Methodology

ReaDepth is a structured, public dataset of life-science and diagnostics labs. Every lab in the launch cohort has a public profile page — a digital twin — drafted from public sources and refined by people affiliated with the lab. Pages cover identity, instrument fleet, assay portfolio, software stack, procurement signals, hiring and install events, and a cohort-relative score (Frontier) that places the lab against its peers on a single axis of technological currency.

The model is closer to Wikipedia than to a vendor CRM. Read access is anonymous and indexable; editing requires an account and a verified affiliation; every fact carries a source URL or an excerpt; every change is diff-able and revertable. The platform is public-by-default, which is what makes it a profile platform rather than a surveillance product.

The unit of curation is the physical lab — a coherent equipment fleet under a single director with a defined budget. What counts as a “lab” varies by archetype, because real fleet ownership varies by archetype:

In NCI cancer centers and AMS, equipment lives at shared cores; PIs use the cores’ fleets and rarely own significant equipment of their own. Modeling each PI as a separate lab would generate hundreds of mostly-empty profiles. We don’t. Cores are labs; PIs are metadata. Where a PI does own equipment (the Fischer-style outlier — a PI with several mass specs of their own), the instruments are allocated to the relevant core with an attributed_pi annotation that displays as a label (“Orbitrap Eclipse — owned by Fischer Lab”) but counts toward the core’s fleet.

Each PI is bound to exactly one primary core to avoid double-counting in funding, output, and Frontier-velocity aggregates. Secondary affiliations are display-only.

Every fact on every twin starts with a guess from a confidence-ranked source chain. The contributor’s job is ratification, not authorship. The chain, in order:

1. Verified contribution from this lab. A claim made by a contributor with active, tier-1-to-5 affiliation to this lab.
2. This contributor’s prior submissions. Earlier statements by the same contributor about adjacent facts.
3. Direct public crawl. Vendor press releases, conference posters, grant abstracts, lab websites, NIH RePORTER, NCI P30 facility lists, CLIA Lab Lookup.
4. Public-source LLM inference. Cached prompts over the same public corpus, used to fill in plausible facts where the crawl is silent.
5. Cohort prior. The median lab in the same archetype × geography × size band cohort.
6. Empty. Last resort, displayed as a faded slot with a “help fill this” affordance.

Each fact is tagged with its source tier and a confidence weight. Confidence is rendered visually: solid type for verified, italic for inferred, faded for stale (older than the fact-class’s freshness window — six months for fleet status, two years for assay menu, etc.). The /wiki/lab/[lab_id]/methodology sub-page lists every source URL and excerpt used to draft the displayed twin.

Public sources are mostly indexed at the organisation level; the resolution layer reattaches org-level claims to specific labs using PI-to-primary-core mapping, departmental context, conference-poster co-authorship, and cohort priors. Claims that can’t be resolved sit pending in a “facts pending lab attribution” section on the org page until a contributor disambiguates.

ReaDepth borrows its editing model from Wikipedia, not Zillow. No single person knows everything about a lab. PIs don’t always know day-to-day instrument status; lab managers don’t always know the full assay menu; bench scientists don’t always know procurement contracts. Forcing edits through one person produces stale data and a bottleneck.

Three governance rules, in tension: public read by default, editing requires an account, and anyone with verified affiliation can edit, with reputation tiers. The tiers determine source weight in confidence-weighted resolution, not who is allowed to edit:

When two affiliated contributors disagree on a fact, the resolver flags it as contested; both versions surface on the lab page with a “contested” badge, and a talk-page thread auto-opens at /wiki/lab/[lab_id]/talk. The higher-tier contributor’s edit shows as the displayed default but does not suppress the disagreement. Discussion happens on the talk page; resolution updates the underlying claim and closes the thread. If a thread sits unresolved for 30 days, ops triage.

