Evidence register · live
What can be shown. With date and limit.
This page shows the current measurement status of DyReP. Other pages refer to it instead of stating a status themselves.
- Current
- 12
- measurement within the agreed shelf life
- Proven
- 9
- one-off trial with a source; not meant to be repeated
- Expired
- 0
- re-measurement needed
- Not yet measured
- 0
- no valid source or measurement
- Candidate
- 3
- built or prepared, but not yet confirmed as current operation
Per evidence entry
- Component
- The public name of the system component.
- Claim
- Exactly what has been shown.
- Measured on
- Date and time of the observation.
- Coverage
- For example test environment, test case file, production connection or real working situation.
- Source
- A readable source name, for example DyReP Gateway production measurement or DyReP release decision · 25 September 2026.
- Limitation
- What the measurement does not establish.
Core and execution
DyReP · Engine · Gateway · current operation · access control- source status
- live reading · valid for 24 hours
- measured on
- · 0 hours old
- coverage
- production environment · catalogue active · bindings on the expected build
- source
- DyReP Gateway production measurement
Measured on via gateway.dyrep.org/gereed: catalogue 1.27.0 active, as the only active catalogue, 17 bindings on the expected build. This is a snapshot, not a continuous reading; 24 hours after the measurement it counts as out of date.
Snapshot, not a continuous reading.
- source status
- local suite run · valid for 14 days
- measured on
- · 4 days old
- coverage
- 856 of 861 passed · 0 failed · 5 skipped · engine db29b8d (Gateway pin)
- source
- Full engine test suite (npm test) at commit db29b8d, the Gateway engine pin, 13 November 2026
A passing suite proves the tested cases, not production use.
- source status
- live in production, synthetic pilot object · one-off trial, not repeated
- measured on
- · 47 days old
- coverage
- approved knowledge → Pack → Engine → two different results from one permitted knowledge state, each with its own recorded proof of execution → read back and recomputed → a later knowledge change does not silently alter earlier results · independently reviewed
- source
- Golden Witness in production, 1 October 2026
Synthetic pilot object. Revoking a session has only been shown in a test environment, not in production. A source change does not yet signal earlier results, and the knowledge does not yet contain interdependencies.
- source status
- live, test profile · valid for 30 days
- measured on
- · 4 days old
- coverage
- reading an own case file allowed · reading a case file outside the own identity refused · in the same session, with the same identity, through the Gateway · earlier measurement 14 September: second identical request no second execution, read done in 5.7 s
- source
- Monitoring probe of access control, 13 November 2026
Synthetic test case files in a test profile. Proves these reads and the access boundary, not a complete client workflow.
Knowledge
Knowledge base · source submission · admission · approval · validity period of the judgement · change impact · snapshot- source status
- live, test and public sources · valid for 30 days
- measured on
- · 6 days old
- coverage
- file through the submission window → directly to storage, not through the model → virus scan → registered as a source, with a source hash equal to the declared hash · 30 days to 13-11-2026: 36 sources registered, 36 clean, 36 hash-equal
- source
- Control measurement of source submission, 13 November 2026
Submitting and registering is not inclusion in or approval of the knowledge. The ChatGPT model does not send files itself; the submission window does.
- source status
- live · valid for 30 days
- measured on
- · 13 days old
- coverage
- submission → seed by the operator → human review per candidate → batch confirmed → task completed with receipt · 30 days to 13-11-2026: 477 candidate decisions (51 adopted, 322 flagged, 104 rejected) · 28–29 October, two own public sources: white paper "Wat geldt er nu?" (Dutch) 105 candidates, 22 adopted; post "Bekijk deze afbeelding — kennis bewaren, kennis bewaken" (Dutch) 33 candidates, 8 adopted and then all 8 approved · all 138 candidates traceable word for word to the source
- source
- Control measurement of the seed route, 13 November 2026
Proves the route and the human review per candidate. Adopting is not yet approval of the knowledge. Traceable word for word does not mean substantively correct; the figures in the post were checked separately against the original source.
- source status
- live, test accounts · valid for 30 days
- measured on
- · 4 days old
- coverage
- a dedicated submission account only submits sources (25 submissions, 0 other tasks since the start) · seeding and reviewing is done by another account
- source
- Control measurement of accounts and tasks, 13 November 2026; knowledge intake evidence record in the canon
All accounts see the same tools; the server decides per call. Refusals for the expert account are proven in tests only. No production authority.
- source status
- live, test source · one-off trial, not repeated
- measured on
- · 50 days old
- coverage
- a registered source revoked · with time and the reason in the person's own words · the registration is kept
- source
- Control measurement of source registration, 28 September 2026
Revoking takes a source out of use; it does not correct knowledge that has already been included. One test source.
