Repository navigation
Releases
Telryn v0.12.0 · Conduct evidence and signed history
Telryn 0.12.0 records hosted desk access by authenticated subject, route and disclosed case IDs. Teams can select dated Scotland and Northern Ireland calendars and see upcoming expiry. The payments rehearsal distinguishes internal registration from PSD/EMD response deadlines. The standalone deadline dataset includes its own version, sources and schemas.
Risk teams can verify signed cross-quarter histories against an earlier checkpoint. Vendors can issue an attestation with a passing agent release test, quarterly calibration, a bank statement and customer control mappings. An independent operator can receive signed opt-in benchmark aggregates and return a frozen result to qualifying peers.
Use the optional read-only warehouse adapters for Snowflake, BigQuery and Databricks, and the scoring adapters for Braintrust and Langfuse. Provider checks use controlled fixtures. The release does not establish customer deployment or certification.
Install or upgrade with pipx. The signed checksum manifest covers the deadline dataset, source archive, example reports, container image and SBOMs.
Telryn v0.11.0 · Complaint audit pilot and hosted QA desk
Telryn checks closed conversations against a firm's complaint policy, register deadlines and owner actions. Version 0.11.0 adds dated England and Wales and Federal Reserve Bank calendars, after-hours receipt rules, confirmed contact linking, and weekly statements that distinguish missed deadlines, late registration and unresolved evidence.
The desk can import review-level scores, attach case evidence, review the policy and issue a bank statement from the browser. Teams can run it behind their own HTTPS identity proxy with lead and reviewer roles. Verified weekly audits feed vendor calibration trends and quarterly committee packs. Benchmark exports require an explicit opt-in and stay local.
Install with pipx install telryn or pipx upgrade telryn. Try telryn demo --qa-passes --out runs/scored, telryn demo --agent-release --out runs/agent-check, and telryn outcome --example vulnerability. The demos keep their rehearsal data under their output directories. Existing routing commands remain callable under their previous names as well as under telryn queue.
A preregistered comparison used Telryn's local cited extraction on 320 authored chats. At the same modeled 255 staff minutes, Telryn found 89 complaints, SQL found 85 and 57 overlapped. The paired 95% range for the four-case difference was −5 to +15. These are synthetic results. No live tenant or field deployment is claimed.
Detection study · Hosted desk · Calendar and clock rules · PyPI
Telryn v0.10.0 · Audit QA passes against dated outcomes
Telryn now checks which closed conversations passed a QA tool while a required complaint missed its registration deadline. The weekly report measures complaint-classification disagreement and missed deadlines among QA passes separately. Both measures carry their own populations and uncertainty.
Install with pipx install telryn or upgrade with pipx upgrade telryn. Run telryn demo --qa-passes --out runs/qa-passes in a new output directory, then follow C102 and C204 from the vendor decision to the human grade, register and owner action. Both cases are synthetic. The existing saved-SQL comparison remains available with demo --complaints.
A bank can receive the aggregate statement and replay a disputed case with its own private report, frozen exports and the issuer’s independently supplied digest. A second worked outcome follows a requested accommodation from its agreed deadline to a dated delivery receipt. Use telryn outcome --example accommodation to run it locally.
This release includes the scored weekly report and bank specimen as signed-manifest assets. Existing signed audits keep their verification path. Score audits require v0.10.0 or later.
PyPI package · Scorecard CSV · Bank statement · Security review
Telryn v0.9.4 · Trace a missed complaint through owner action
Run pipx install telryn or upgrade with pipx upgrade telryn, then run telryn demo --complaints --out runs/complaints-pilot in a new output directory.
Open C102 in the first synthetic weekly report and C204 in the second. Each case links the customer's cited words and the reviewer's classification to the policy deadline, owner acceptance, registration and recorded completion. Mira’s C108 shows a referral still awaiting acceptance.
The installed command prints both weeks' results and saves the actual selections, staff time and owner outcomes in comparison.json. Conflicting owner histories remain unresolved. Both weekly reports use the rehearsal trust store, and the separate grading packets stay blinded. Reports carrying the new source citations require v0.9.4 or later. Existing reports remain readable using their original trust store.
Telryn v0.9.3 · Recover scheduled audits and verify release provenance
The weekly scheduler now checks the delivery date against the plan's audit week before issuing a sample. Pass the original Monday delivery date to recover a missed run. The private browser desk can open scheduled audits and accept grades for them. A failed output-directory publication leaves the report count unchanged, and verifying the first signed report completes an interrupted registration.
The complaints rehearsal uses separate private trust storage. The installed CLI, generated reports and operator guides use commands that work with pipx. Existing signed reports can still be verified using their original trust store.
Python publication runs from protected main. It verifies the signed source archive, CI-qualified tag and exact distribution hashes. A retry after a partial upload can deliver the missing distribution while preserving the original files.
PyPI package · Weekly operations and backfills · Release verification
Telryn v0.9.2 · Install the complaints audit CLI with pipx
Install the complaints audit CLI with pipx install telryn on Linux or macOS with CPython 3.12–3.15 and SQLite. Rehearse a two-week complaints audit from any directory with telryn demo --complaints --out runs/complaints-pilot. Open the first week's report and find case C108 to trace the original passage, deadline and follow-up. The two weeks use synthetic records and staff times.
This version packages the offline desk, weekly engine, built-in policy packs and qualification fixtures. The base CLI has no third-party runtime dependencies. For regional Bedrock use pipx install 'telryn[bedrock]'. The weekly guide includes a cron example, a weekly delivery directory and persistent private trust storage.