Per-contributor reputation accumulates across all labs a contributor edits. Corroborated edits raise a reputation multiplier (range 0.5×–1.5×); overturned edits lower it. The displayed source weight is tier_weight × reputation_multiplier. Sustained false claims drop weight; persistent abuse loses verified-affiliation status.

Frontier is a single quantitative score per lab, measured in bits, that answers a question PIs ask but can’t compute from inside their own lab: am I keeping up with my peers? It is information-theoretic by design. The unit reads as legitimate methodology to anyone fluent in Shannon information; the output is hard to game because the cohort distribution is the anchor.

For each fact F (instrument model, assay, software, primary consumable vendor) attached to lab L:

rarity(F)       = -log₂ p(F is in a comparable cohort lab right now)
velocity(F)     = d/dt p(F in cohort)        ← linear regression slope over the
                                               last 4–8 quarters
contribution(F) = rarity(F) × velocity(F)    ← velocity bounded to avoid
                                               pathological behaviour

Frontier(L) =  Σ contribution(F) over facts L has
            −  Σ contribution(F) over high-velocity facts L lacks

Direction is built in. Rare-and-rising tech contributes positively; rare-and-declining tech contributes near zero (low velocity); common tech contributes near zero (low rarity); high-velocity tech the lab lacks contributes a penalty. A lab leading the cohort accrues bits; a lab trailing the cohort on rising platforms loses them.

Worked example.

Take a hypothetical NCI cancer-center genomics core in mid-2026 that runs Roche SBX (sequencing-by-expansion). At that time, SBX is present in roughly 3% of cohort labs and cohort adoption is climbing at about +8 percentage points per quarter. The contribution of that single fact is:

rarity(SBX)       = -log₂(0.03)            ≈ 5.06 bits
velocity(SBX)     = +0.08 / quarter        (normalised: 1.07×)
contribution(SBX) = 5.06 × 1.07            ≈ 5.4 bits

The same core also runs an Illumina NovaSeq 6000 (present in ~78% of cohort labs, cohort velocity slightly negative as the X-series replaces it): rarity ≈ 0.36 bits, velocity weight ≈ 0.95×, contribution ≈ 0.34 bits — small but non-zero. It lacks a NovaSeq X Plus, which is in ~31% of cohort labs and rising at +6 pp/quarter: penalty ≈ −1.7 bits. Summed over its ~25 fact slots, the core lands at, say, +12.4 bits versus a cohort median of 0; that translates to a percentile rank of top 22% of NCI cancer-center genomics cores (US).

Display rule. Frontier is computed for every lab but only displayed when the lab’s primary cohort has at least 30 active members. Below that density, the rarity and velocity estimates are too noisy to be defensible, and the score reads as marketing rather than measurement. Pre-density labs see the per-fact breakdown internally; the headline percentile is suppressed until the cohort qualifies.

Time and velocity. Velocity is a cohort property, not a lab property — it is measured against the cohort as a whole over the last 4–8 quarters of resolved twin state. A lab’s own historical churn does not affect its Frontier; only the current fact set, evaluated against the current cohort distribution and its recent trajectory, does. That makes the score comparable across labs and stable under honest fleet maintenance.

Frontier is computed nightly per lab and stored in the materialised twin view. The full per-fact breakdown is expandable from the headline percentile on every lab page; cohort distributions are downloadable for replication.

A lab is a member of multiple cohorts; the resolver picks the densest one (highest member count) for Frontier computation. Bucketing axes:

• Archetype. academic_research_lab, academic_medical_system, clinical_test_vendor, service_lab, biotech_lab.
• Geography. US, EU, APAC, RoW.
• Size band. small (<10 staff), medium (10–50), large (50+).
• Research domain. oncology, immunology, neuro, agbio, infectious-disease, developmental-bio, multi-domain.

Two predefined cross-axis cohorts carry their own membership lists rather than being computed from buckets:

• NCI-designated cancer centers — membership from the NCI public list (~71 centers).
• AAMC member academic medical centers — membership from the AAMC public list.

