About & methodology · updated September 11, 2026
A research tool you can question.
Patefacta brings Canadian drug records and their source links into one research workflow. Every result has a scope, a date and limits.
Created and maintained by Diego as an independent student research project. It is an ongoing prototype, with no claimed endorsement by Health Canada, a university or a pharmaceutical company. This site does not offer an investment product or claim enterprise service guarantees.
The aim is to make evidence easier to trace: identify a product, inspect its regulatory sources, distinguish a study reference from a verified product–study link, and retain a reproducible research record.
AI assistance & accountability
How AI is used here.
AI assistants helped develop code, prepare evidence packets, inspect source fields and run automated browser checks. The current product–study matcher runs explicit, versioned parsing and identity rules; it does not ask a generative model to invent a relationship at search time.
AI-prepared evidence is a proposal. It is not independent human review, clinical validation or a regulatory-compliance determination. The exact-product approval workflow requires an authorized human to inspect the source passages, product identity and study scope and record a decision. A second AI assessment does not satisfy that requirement.
Source-reported facts, automated interpretations and human decisions remain separate. A checksum verifies retained bytes; a source link allows inspection; neither proves that an interpretation is correct. Unsupported dates, product relationships and formulation assumptions remain unresolved.
Diego maintains the project and its published corrections. Report a source-linked correction and retain the receipt to follow its review. Submission does not automatically change a record, and no response-time guarantee is claimed.
What the checks establish
Four different questions. Four different standards.
| Question | Current evidence | Limit |
|---|---|---|
| Is the catalogue structurally consistent? | All 58,034 frozen records checked on September 10: unique product codes, identifier syntax, required field presence and official product/status URL binding. Zero failures under these defined checks. | A valid link or nonblank field does not prove current or scientific accuracy. |
| Does the snapshot agree with current sources? | The September 10 stratified check compared 50 records with 250 complete official API responses: 394 of 400 comparable field checks agreed. Five products had differences; DIN ownership was not assessed by these API responses. | This deliberately disproportionate sample does not estimate catalogue-wide accuracy. The differences affect four DINs and one ingredient-name description; the 40 mg amount is unchanged. A source difference is not a finding of regulatory noncompliance. |
| How much regulatory history is available? | The 2026-09-10 checkpoint acquired 671 modern documents at 270 of 270 index positions. Regulatory index references were found for 11,432 of 14,699 marketed or approved human products. | Index availability is separate from full-document acquisition and trial-link verification. Coverage is incomplete. |
| Does a trial support this exact product? | Exact product identity, formulation, source excerpts and study/arm evidence are evaluated separately. Ingredient-only and unresolved candidates remain visibly limited. | An official study reference does not establish formulation equivalence, comparative efficacy or an approved exact-product link. Independent review remains required. |
Download the structural census · JSON. Catalogue snapshot: August 16, 2026. The report includes the input checksum, script checksum, release ID, field counts and scope.
Download the September 10 source-fidelity audit · JSON. It includes the fixed sample method, field denominators, five differences, official URLs, retrieval dates and response checksums. The targeted bulk-file follow-up retains exact raw-row locators and confirms that the five snapshot values also occur in the currently served official extracts; the pantoprazole 40 mg amount is unchanged.
Download the regulatory-history coverage · JSON. The verified normalized records include 468 records with explicit DINs and 201 records whose designated DIN field reports not applicable. 31 records without decision dates have separately recorded cancellation dates. Original missing-field states remain in the report, alongside separately identified derived corrections.
Three SBD records have an empty parser-v9 authorization field. A September 10 correction supplies their original NOC dates from three corroborating passages per record, only for the exact retained source hashes. Original bytes and authorization conditions are preserved. This does not establish a new approval or a verified trial link. Download the dated SBD date corrections · JSON.
The complete acquired-record audit also supports two DIN-field corrections, two decision-date corrections and four cancellation-date corrections. Each is bound to the exact source version and preserves its original values. Xeljanz's malformed DIN string is repaired only for that captured record with corroborating product rows; the Rybrevant and Rybrevant SC product groups remain separately identified. Download the additional field corrections · JSON.
