Loading…
Choose the existing term that is another name for. Documents that use it stay tagged as they are, and are found under the term you choose.
While no plan is chosen your documents are not shown here. Nothing has been deleted, and you can take a copy of every document away.
Your organisation administrator chooses the plan.
Your organisation administrator or an analyst can download the documents.
No document open. Use Open from QMS Library in the palette on the left.
Loading the document sent by the calling system…
Enter your username and registered email address and we’ll send you a reset code.
Too many failed login attempts.
Your organisation requires MFA. Scan the QR with an authenticator app (Google Authenticator, 1Password, Authy, etc.) and enter the first 6-digit code to complete setup.
The only way to add someone to an existing organisation. They set their own password from a link that works once.
A RegPack is a set of documents from one publisher or from this organisation. Choose a pack above, add its documents, and drag one document onto another to read them as one framework item; right-click an item to unlink it. Each framework item has a force and is what people tick in the Framework selection.
Currently logged-in users. Revoke terminates all sessions for that user immediately.
API keys allow external systems and the Python client to submit documents and retrieve results without a user login. Keys are scoped to this organisation.
Audit records are kept until an administrator archives them. An archive writes the records to a retained file before they leave the live trail.
Shown in the Messages column of the welcome panel at every login. All the messages that apply to a user are shown together, widest first — yours does not replace anyone else's. Leave a box empty to take that message down.
One message, to one address, sent now and recorded in the audit trail with the recipient.
Set in your .env file. Required for welcome emails and password resets.
RESEND_API_KEY
RESEND_FROM (must be an address on a domain verified with Resend)
Sent when creating a new user with "Send welcome email" checked.
Use {username}, {login_url} and {brand_name} as placeholders. The password is never sent by email.
What qualified people did with the findings. In an opinion-led domain the action rate is the accuracy measure — nobody can mark an analyser’s opinion right or wrong in the abstract, so what a reader did with it is the only signal there is. Every number states what it is partitioned by.
Send a test card to the Teams notifications channel or a test email via Resend to verify each channel is configured correctly.
Token & external-API spend from the billing_usage ledger, plus the
per-document ingest rollups. Defaults to the current calendar month.
Loading…
Loading…
Loading…
Minimum Jaccard overlap (0–100%) required for the consistency test to pass.
Set in the admin panel or override with CONSISTENCY_THRESHOLD in .env
(the .env value takes precedence).
Pick two saved analyses for this organisation and compute span-overlap between their findings. Useful for evaluating model upgrades or comparing SOP draft vs. approved versions.
Runs the analyser N times in parallel against the same document and scores pairwise span overlap. Bytes are loaded once and shared in memory by all runs — no document-store renames, no filename collisions.
Compares two consistency batches against the same SOP — typically one per model — to measure how much the models agree on findings, severity, and category. Both batches must be on the same document and (ideally) the same framework / depth.
Every model the system can call lives here. Two-axis classification: vendor (the company — anthropic / openai / xai / google / client-X) and wire protocol (the API shape — anthropic_messages / openai_chat_completions / google_generative). chat_handler dispatches on protocol, not vendor — so a Grok model registered as vendor=xai, wire_protocol=openai_chat_completions takes the OpenAI streaming path automatically. Each model declares its own key_env / url_env environment variables; the model is runtime-available iff enabled=true AND the key env var is non-empty.
Loading…
Every call is costed at the rate that was in force on the day it was made, so changing a price does not restate what earlier months already reported. Two different things can happen to a rate and the system needs you to say which: the price changed applies from a date forward and leaves closed periods alone, while we held the wrong rate recalculates every period that rate covered. Nothing is ever deleted — a superseded rate stays on the record.
Loading…
These are tuned by what the call is doing, not by which model runs it — the same model wants a different ceiling for a standard analysis than for a debate turn. Output ceiling is the enforced limit and caps the model's reasoning and its answer together, which is why raising effort without raising the ceiling is what cuts answers off. Effort is how hard the model reasons; leave it blank to take the vendor's own default. Pacing budget is advisory and only helps a multi-turn loop such as a debate.
Loading…
pgvector-backed semantic-search foundation. Three namespaces share one vector store: sops (SOP paragraphs, org-scoped via RLS), regulations (CFR / ICH / ISO chunks, global), corpus (adjacent regulatory material — MHRA blog, FDA guidance, EMA, EU Commission; Phase C consumer). Per-persona retrieval profiles arrive in Phase B; inverse-RAG regulation-coverage queries in Phase D; v4.0.0 close at Phase E. See project_v4_program.md.
Loading…
Each persona is one AI call site in the analysis pipeline. These are ParaQualis’s choices, and they are the seed copied into a NEW organisation. From that point the organisation’s own settings rule — so changing a value here does not move an organisation that already exists. To see what an organisation is actually running, use Your organisation’s models below.
Loading…
Which AI models this organisation has approved for use, and which model each persona runs on here. Approving a model records a quality decision; it does not change what is running. ParaQualis recommends a model for every persona — you can accept the recommendation or choose any model you have approved.
Each prompt below is a system prompt the analyser uses at a specific call site. Edit any prompt to override your org's behaviour for that call site; reset to fall back to the parent domain's template. Edits take effect on the next analysis run — no restart required.
Loading…
Each lens is a complete, standalone QE persona prompt. The lens picker on the analysis page lets the user pick which lens drives the QE chat for a given SOP. Click the empty box in the Default column to promote a lens to default — that's the lens new SOPs land on.