Cohort definitions, member lists, and the version history of both are public and downloadable. If you think a lab is in the wrong cohort, you can challenge it on the lab’s talk page.

Every claim in the dataset carries a source URL or an excerpt, plus the source tier (verified contribution, prior submission, public crawl, LLM inference, cohort prior). The full source list for any twin lives at /wiki/lab/[lab_id]/methodology. The full edit history — every change, attributed and timestamped, with before/after diffs — lives at /wiki/lab/[lab_id]/history. There is no privileged channel; anyone can audit any claim back to its source.

Inference outputs are persisted with explicit source = ‘llm_inference’ and state = ‘guess’ markers; they never masquerade as direct evidence.

If you work at a lab represented on ReaDepth, you can claim and edit its twin. Account creation is light (email plus a verification click). To declare an affiliation, you supply your role at the lab; verification places you into the relevant tier:

• Email-domain match — institutional email matches the lab’s parent organisation. Default verification path; covers most academic labs.
• LinkedIn-DM verification — slower, higher-trust path used when domain match is unavailable, contested, or insufficient (e.g., when claiming Tier 1 PI/director status).
• Institution-admin override — bulk verification by an institutional admin (e.g., cancer-center IT vouching for affiliations).

Tier 1 (lab director) requires email-domain match plus role attestation. Tiers 2–3 require email-domain match. Tiers 4–5 are accepted on email-domain match alone. Tier 6 (verified-but-unaffiliated) is the default for vendor reps, peer PIs, and attached PIs whose primary core is elsewhere; their edits are accepted but require corroboration to elevate.

Corrections can be made to any fact on a lab you have affiliation to. They flow into the claims layer immediately with state = ‘confirmed’, attributed to you, and are visible in the edit history. Reverts are one click for any active affiliated contributor.

Different facts have different sensitivity. The twin layers by tier:

• Public-grade. Fleet, assay menu, archetype, PI, key personnel, software stack. Collected raw, attributed.
• Open-secret-grade. Sample-volume bucket, primary consumable vendors, hiring/layoff events. Collected raw to the contributor’s cohort; bucketed in public display.
• Contract-sensitive. Exact throughput, exact spend, contract terms, proprietary panel content. Collected only at aggregated bucket level (N≥5), with a 6–12 month time delay.
• Strictly proprietary. Revenue per assay, internal R&D programmes, unannounced products. Never collected. No tier of contributor and no payment unlocks this layer; it is out of scope by design.

For the launch cohort (academic research and cancer centers) almost everything is public-grade or open-secret-grade. Sensitive tooling — differential privacy for numerics, per-fact disclosure toggles, time-delayed aggregation — is required for a corporate cohort and is deferred until that cohort is in scope.

Contributors set a citability default for their own identified contributions: full attribution, aggregate-only (default), or confidential. Secondhand claims about other labs default to confidential and never appear with public attribution. Redaction requests are honoured without question.

Cohort definitions, the Frontier formula, source-tier weights, and the cohort distributions used in the most recent score computation are all public. Anyone with the public dataset and this page can replicate any displayed Frontier score for any lab whose cohort meets the N≥30 display threshold. We treat methodology as a load-bearing public artifact: if you can’t check our work, you shouldn’t trust the numbers.

Peer review is welcome. If you find a methodological error — cohort mis-bucketing, a velocity term that swings on a single noisy quarter, a source-weighting choice that biases a class of labs — open a talk-page thread on any affected lab, or email below.

Source coverage is biased toward English-language US/EU labs with strong public footprints; smaller, non-English, and emerging-region labs are underrepresented in the public crawl and rely more heavily on contributor verification. Frontier is gated to the N≥30 cohorts; for smaller cohorts the score is internal-only until density is sufficient. LLM inference is a fallback layer, not an evidence layer; its outputs are explicitly marked and weighted accordingly. Honest acknowledgment of these limits is part of the methodology, not a flaw to hide.

Methodology questions, peer-review notes, corrections, or redaction requests: alex@readepth.com.