The captured Yesafili and Epsolay reviews have blank decision-date values. Those dates remain not reported. Cancellation records retain their separate event type and any preceding regulatory actions described by the source.
Why some values remain unresolved
The full catalogue contains 101 explicit “Not Applicable/non applicable” DIN fields, all radiopharmaceuticals. There are 108 repeated numeric DIN groups across the full catalogue. These snapshot values are retained; product code is the record key. Four sampled radiopharmaceuticals have explicit numeric DINs in the current official API. Their source records are linked in the dated audit below; the government bulk extracts listed September 2 also retain those four not-applicable values. The dated bulk and API representations disagree; the original August raw inputs remain unavailable. Within the 14,699 marketed/approved human rows, there are no nonnumeric or repeated DINs. Two of those DINs also appear on historical cancelled records elsewhere in the catalogue.
The dedicated DIN-owner column is empty in this snapshot. Associated company names are available, but a company name alone does not prove DIN ownership. The explorer labels its fallback “Company (snapshot).” A targeted follow-up found an explicit DIN_OWNER field in the current government bulk rows for the five discrepant records. That separately dated source role does not establish ownership throughout the catalogue or reconstruct the original snapshot inputs.
A separate replay of 103 modern regulatory documents identified one recoverable DIN parser omission and 31 documents whose designated DIN field explicitly says not applicable. The published document correction recovers Tremfya DIN 02570890 from its designated source field. That DIN is absent from the approved August 16 catalogue, so this correction does not add a searchable product or a trial link. The explicit not-applicable fields remain unlinked; they are not treated as missing parser output. Read the official Tremfya decision or download the correction record.
How evidence moves
Preserve the source. Review the conclusion.
- Acquire. Newly acquired regulatory-source records are retained with retrieval time, source URL and checksums. Modern document acquisition uses a resumable queue and exclusive work leases.
- Normalize. Versioned parsers extract designated fields. For these newly acquired records, original bytes remain separate from the interpretation so parser changes can be replayed. The original August catalogue source inputs remain unavailable.
- Check identity. A matching ingredient or brand name alone cannot establish the same formulation. Ambiguous DIN fields, missing strength/route evidence and unsupported kit relationships remain unresolved.
- Review. Candidate evidence retains its source locators and decision history. Importing a packet does not approve a product–study relationship.
- Publish within scope. Public records expose supported fields, uncertainty and official links. Restricted source text remains separate from public summaries.
Acquisition and clinical evidence observations run on demand. Patefacta does not currently provide unattended clinical monitoring or automated regulatory-compliance determinations. A failed fetch, unavailable document and explicit source “not applicable” are different states.
The active release is verified before each search. A private, checksum-verified index avoids repeatedly parsing the full catalogue; bounded caches reuse results and visited study pages. The index must match the approved release. These controls do not establish large-scale capacity or remove initial connection delays.
Updated September 12, 2026
Results, with their denominators.
Formulation challenge and search fixes
A targeted eight-pair challenge added five source-complete formulation descriptions and three naturally retrieved near-matches. Matcher 8 surfaced one of the five; matcher 9 surfaces three after a narrow repair for explicit ingredient–strength–oral-tablet names. All three near-matches remain unpromoted. Of the two source-complete descriptions still withheld, one is an observational cohort outside the formulation tier’s eligibility rules and one has ambiguous acronym ordering. Neither is silently approved.
These provisional AI labels were frozen before the first run. The repeat run used the same cases to check a repair, so it is not independent validation. No human accuracy, clinical exposure, or registry search-completeness estimate follows from these counts. Original results · Repair results · Cases and method.
Search now permits typing while the release check runs, focuses results after they appear, offers a retry for a failed release check and stops a stalled search after 20 seconds. Reopening the selected drug preserves its loaded details. Study evidence mounts on first expansion and remains available on reopening. Independent storage reads overlap, while every release and closing artifact check remains in place. These changes do not establish a new production latency target.
Eight local browser checks covered these interactions, correction receipt recovery and desktop layout. Actual mobile/Safari sessions, human usability work and production correction submission remain unverified. Browser results and limits. Separately, recent production records confirm successful search and correction-page responses; their server timings exclude the network and browser. Bounded server observations.