- source status
- live in production · valid for 30 days
- measured on
- · 14 days old
- coverage
- from under review to approved only through a recorded human decision, through the Gateway · 51 approvals in production in 30 days to 13-11-2026 · completed execution with a traceable trail · an expired decision is not applied
- source
- Control measurement of knowledge approval in production, 13 November 2026
Approval records who decided and when; it says nothing about substantive correctness. Part of the approved knowledge is a synthetic test object.
- source status
- live in production, synthetic test object · one-off trial, not repeated
- measured on
- · 39 days old
- coverage
- if a qualification on the knowledge is still under review, the representation is refused · once that qualification is approved by a person, it appears in the result as a condition, alongside the core · every result with a receipt and a traceable trail
- source
- Control measurement of representation in production, 9 October 2026
Synthetic test object. Proves a qualification recorded as a separate condition, not uncertainty within the text of a claim. Refusing is not a nuanced representation. Says nothing about substantive correctness.
- source status
- function present · valid for 30 days
- coverage
- dependencies of a source across several steps
Candidate until a measurement with a source has been designated.
No designated source: this entry cannot carry a status.
- source status
- dry backtest on public statutory text · valid for 30 days
- measured on
- · 19 days old
- coverage
- all 78 versions of the Dutch Bankruptcy Act since 2002 (wetten.overheid.nl) · at paragraph level 0 missed changes and 0 false alarms across 76,973 comparisons · at article level 211 false alarms (8.6 per year), 208 of them from a single text placement (Stb. 2005-600) · a first, simpler rule failed: 306 missed changes, which is why the rules were fixed before the repeated run · replay on article 370 as of 1 January 2023 (Stb. 2022-491): paragraphs 3 and 4 changed, paragraph 1 did not; a claim on paragraph 3 or 4 would go under revision, the existing approved claim on paragraph 1 rightly stays valid
- source
- Source watch Bankruptcy Act, dry backtest, 29 October 2026; independently repeated with a byte-identical result

Dry backtest, not a live revision. Putting a claim 'under revision' does not yet exist as a controlled function; that follows with a real change, through a human decision. The daily watch (read-only) is being built and is not yet active. One act, one claim.
- source status
- measurement on one public test source, two runs · valid for 30 days
- measured on
- · 19 days old
- coverage
- splitting a document into knowledge objects (way of working, principle, fact, case, statement); a language model only labels and groups, the software rejects any answer containing its own text · on a hand-made test fixed in advance (VNG guide, 59 objects): facts found 5% → 82–91%, principles correctly identified 47% → 83%, step plan without errors, false cases 2 → 0 · costs: facts split more finely (64–71% exact match), more separate statements
- source
- Object proposal measurement, 29 October 2026; prompt and code fixed before the test
One test source and two runs, not a general reliability claim. Cases not tested. The model step does not yet run in the working system. The web lane checked the measurement tables and the order of fixing and measuring, not the locally stored test itself.
Packs
Seed · Expert · GEO- source status
- live, public source · valid for 30 days
- measured on
- · 14 days old
- coverage
- source → candidate → human admission decision → recorded entry → human knowledge approval, with origin and evidence entry, each with a receipt in the gateway:production chain · 30 days to 13-11-2026: 14 seeds, 50 inclusions, 51 approvals
- source
- Control measurement of the Seed route in production, 13 November 2026
Admission and approval apply per claim, not to the whole source. No production authority outside the own project.
- source status
- live, synthetic real-world measurement · one-off trial, not repeated
- measured on
- · 47 days old
- coverage
- case creation · document processing · source validation · intake · analysis · human judgement on exactly one result version · governed representation · result + Result Passport, executed within one continuous synthetic workflow
- source
- Full real-world workflow closure evidence, 1 October 2026
Proves that the Expert Pack methods can be executed together in this synthetic workflow. No real client data and no substantive validation of a financial, legal or financing judgement.
- source status
- live, through the Gateway · one-off trial, not repeated
- measured on
- · 33 days old
- coverage
- baseline (geo.baseline), 8 October 2026 · representation of approved knowledge (geo.represent), 15 October 2026 · each with a receipt in the gateway:production chain
- source
- DyReP control measurement of the GEO methods · 28 October 2026
Proves execution of the methods and a traceable representation. Says nothing about findability or citation in AI answers.