The signed container, both SPDX inventories and security review pack are included below. GitHub distributes a snapshot of the repository. PyPI distributes a wheel and a source package under telryn 0.9.2, with its own file hashes and publisher attestations.
Telryn v0.9.1 · Signed private deployment and complaint policy packs
Run the complaints desk and weekly audit inside your network. This release ships a signed Linux image for amd64 and arm64, a portable OCI archive, a software inventory for each architecture and a downloadable security review pack.
- For IT and vendor risk. Verify the signed checksum manifest before importing the image. The review pack maps each data flow and answers questions about egress, credentials, storage, retention, report signing and recovery.
- For the complaints lead. Issue a weekly audit using an exact-version FCA DISP, US Regulation E, CFPB Company Portal or firm-defined policy pack. Reviewers grade the qualifying complaint independently of the later register check. Reports preserve the rubric, category order and clock policy in force when the audit was issued.
- For the cloud owner. Use local Ollama, a specified Azure OpenAI deployment, or an in-region Bedrock model in the firm's AWS account. Each definite fact needs a locally verified citation to a redacted passage. The default export workflow requires no inference service and runs offline.
The provider contracts were exercised with synthetic records and controlled transports. The included policy files require the firm's review before use on a regulated queue. The synthetic and public-data examples show how to reproduce the audit.
Container verification and deployment · Security review · Policy packs · Model endpoints
Telryn v0.8.0 · Complaint episodes and AI agent release checks
A complaints lead can now follow one issue across human and AI support and see which owner action remains outstanding. The signed weekly report links the customer’s contacts to policy citations, deadlines, dated register records and case-manager actions. It also names uncovered channels, their owners and review dates.
- Track the episode. Join dated contacts using team-keyed customer and issue references. Distinguish a registration deadline from a final-response deadline. A changed policy clause can flag earlier signed reports for another review.
- Review the classification. A complaint decision, category, vulnerability and redress each have their own fields. Reviewers document the reasons for a downgrade.
- Test an agent change. Send synthetic complaint phrases to the agent’s test workspace and check a separate, dated register snapshot. The gate requires an on-time entry for each complaint.
- Read the field comparison critically. In the first synthetic draw Telryn found six accepted episodes in 131 stipulated staff minutes. The saved SQL search found five in 125. Enumerating all 1,350 possible selections on these authored cases shows SQL matching or beating Telryn in 96% of draws. Customer effort, earlier action and paid demand remain unmeasured.
Open the synthetic complaints report to inspect Mira’s chat, call and policy deadlines. Download the source to run the two-week rehearsal, both seeded release checks and the qualification suite. Verify the release files with shasum -a 256 -c SHA256SUMS.txt. The Ubuntu report is a separate public-data support example.
Worked complaints case · Policy and episode evidence · Agent release check
Telryn v0.7.0 · Complaints QA from conversation to owner action
Telryn helps a complaints quality lead trace a sampled conversation from a human policy grade to a register deadline, a dated entry and an owner’s response. The register and work records attach to the existing signed HTML report. Reviewers keep their original blinded packet.
- Rehearse two weeks on a synthetic payments queue with a saved SQL search and Telryn’s committed draw. Grade the union once and compare accepted complaint episodes, method overlap, elapsed action time and staff minutes.
- Inspect four queues in the same report. See registrations awaiting acceptance, work an owner accepted, completed work still missing a register entry, and cases needing better evidence.
- Plan from a tolerable miss rate. Enter the number of cases the lead needs to detect, the desired probability, and time for second review and adjudication. At 25 qualifying cases among 5,000, 50 uniform first grades give a 22.3% detection chance under perfect recognition. Reaching 95% takes 564.
- Compare archived Intercom 2.17 scorecards with an independent ledger locally. The rehearsal distinguishes human submissions, AI results, assignments and later observations. Live scorecard ingestion remains contingent on a measured customer week.
In the release rehearsal both methods found the same three owner-accepted complaint episodes. The saved-search workflow used 102 simulated staff minutes across both weeks and Telryn used 118. This run fell short of the proposed investment rule. Draws can vary with the team key. Earlier action, actual staff effort and paid repeat use require customer evidence.
Download the source to run the rehearsals, the synthetic complaints HTML to inspect the workflow, or the Ubuntu report for the public-data census. Verify the files with shasum -a 256 -c SHA256SUMS.txt.
Telryn v0.6.1 · Weekly QA evidence and connector fixes
Weekly audits now keep a confirmed serious finding when reviewers disagree about prior handling. The QA desk also keeps adjudication drafts with their case and invalidates sample previews as soon as the budget changes.
- Freshdesk transcript imports distinguish the requester, assigned agent and unknown authors. Salesforce excludes draft email. Intercom validates conversation parts, handles week boundaries and checks each tag handoff against the current record.
- Source adapters reconcile record identities across pages. An overlapping page can no longer satisfy an expected count while hiding a missing record.
- Local transcript screening checks the installed weights before saving resumable results. The first resume from an older checkpoint reprocesses it once. Existing signed weekly reports retain their original evidence.
The release was checked with 150 qualification cases, real Chrome, a local Ollama screen, a public-data census, and a held-out fintech example. Connector checks used synthetic vendor responses and local services. Live tenant access and downstream assignment remain customer-specific.
Download the source archive to run Telryn or open the Ubuntu report to inspect the public census. Verify the downloads with shasum -a 256 -c SHA256SUMS.txt.