A hash-bound Yesafili note connects the same DIN and original submission to the supporting SBD’s reported NOC event, retaining the RDS’s blank date. Epsolay now has a separately sourced NOC date of 27 August 2025 from Health Canada NOC 35909, explicitly naming DIN 02560720. Its RDS date stays blank. The renewed Adheroza source assessment clarifies packaging versus dosage form but does not resolve the 150 mg trial-vial to 420 mg Canadian-kit bridge. History source note · Adheroza source assessment.
Curator controls now support explicitly selected legacy review bodies, up to 50 per selection, with immutable selection receipts, separate resume/checkpoint steps and preservation of already verified or failed records. The original modern portal checkpoint remains 671 acquired reviews. A separate September 12 recovery captured all 1,704 frozen-index legacy narratives from the official bulk APIs and attempted every individual source. Of these, 1,697 individual responses passed identity checks and seven returned another record. Their matched bulk narratives remain usable; wrong individual responses do not supply child tables. This is acquisition integrity, not a clinical accuracy estimate. Complete current portal inventory reconciliation remains open.
The preceding update passed type checking, the verified build and all 358 automated tests, including new capsule, administration-negation, source-supplement and legacy-queue regression checks. Verification scope.
Legacy recovery and readiness checks
Read the completed acquisition, measurements and remaining independent checks.
The retained API representations supply 616 milestone tables and 262 nonempty later-activity tables; 354 further activity tables are explicitly empty. Six wrong-response cases also have separately captured, identity-verified portal facts. Three supply the remaining SBD milestone tables (26 dated milestones), bringing milestone availability to all 619 legacy SBD records; those portal versions also supply 29 activity rows. The explicit Maviret portal DIN 02467550 is retained separately from the malformed original bulk field 2467550. Portal and bulk dates keep their own source versions. Separately pinned notes retain the original NOC, SBD issue and PAAT update dates for Rosiver, Maviret and Mektovi; these labels are not interchangeable. Narrative fields omitted by the previous RDS display are now included. Field origins, original acquisition times and hashes are retained. The archive is activated only after all records and complete sources are verified in storage. Current archive publication state · Every-record identity verification.
Epsolay’s NOC date evidence · Adheroza’s regulatory citations and remaining presentation gap. Independent human approval is still required for 25 prepared packets; the two exact-presentation Adheroza packets remain incomplete. An AI source assessment is not recorded as human approval. The independent review and real-device task protocol states how the remaining checks can be completed and measured.
Three local responsive task checks exercised 320, 390 and 768 CSS-pixel layouts, including search, product selection and the Epsolay note during a live-source failure. These were Chromium iframe checks, not real phones or participant testing. Tasks and limits. A bounded production measurement used 18 serial authenticated requests, with connection time included and server work recorded separately. Production baseline · Deployed search measurements. Search can now be submitted as soon as the page is interactive, while the initial status lookup is pending; every search still verifies the approved release on the server. No large-scale capacity or clinical accuracy claim follows.
September 11 baseline
A separate evaluation used 16 product–study pairs across four products: 12 newly retrieved studies and four deliberately mismatched product controls. Selection and AI source assessments were frozen before running matcher version 8. Six pairs supported a qualified ingredient link and all six were surfaced. The remaining ten were withheld, including four mismatched controls and six relevant records without sufficiently clear intervention links. No exact-product or complete-formulation links were promoted.
This is agreement with provisional AI source assessments, not measured clinical accuracy. Human labels are still pending. The small, deliberately selected sample cannot estimate catalogue-wide accuracy or registry search recall; it also contains no positive full-formulation cases. Existing source-level relevance can remain visible even when a qualified link cannot be established.
Download the 16-pair results · Read the source assessments and remaining gaps · Read the selection method.
