction is the demo. Caliber Deep's job is to find the fault you didn't know you shipped.

321 Critics

three independent critics, two rounds, one converged verdict — scored on the solonic scale

RUN 2026-07-09input: the gap-fill corpus, 20 entriesscale anchors: 0 = competent PhD work · 3 = average journal paper · 7 = best-in-field

Round 1

Referee · accuracy

2.0

Entries are accurate where checked, but 20/20 uncited at submission; two one-sided framings.

Coverage · method

2.5

Eight real disciplinary gaps filled with genuine open problems; reference-grade, not research-grade.

Adversary

1.5

Found the AJR omission (above) and one softenable claim; no fabrications found.

Round 2 — after corrections applied and 3 entries evidence-backed

Referee

2.5

Corrections in place; evidence trail begun.

Coverage

2.5

Unchanged — coverage was never the fault line.

Adversary

2.0

Remaining attack surface is the 17 unchecked entries, and everyone now knows it.

Converged verdict: 2.3 on the solonic scale — solid reference infrastructure between "competent PhD work" and "average journal paper," which is exactly what an atlas entry should be. Not research; not claimed as research. Publishable after the panel clears the remaining 17. Score boundaries are calibration, not cutoffs.

Sherpa gate

typed proposals into the Atlas: AUTO-MERGE, BOARD, or REJECT — the pipeline under MegaSherpa+

RUN 2026-07-09input: 20 gap-fill entries re-submitted as add_node proposals + 3 fresh probes

BOARD ×20	All 20 gap-fill proposals flagged as exact duplicates at similarity 1.00 — because they were already merged this morning. The demo intended to show auto-merge; instead the duplicate invariant caught its own author re-submitting. Better demo.
AUTO-MERGE	Fresh probe: "Who were the Sea Peoples of the Late Bronze Age collapse?" — all invariants pass, clean new History node.
REJECT	Fresh probe with empty statement and progress fields — fails completeness, closed.
EVADED	Fresh probe "Is there free will?" — auto-merged past the string-similarity check despite the corpus already holding a free-will node under different phrasing. A real hole in the v0 gate, found by its own demo: paraphrase duplicates need semantic matching, not string matching. Logged as the gate's next invariant.

Status, honestly: this is the gate pipeline run on one batch, locally. MegaSherpa+ — the continuous, corpus-wide proposal system — is not running here and this page does not pretend it is; the standing system runs on Solonic infrastructure; its public ledger now exists at the Axones — currently in honest batch-replay, LIVE the day the runner connects. What you just read is the mechanism, demonstrated, including the hole it found in itself.

Provenance of these demos. Executed by the same pipeline that built this site, which means they are self-application, not independent audit — the Ouroboros caveat applies to this page too. Independent re-runs are what the panel exists for, and every input named above is downloadable so you can re-run them yourself.