Connections
Claude/Cowork · ChatGPT/DyReP app · other connected environments- source status
- live, test profiles · valid for 30 days
- measured on
- · 4 days old
- coverage
- OAuth client of the Claude connector · 37 successful tasks in 30 days to 13-11-2026, 21 of them with a receipt in the gateway:production chain
- source
- DyReP control measurement of Claude calls · 13 November 2026
Proves use of the client used by the Claude connector, not a separately established Claude identity. Test profiles and synthetic test data; says nothing about the correctness of the professional method.
- source status
- live, trial profile · valid for 30 days
- measured on
- · 38 hours old
- coverage
- OAuth client DyReP – ChatGPT · 12 successful tasks in 30 days to 18-11-2026, 12 of them with a receipt in the gateway:production chain
- source
- DyReP control measurement of ChatGPT calls · 18 November 2026
Synthetic trial data in a trial profile: proves call, identity and receipt, not production use with real client data.
- source status
- live, single entry point · valid for 30 days
- measured on
- · 6 days old
- coverage
- OAuth client of the DyReP Orchestrator · 573 successful tasks in 30 days to 13-11-2026, 112 of them with a receipt in the gateway:production chain · source submission, seed, review and knowledge approval through one entry point
- source
- DyReP control measurement of Orchestrator calls · 13 November 2026
Proves calls through the Orchestrator client under a person’s own login. The Orchestrator operates the route; a person decides on knowledge.
- source status
- live, trial identity · one-off trial, not repeated
- measured on
- · 35 days old
- coverage
- three admitted OAuth test clients: Hermes, OpenAI Agents SDK and Claude Agent SDK · each 12 successful tasks with a receipt in the gateway:production chain (12–13 October 2026) · case file, case status, knowledge reading and representation
- source
- DyReP control measurement of agent calls · 28 October 2026
Trial identity and synthetic case files. Proves that an agent runtime works through the same entry point and within the same boundaries; the agent gets no authority of its own.
Expert work
Case file · intake · analysis · human judgement · full real-world workflow- source status
- live, synthetic real-world measurement · one-off trial, not repeated
- measured on
- · 47 days old
- coverage
- one new synthetic case file through case creation → multi-file upload with type gate → extraction → human extraction review → intake → analysis → human judgement on exactly one result version → governed representation → result + Result Passport; lineage via analysis → review → representation
- source
- Full real-world workflow closure evidence · 1 October 2026
Synthetic case file; no real client data and no financial, legal or financing judgement. For this trial, representation lineage is proven through the exact analysis result → human review → receipt predecessor; direct binding to case-file extractions is not a criterion of this trial.
- source status
- live, synthetic test case file · one-off trial, not repeated
- measured on
- · 48 days old
- coverage
- three source files read → each source validated by a person → analysis in three versions → human judgement bound to exactly one version (revised twice, then accepted) → earlier versions and judgements are kept; unknown data stays unknown
- source
- Control measurement expert case file, 30 September 2026
Synthetic test case file; not a real-world measurement and not a financial, legal or financing judgement. A governed representation result, including product and Result Passport, can now be read live through the Gateway; the separate full real-world workflow was measured in PW-001.
Website
Contact · public measurement layer · representation- source status
- production · valid for 30 days
- measured on
- · 13 days old
- coverage
- NL 14-09 · EN 24-09 · one submission on repetition · storage in the EU 28-09 · latest receipt in the EU 05-11
- source
- Control measurement of contact form storage, 13 November 2026
Proves receipt and storage in the EU; not the handling of replies by email.
GEO
Baseline measurement · diagnosis · any re-measurements- source status
- live measurement, sealed · one-off trial, not repeated
- measured on
- · 40 days old
- coverage
- 30 questions, each asked twice: 60 of 60 answers complete · DuckDuckGo Assist · 118 source citations to 45 domains · 0 to dyrep.org · completed and recorded through the Gateway
- source
- GEO baseline measurement, 8 October 2026
One answer engine and one question set. A baseline measurement is not a measurement of effect. The earlier baseline of 8 September (ChatGPT, Google AI Mode, Perplexity) has not been repeated. The measurement queried DuckDuckGo Assist automatically; new runs by that route are paused until the terms of use have been clarified. A next measurement runs through permitted APIs and gets its own baseline; this baseline is not a point of comparison for it.
Ground rules
- 01
No source, no status.
An entry without a designated evidence source is marked Not yet measured.
- 02
Measurements can expire.
A snapshot does not remain current simply because it was accurate when recorded. The register entry shows when re-measurement is needed.
- 03
A test stays a test.
A test case file does not become a real client case through wording.
- 04
One green entry proves only that entry.
It does not establish client deployment, routine use or market impact unless those have been measured explicitly.
Test DyReP on one real result.
Describe one recurring task. That is enough to begin.