The pipeline behind every grade.

What this is, and what is new.

A rule-based, verification-gated pipeline that grades the evidence that a longevity intervention engages the biology of aging, one hallmark at a time. Language-model agents read, extract and propose mechanism scores under written rules; deterministic code computes the evidence scores, every rating and every displayed rating word; every cited study is verified against PubMed on entry. It runs on 107 interventions today.

Each intervention gets a score on each of 13 hallmarks of aging from two axes that are multiplied: a mechanism score anchored to a curated pathway map, and a replication-gated evidence score built only from studies that found an effect. Studies that found no effect can hold a rating down but never add to it. Funding, circulating claims, safety and absorption are reported beside the score on separate rails. The contribution is a methods-and-application framework; its parts build on established techniques (controlled-vocabulary curation, literature triage, GRADE-adjacent appraisal, funding extraction, workflow provenance). What we claim is the integration, at scale:

1. (i)Three parallel controlled vocabularies (biomarker, disease, pathway) route any study to one 13-hallmark schema by independent evidence signatures, combined in one deterministic score.
2. (ii)An evidence score with explicit null handling: one vote per trial, a like-for-like Mixed-results hold, and a rule that withholds credit for unstated design facts without asserting them.
3. (iii)A conflict-of-interest layer that checks every human study in up to twelve places and cannot publish a label, a conflict or independence, that it cannot trace to a field or a verbatim quote in the source.
4. (iv)An operating model: writer and verifier in different sessions, independent mechanism readers, fresh-session audits with a separate review seat, and every method change shipped as a dated, approved release.

Seven rules the rest follows from.

Each principle names how it is enforced: code a script exits or recomputes, procedure a run-sheet step or a fresh checker carries it, or both.

Aging biology as a fixed set of thirteen targets.

The hallmarks of aging are a consensus set of processes whose disruption contributes to aging and whose modulation can in principle slow it. Here they serve as a fixed, low-dimensional target schema: every finding, whatever its endpoint (a methylation clock, a lipid panel, a hazard ratio), is routed onto these thirteen cells. That shared coordinate system is what lets heterogeneous evidence be combined. Hover or tap a hallmark.

Four hallmarks marked frontier (H2, H5, H9, H13) are left out of the breadth count in the stack grade (section 11), because today’s evidence cannot reliably move them across a routine. They are rated like any other hallmark.

Two sessions, a gate on every step.

Every new intervention runs in two sessions, so the session that writes the research never verifies it. The research session dispatches one subagent at a time and stops. A separate session verifies, maps, scores and builds. Runs are unattended: judgment calls take the conservative option, record both sides, and are reviewed after the run. Tap any step for what it does, its gate, how that gate is enforced and the failure it exists to prevent. Most of those failures happened once; the gate came after.

Where the models work, and what checks them.

The system is best described as verification-gated, human-supervised automation. Language-model subagents do the reading-heavy work, including proposing mechanism scores and pooled-analysis flags under written rules. Deterministic code computes the evidence scores, ratings, trial flags and rating words. A person approves rules, never an individual score. In unattended runs subagents read and write files only; the orchestrator fetches PubMed records and web pages into dated files that subagents cite from.

Failure modes designed against

Three independent routes to the same hallmark.

These are expert-curated controlled-vocabulary mapping tables, not ontologies in the OBO or OWL sense: no formal axioms, no reasoner, and not yet aligned to GO, MONDO, DOID or Reactome identifiers (named work in section 15). Their useful property is parallelism: a study can reach a hallmark through a measured biomarker, a disease endpoint or a documented pathway engagement.

How one paper becomes part of a rating.

Follow one randomized trial from retrieval to a displayed word. Step through the stations; the funding lane below runs on the same paper and never touches the score.

Replication-gated, positive studies only.

The evidence-quality score (EQ, 0 to 5) is set by how many qualifying studies replicate, not by the single best one. EQ 1 and 2 are preclinical; human evidence begins at 3. Each study is first placed on a seven-level ladder by design.

A ladder of engagement, anchored to a map.

The mechanism score (MS, 0 to 5) asks whether there is a documented biological route from the intervention to each hallmark, using the most direct pathway when several exist. Every MS of 3 or more must name a pathway on the 52-row map at the claimed tier or stronger; an off-map pathway can support at most 2 until it is promoted after review, and a pathway that lists the hallmark only as a contributing one caps the score at 3. One molecule may hold MS 5 on at most two hallmarks.

A transparent function, then gates, then words.

The two axes are multiplied, so a strong mechanism with no human evidence and strong human evidence with no mechanism both stay low. The rating is then gated on study anchors, not thresholded on the product alone. Toggle between values and ratings, and between molecules and the lifestyle archetype.

IES = MS × EQ / 5            0 if MS = 0 or EQ = 0;  0 < IES < 0.5 is raised to 0.5

Strong     IES ≥ 4  and  MS ≥ 5 (MS ≥ 4 lifestyle)
           and ≥ 2 positive L4 to L7 studies (one vote per trial)  and  one with n ≥ 100
Supported  MS ≥ 4  and  IES ≥ 2  and  ≥ 1 positive L6/L7 trial vote (GREEN, YELLOW or RED)
Emerging   IES ≥ 0.5
None       IES = 0

Hold       mixedResults  =>  Strong → Supported,  Supported → Emerging   (IES unchanged)

The Mixed-results hold, as code

Good no-effect trials can hold a rating down; a pile of weak ones cannot, and a pile of weak positives cannot outvote good no-effect trials. Set the trial counts and watch the rule fire.

