[ METHOD ]

the evidence standard

We claim two things, and we measure both: the evidence in a Cesure review is real — every quote verbatim, every quote checked against its source by a program — and the agent that finds it actually finds the papers that matter. This page shows the pipeline, then the current numbers, then exactly how to re-measure them. Nothing here is asserted without a harness behind it.

[ THE PIPELINE ]

Agentic discovery, measured on curated benchmarks

The engine plans diverse query arms (search, citation walk, reference walk, recency), grades every candidate, and stops when coverage flattens. We don't ask you to take that on faith — §3 shows its recall on benchmark briefs with known key papers, run through the production modules (scripts/eval-discovery.ts).

Verbatim extraction with offsets

Evidence is a quote copied character-for-character out of the source, stored with its character offset and surrounding excerpt (src/lib/verbatim.ts). Nothing paraphrased is allowed to call itself a quote.

Mechanical verification, not model vouching

Each quote is re-found in the source text by offset — a code check. A quote the code can't find in the source is inadmissible and is dropped before it ever reaches a document.

A firewall inside the reconciler

Before any proposed change is returned, the firewall drops unknown ids, out-of-set papers, and non-verbatim quotes (src/lib/reviews/reconcile.ts). Fabricated proposals do not survive it; the count in §3 is post-firewall.

Independent judgment, then an audit

The strongest labels (contradicts, supersedes) get a second, independent judge pass or are downgraded. A faithfulness audit (src/lib/reviews/audit.ts) samples claim↔quote entailment with a strict judge and writes the score onto the review, tagged with the version it was computed against — stale scores never render.

[ MEASURED RESULTS ]

Faithfulness audit — every public review

SOURCED counts claims carrying at least one verbatim quote. GROUNDED is the share of a judged sample whose quotes actually entail the claim, per a strict second model. Read live from the review rows on every render.

LLM Evaluation Awareness and Strategic Behavior

10/12

SOURCED (83%)

100%

GROUNDED (10/10 judged)

FLAGGED CLAIMS

—

SOURCE-VERIFIED QUOTES

AUDITED 7/24/2026 · VERSION 1 · SAMPLE 10

Discovery recall — curated benchmarks

The production discovery modules run against briefs with hand-picked key papers. Recall is the share of those papers the engine surfaced and graded relevant. Misses are listed, not hidden.

SAE interpretability (arXiv-heavy ML)

73%

RECALL (8/11 key papers)

293

CANDIDATES SCREENED

GRADED RELEVANT

SEARCH ROUNDS

MISSED · Towards Monosemanticity: Decomposing Language Models With Dictionary Learning; Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders; Codebook Features: Sparse and Discrete Interpretability for Neural Networks

COVID-19 clinical + structural (OA biomedical)

80%

RECALL (8/10 key papers)

427

CANDIDATES SCREENED

243

GRADED RELEVANT

SEARCH ROUNDS

MISSED · Clinical features of patients infected with 2019 novel coronavirus in Wuhan, China (found but graded <2); Efficacy and Safety of the mRNA-1273 SARS-CoV-2 Vaccine

Efficient long-context LLMs (fast preprint field)

42%

RECALL (5/12 key papers)

385

CANDIDATES SCREENED

GRADED RELEVANT

SEARCH ROUNDS

MISSED · RWKV: Reinventing RNNs for the Transformer Era (found but graded <2); Retentive Network: A Successor to Transformer for Large Language Models (found but graded <2); Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models; Ring Attention with Blockwise Transformers for Near-Infinite Context (found but graded <2); LongNet: Scaling Transformers to 1,000,000,000 Tokens (found but graded <2); Gated Linear Attention Transformers with Hardware-Efficient Training; Efficient Streaming Language Models with Attention Sinks (found but graded <2)

SNAPSHOT 7/24/2026 · 4 ROUNDS · GRADER default

[ LIMITATIONS ]

Audits are sampled, not exhaustive.

At most 40 claims are judged per audit run; this page shows the real sample size next to every number. A grounded% describes that sample, not a guarantee over every claim.

The judge can be wrong.

Groundedness is an LLM's call on whether a quote entails a claim, and it can misjudge — especially on hedged or comparative claims. Judge errors are dropped from the denominator, not counted as passes.

Verified is not correct.

Mechanical verification proves the quote exists in the source at that offset. It says nothing about whether the claim's interpretation is right — that judgment rests on the audit and, ultimately, on you.

The public sample is small today.

The audit table covers every public review — currently one. The discovery benchmark is three curated briefs with hand-picked key papers: real, but small. We publish the misses, not just the rate.

Numbers age.

Every figure here carries its run date. The audit re-renders from the review row each re-audit; the discovery snapshot is refreshed by re-running the harness. If you catch a stale date, that's the truth too.

[ REPRODUCE ]

Every number on this page is the output of a harness in the repository, run against the production code paths. With the repo and keys:

# faithfulness audit (writes the badge + this table)
npx tsx scripts/faithfulness-audit.ts

# discovery recall on curated benchmarks
npx tsx scripts/eval-discovery.ts

# keyword baseline (same grader, fair comparison)
npx tsx scripts/eval-baseline.ts

# full back-test replay (real LLM spend; budgets bound it)
npx tsx scripts/replay-review.ts --asof 2025-01-01 --step-days 90 "<question>"

The method in prose, with the same honesty: how we measure trust. The benchmarked account, generated from the run artifacts: whitepaper (PDF).