The September 11 recheck of all 27 review packets agreed with the retained September 10 queue readbacks: 25 await independent human approval; two Adheroza 420 mg cases still need a source-supported bridge from the trial presentation to the Canadian kit. This was an AI-assisted recheck of retained evidence, not a fresh private-queue query or a human approval. Download the packet recheck.
| September 11 API check · deployed version 72 | Observed result | Scope |
|---|---|---|
| All requests | 18 of 18 returned correct expected responses; zero failures and no retries. | PLAN B and levonorgestrel queries; at most three simultaneous requests. |
| Reused connections | 15 requests: median 0.452 s; slowest 1.688 s. | Catalogue and saved-source study responses. No page-render timing. |
| New connections | 3 requests: median 10.897 s; slowest 10.907 s. | Connection setup contributes substantially. New connection does not mean empty server cache. |
These are bounded client measurements from one environment, including network/proxy time. Cold source acquisition, real-user page speed and larger-scale capacity remain unmeasured. A small-sample p95 equals the observed maximum here. Download methods and timings · Read every request observation.
The September 11 release passed 356 automated checks. Nine local browser checks covered form validation, receipt/status handling, keyboard navigation, desktop layout and reviewer access. These verify tested behavior only. Read the browser-check scope and limitations.
A human usability pilot has not been performed. The prepared protocol asks participants to identify an exact DIN, distinguish an ingredient link from a verified product link, interpret missing history and submit a correction. Download the independent-review and usability protocol.
Earlier September 10 measurements, retained for comparison
The fixed source sample had 394 agreements in 400 comparable field checks; 45 of 50 products agreed on all eight fields. All 27 revised study-evidence packets were imported and read back. Of these, 25 await independent human approval and two need more presentation evidence. No human approvals or trial-link precision estimate are claimed.
Eleven agent browser functional checks passed, covering search, pagination, exports, source notices, keyboard behavior, narrow-screen layout and failure states. A human usability pilot has not been performed.
After a concurrency repair, a bounded production recheck of version 67 returned ten correct responses, including five concurrent searches. Client waits ranged from 5.7 to 18.1 seconds. Earlier failed attempts are retained in the measurement summary; latency and larger-scale capacity remain unresolved.
Download the dated measurement summary · JSON. It separates sample agreement, review status, functional checks and performance, with methods, denominators and limits.
Reproduce & correct
Keep the exact question with the answer.
After searching, choose Download research record · JSON. The file contains the displayed page, its submitted filters and official links. Editing the search box does not change the query attached to existing results. Pagination uses the submitted query.
| Field | Meaning |
|---|---|
release | Frozen catalogue release ID, snapshot time, attribution and licence links. |
query | Submitted text, search mode, market status, page and page size. |
rows / rows_in_this_file | Only the displayed page, up to 50 records. Each includes product code, DIN, recorded ingredients/strengths, company, form/route, status and official source links. |
total_matches / all_matches_in_this_file | The query count and an explicit indication of whether this file includes every match. |
captured_at_utc / exported_at_utc | When the page was received and when this file was downloaded. Neither date makes the original source newer. |
reproduction_url | Reruns the query against the active release. Check the returned release ID; later releases can produce different results. |
csv_export | The separate CSV cap and whether the query would exceed it. Downloading JSON does not create a CSV. |
correction_template | An unsubmitted draft for the exact product, field, observed value, source URL, date and explanation. |
Filtered CSV exports remain limited to 500 rows. Result CSV links carry the displayed release ID; if the active release changes, search again before exporting.
To document a possible correction, keep the original research file and use the exact official product or document link. Do not overwrite an old observation with a newer company name. The research download itself submits nothing. Use Report a correction to submit, then download the private receipt to check its status and decision history. Reports and decisions are retained separately from source records. Response timing is not guaranteed.
Sources & reuse
Official links remain the point of reference.
- Health Canada Drug Product Database: official product records and monograph access. Health Canada describes its database as updated nightly; Patefacta serves a separately dated snapshot.
- DPD open-data catalogue and Open Government Licence – Canada: source-specific attribution and reuse information for the DPD extract.
- Drug and Health Product Portal: regulatory decision documents. Each acquired record keeps its own source link and retrieval state.
- ClinicalTrials.gov: trial registration, reported study design and available results. Registry inclusion alone does not prove an exact marketed-product relationship.
Public availability does not grant blanket reuse rights for every linked document. Patefacta’s approved DPD extract and linked product monographs have different distribution scopes. Full monographs and restricted evidence are not included in public research downloads.
Contains modified and reformatted Health Canada information licensed under the Open Government Licence – Canada. Health Canada and the Government of Canada do not endorse Patefacta. For research and information; verify current details with official sources.