Every word on the page, computed

The page’s Evidence, Verdict and Independence words are computed from the data. Ratings, Mixed-results tallies and the Evidence and Verdict words are computed at deploy; the Evidence, Verdict and Independence words are computed again when the page is shown, and the shown values win. Form-dependent is the one label a writer sets.

Funding, claims, safety and absorption, on separate rails.

Funding and conflicts of interest. This is the layer most evidence summaries leave out, and the one we built to fail safe. Every human study (L3 to L7) is checked in up to twelve places, each label must survive the verify-or-drop gate, and the governing rule is to under-assess rather than falsely accuse. That rule cuts both ways: Independent needs a source just as a conflict does, and silence is reported as a disclosure state, never read as a label. Per-study labels are never read by the scoring code.

Claims. Up to six per audited page (three pages built before the cap carry more), label claims always included, the rest ranked by independent repetition. Each card reports what was tested and found, attributed to the trials, under one of six locked statuses; it never confirms, denies or grades a claim and never names who made it. A fresh checker re-opens every source before a claim ships. 35 pages carry a checked claims section today.

Safety and cautions. Ingredient-level safety from about a dozen source types (FDA, EFSA, Health Canada and TGA actions; LiverTox; upper intake limits; NIH ODS; systematic case reviews; the page’s own trial adverse events), quoted exactly from dated fetches. Single case reports and adverse-event tallies without a denominator are excluded. Cautions carry one of four severities and one to three stating sources, re-checked at Step 8.5; older pages carry an earlier format until audited.

Absorption and form. A GREEN, YELLOW or RED absorption flag (shown as High, Moderate, Low) with the main barrier named, and a product-form note where the source decides the outcome. Absorption never enters an intervention’s score; it enters only the stack engine.

Breadth times quality, minus redundancy, gated on exercise.

The member’s Longevity Intelligence Report grades a whole routine in five words (Exemplary, Solid, Mixed, Thin, Speculative), with no number shown. The grade reads each item’s per-hallmark rating and applies the rules below, in order; caps only lower a grade. Size is neutral: the penalties are shares of the stack.

grade(stack):
  if items < 2                      -> no grade
  if no Strong rating anywhere      -> Speculative if nothing Strong/Supported, else Thin
  clean = redundancy < 0.12 and pileMax < 3 and conflict < 0.30
          and redAbsorb < 3 and redAbsorb/items < 0.25 and strongBeyondExercise >= 3
  if breadth >= 0.8 and q >= 0.65 and clean   -> Exemplary
  elif breadth >= 0.4 and q >= 0.45           -> Solid
  else                                        -> Mixed
  cap at Mixed if redundancy >= 0.30 or pileMax >= 4
  cap at Mixed if no exercise

Seventeen gates, and how each one is enforced.

Software-style quality assurance applied to evidence synthesis. Code gates halt the step that runs them. Several gates are scripts a person runs before deploy, and stop the run by procedure. Some checks are warnings by design. Tap a gate.

Every page re-audited by a session that did not build it.

The audit program brings each deployed page to the current bar without a fresh deep-research pass. One page per fresh session; a separate review session checks each finished audit; a page whose problems are too broad to patch gets a full re-run instead. As of 22 October 2026, 35 of 107 pages were done; the live count of pages revised since their first build is 37.

1. 1TriagePage checker, synonyms, legacy test. A pre-schema page stops here and is queued for a full re-run (major version).
2. 1.5Data integritySizes, durations and blinding re-decided from the live abstract; one trial key per trial; within-group positives nulled; comparators checked; registry-industry trials relabeled; out-of-scope studies dropped.
3. 2Missed studiesFixed searches minus the catalog. Small literature (40 or fewer human trials): add every in-scope study. Large: up to 20, at least 5 to large no-effect trials when the pool has them.
4. 3AdditionsNew studies run through the same steps: cache, relevance, tier, router, mapping, summaries, funding.
5. 4Mechanism checkA fresh checker re-verifies every MS 4 and 5 from a dated PubMed packet and can only confirm or lower.
6. 5Claims and cautionsClaims research and its checker; cautions and their checker.
7. 6RewriteOne writer and a fresh fact-checker.
8. 7ShipVersion bump, validator and page checker clean, every rating move explained in a dry run, deploy, live check, run record.

Method change log

Novel integration, not novel parts.

Each capability has prior art. Network-medicine methods map drug targets to hallmark gene modules. Automated tools extract risk-of-bias judgments and track trials for systematic reviews. Classifiers detect industry ties in the literature. Curated aging resources (the HAGR family, HALD, Aging Atlas) catalog genes, compounds and associations. What we did not find is one system that combines hallmark-mapped, replication-graded clinical evidence with a per-study verify-or-drop funding layer inside one provenance-tracked pipeline.

What is not yet done.

The honest gap is validation. Everything above shows the pipeline is self-consistent and traceable; whether its mappings and scores are correct is a separate question, not yet answered empirically. Three studies would answer it.

Known gaps in the running system

The plain-language version is How the grading works. The source of record is the canonical pipeline (BOOTSTRAP, AUDIT, the STEP and REF documents, the research prompts and the Pathway Map), the scoring specification, and the scoring code itself, which governs where a document and the code disagree.