For clinicians, researchers, advocates, and contributors
Held to Standard cross-references the evidence and the official standards of care behind transition-related treatment decisions, and pins every treatment profile to an archived, dated, primary source. It is an evidence-and-provenance layer, not a guideline, and explicitly not patient-directed advice. The code is open source: transparency about how treatment profiles are built is as important as transparency about the profiles themselves.
What the project asserts, and what it does not
Each evidence profile is a structured record pairing a
standard of care (what an official body actually requires or recommends)
with the evidence behind it, graded for certainty, with both sides citing
sources held in a local archive. This is the GRADE sense of the term: a paired,
sourced summary of standard and evidence. Patient-facing, the same record is called
a treatment profile; internally it is a claim YAML file.
An evidence profile describes the state of the evidence and the regulatory position. It does not issue a recommendation in its own voice. The informing-versus-advising line is the single most important boundary in the project, and the tooling actively enforces it.
This project is compliant and above-board by design, not adversarial. It respects robots.txt and rate limits, identifies itself, and treats official bodies' published positions as the standard of care it cross-references against. Where it records a gap between a standard of care and the underlying evidence, that gap is stated neutrally and with citations.
GRADE: how evidence certainty is rated
HeldToStandard uses the GRADE framework (Grading of Recommendations, Assessment, Development and Evaluations) to rate the certainty of evidence for each finding. GRADE is the international standard used by the WHO, Cochrane, and most national guideline bodies, including Socialstyrelsen.
High
We are very confident the evidence reflects the true effect. Further research is very unlikely to change the finding.
Moderate
We are moderately confident. The true effect is likely close to what the evidence shows, but there is a meaningful possibility it differs.
Low
Our confidence is limited. The true effect may be substantially different from what the evidence shows.
Very low
We have very little confidence. The true effect is likely substantially different from what the evidence shows.
GRADE certainty reflects the quality and consistency of the evidence, not whether a treatment works or is recommended. A moderate-certainty finding can still support a strong clinical recommendation if the expected benefit is large and harms are small.
Certainty is assessed per outcome, not per treatment overall.
Provenance model
Two sanctioned fetch paths are built. A third (EMA bulk download) is in progress. Each is documented in a decision record. All identify honestly and pace politely.
-
Regulator and institutional web pages
Fetched via the compliance-first crawler: robots.txt re-checked on every redirect hop, per-domain rate limiting honouring Crawl-delay, honest User-Agent, and snapshot-and-index under the final URL with a content-addressed slug so nothing silently overwrites anything. Refusals are logged as carefully as successes. A hardening pass added an SSRF address gate (redirect targets validated against resolved IPs before fetching; loopback, private-range, link-local, and cloud-metadata addresses refused), a 10 MiB streaming body-size cap, and a content-type allowlist.
-
Bibliographic evidence metadata
Fetched via an academic API layer: Europe PMC primary, PubMed E-utilities fallback, identifying with the project User-Agent and honouring each API's rate limit. Output is a BibTeX entry with provenance fields (API, identifier queried, retrieval date) so a reviewer can re-fetch and confirm. This is a deliberate exception to the web crawler, because academic APIs are not crawlable pages and their output is a citable record, not a snapshot.
-
EMA bulk dataset download (in progress)
Periodic bulk pulls with a per-dataset license-first gate. No build proceeds until reuse terms are confirmed for each dataset.
A separate integrity layer (retraction status via Retraction Watch and Crossref, venue indexing via PubMed, Europe PMC, and DOAJ, and citation-resolution fraction) is designed and in progress. Signals will be reviewer-facing and inform the GRADE call. A retracted source on a reviewed or published treatment profile will be a hard validation error. AI-text detection is deliberately excluded.
A reviewer can trace any treatment profile: regulator-side via the page archive, evidence-side via the API-stamped BibTeX entry. Once the integrity layer is built, integrity signals will also be independently re-fetchable. The chain runs back to what was published, by whom, on what date, as it actually appeared.
No silent failures.
A structural validation layer runs both standalone and as a git pre-commit hook. It is severity-aware.
On a reviewed or published evidence profile, a missing citation key, an empty source list, an invalid GRADE tier, an unknown regulator, a retracted source, or advice-voiced language is a hard error, not a warning. The hook validates the staged content (exactly what gets committed) and gates draft profiles off the main branch.
The goal is to make silent failure structurally difficult. If an evidence profile cannot show its sources, its provenance, its evidence assessment, and its boundary against advice, it does not get published.
Not yet public. Contributions welcome.
Pre-MVP. Sweden first, with a Scandinavia rollout planned. The crawler, compliance layer, academic ingestion, structural validator, and git pre-commit hook are all built and tested. Source integrity vetting and EMA bulk download are designed and in progress. Ingestion of regulator and evidence sources is paused pending legal counsel confirmation of the redistribution model.
Contributing to the tooling
The repo is private and pre-MVP. Tooling contributions (crawler, validators, supporting scripts) are welcome, with the same bar applied internally. Changes to the validation layer should come with decision-record reasoning and the tests that prove the behaviour.
Evidence profiles are the integrity-critical content the tooling exists to protect. Contributing to evidence profiles requires an established level of trust that cannot yet be earned through a pull request alone. The bar will be named explicitly when the project is ready to accept them.
This project is built openly, under my own name, out of lived experience and prior advocacy work. More on the motivation is on the About page.