Loading…
Edit the prompt templates that ship with each domain. Existing orgs are NOT auto-updated — they keep the snapshot they were seeded with at creation. Each org's admin sees a banner on the AI Models tab when their domain's template version advances; they opt in to the new text by clicking "↻ Reset to domain template" on individual prompts.
Pick a domain above to load its template prompts.
Tuning surface for the level-2 activity derivation — the readable "management-level" map derived once at ingest over the level-3 step graph. Two knobs: the per-process grouping prompt the LLM runs, and the singleton threshold that governs how coarse a large process is allowed to be. Edits take effect on the next ingest / re-derivation — no restart required.
The system prompt for the per-process grouping call. One call runs per level-1 process; it receives the ordered node list and must return the activity grouping JSON. Retuned to coherence with no min/max — the model decides the count. A failed call never blocks ingest (the process falls back to deterministic segmentation).
max_nodes_for_single_step_activity — single-member ("one-step") activities are allowed only
when a process has this many grouped nodes or fewer. In a larger process a deterministic
post-pass merges every singleton into its sequence-adjacent activity, so a big process is never fragmented
into one-step blocks. Higher = more one-step activities tolerated; lower = coarser maps.
Loading…
Select a domain to edit.
The base colour for each finding category. This is the single source of truth — it applies everywhere: the count pills above the document, the in-document highlights, the finding cards, and the process graph. Pick a base colour; the light background and dark text are derived automatically for contrast. Changes preview live; Save persists them for the domain.
Appears in the About dialog, visible to all users.
Read-only. Every store the system keeps, with how many rows are in it right now — and, for each, what an empty one would mean, because “0” from a bulk load that has never run and “0” from a table that fills as people work are different facts. Open one to see the most recent rows.
Toggle numbered debug-instrumentation components live. Each component
gates a family of [instr:<n>:<name>] log lines on the
backend. Changes apply immediately to all subsequent log output but are
in-memory only — on server restart the active set reverts
to the RC_INSTRUMENT environment variable. Enabling
all (component 0) prints every component.
Loading…
The structural checks run against every document in this organisation. Turn a rule off and it produces nothing at all — no finding, nothing to disposition, no hidden backlog waiting if you turn it back on. The action decides what an outstanding finding does to a promotion:
Rules marked seeded are running the default this build shipped; the moment you change one it becomes this organisation's own setting.
What each plan includes, and what each functionality costs on it, is set in one place: the cost-and-permission matrix, in Content management, section ABC. The table below is what does not vary by plan: administrators' permissions, which apply universally within their own scope. An untick takes effect on the next request, with no release.
Edited at /content_mgt, with the rest of the public wording. Dev and test only — these words are authored here and are not editable on a deployed arm.
A periodic review records that a person looked at a set over a period and found it acceptable, or did not. The system never records one for you and never infers one from the data looking healthy — a record the system wrote would attest to nothing.
The period each set was last reviewed over. Whether that is overdue depends on how often your organisation reviews it, which is your decision and is not held here.
State:
Per-principle rules — plain English, one rule per line is clearest. These tell the AI what to always or never flag for this organisation's SOPs.
The system prompt used whenever the SUCCESS analysis runs. The default is a management-consultant persona driven by the Trusted Advisor model, communicating top-down with the Minto Pyramid Principle, focused on actionable advice and willing to ask clarifying follow-up questions when the SOP is ambiguous.
These operations permanently delete data from the database and cannot be undone. They are intended for resetting a test or development environment, or wiping a fresh installation before go-live. Do not use in production with real data.
These actions are irreversible. All selected data for the chosen domains will be permanently deleted.
Quality Expert Review
This Visio file has multiple pages. Select the page to convert to a SOP:
This SOP is Effective. To make changes, the system will branch a new Draft from the current Effective version. The Draft inherits the current content and remains editable until you promote it back to Effective (the auto-supersede step will then move the current Effective to Superseded).
Why are you making this change? (required)
Change type
The rationale is recorded as part of the audit trail and travels with the new revision through its lifecycle (spec §4.19). A formal Change Plan can be attached later via the QMS curation panel.
Changing this document's status is a controlled action. Re-enter your password to apply your electronic signature (21 CFR Part 11 §11.200).
What you are approving
Password (required)
Reason (optional)
Rejecting records a decision not to act on this finding — kept with the document's change history (21 CFR Part 11) and used to calibrate the analyser.
⊘ Pre-filled from the expert debate's DISMISSED conclusion — review, adjust if needed, and sign.
Reason (required)
Abandoning discards the in-progress Draft. This is an allowed business choice — it is NOT a deviation. A reason is recorded for traceability.
Reason (required)
Working versions are uncontrolled drafts (saved analyses). Deleting them does not affect the Effective or controlled Draft version.
| Number ⇅ | Title ⇅ | Type ⇅ | Status ⇅ | Version | Effective date ⇅ | Analysis | Owner | Groups | Used by |
|---|
This is the organisation's agreed vocabulary — what document tags are filtered and reported against. A document is always tagged with what it actually says, so removing a term here does not change any document's tags; it only stops the term being part of the agreed set. New terms found in imported documents wait below until you decide on them.
Templates define the structure for each document type. Upload a .docx master and define its section list — used when reformatting existing SOPs into the template structure.
Only documents with a stored file are shown. Select one to load it into the Inspector.