v1.12.0minor
Provenance shading
Released May 1 · 2026Confidence is now visually present in the prose, not buried in tooltips. Every claim Pulse synthesises is rendered with a subtle underline whose intensity tracks the model’s calibrated confidence. Hover for the chain. The shading reads as ink, not as decoration, we tuned it for 60+ minute reading sessions.
addedProvenance shading in synthesised prose
Underline density tracks calibrated confidence. Off by default for new tenants; admins can switch in Settings → Reading. Power users will probably want it on.
c3a4f9e · #2814 addedPer-claim “where did this come from” inline trace
Hover any shaded span to see the chain, sources, retrieval scores, the prompt that synthesised it. Click to pin the trace open while you keep reading.
7b21d4a · #2820 improvedCalibration on retrieval-poor queries
When fewer than 3 high-quality sources are retrieved, Pulse now leads with “I’m working from limited material” rather than padding the answer. Brier score on the small-source bucket dropped from 0.18 to 0.11.
f0c819d · #2806 fixedBriefings late by 6+ minutes for tenants in APAC
Schedule queue was using UTC offset of the workspace creator instead of recipient. Apologies if your 8:30 brief was a 10:30 brief for the last week.
d44e6c1 · #2818 changedDefault confidence threshold for “publish without flagging” raised from 0.78 → 0.85
You’ll see slightly more “I’m not sure” tags, especially in the first weeks after a new connector goes live. This was deliberate.
a1b22f0 · #2823 securityCVE-2026-0314, patched, not exploited
A regex in the policy engine had catastrophic backtracking on adversarial inputs. Fixed in policy-engine v0.32.1. Hosted tenants got the patch automatically; self-hosted operators should upgrade.
CVE · advisory-0314 23 commits11 contributors0 rollbacksHealthy: 100% rollout, no incident
v1.11.0
Skills v2: forks, fixtures, finer scopes
Released Apr 24 · 2026Skills (the Network’s reproducible workflows) get a v2 manifest. Fork lineage is first-class, capability scopes are 14 finer-grained verbs instead of 4, and skills can ship with fixture data so you can dry-run without touching production.
addedFork lineage in skill registry
Every skill now shows its parent chain back to the original. Forking is one click; your version automatically credits the upstream author and tracks divergence.
addedSkill fixtures for dry-runs
Authors can ship synthetic data alongside a skill. Importers can run the skill against fixtures before pointing it at their own tenant. No more “let’s just see what happens”.
changedCapability scopes split from 4 verbs into 14
The old read / write / tool / net scopes are now read:map, write:tasks, etc. Skill manifests on v1 are auto-migrated; the upgrade is non-breaking.
improvedSkill execution is now fully deterministic on replay
Same inputs, same retrieval snapshot, same model version → byte-identical output. Useful for compliance archives and bug repros.
removedSkill global variables (deprecated v1.7, gone now)
If your skill still uses $global.*, it stopped working today. Migration guide is on the docs site.
41 commits14 contributors0 rollbacksHealthy
v1.10.0
Map quality: and a smaller graph
Released Apr 17 · 2026A concentrated quality release. We rewrote two retrieval paths, deprecated 4 stale relation types we’d been carrying since beta, and tuned the entity-resolution ratchet. Net: average map graph for a customer dropped 14% in node count; retrieval relevance went up 9%.
improvedEntity resolution: stricter ratchet on people
“M. Chen”, “Mai Chen” and “mchen@” used to merge cautiously. Now we merge confidently on triangulated email + display-name + IdP claim. Wrong merges are rare and reversible.
improvedRetrieval recall on cross-doc topical questions
Two-pass retrieval (BM25 + embeddings, then a re-ranker that has access to the full map context). Recall@10 up 12% on the internal eval, latency unchanged.
removed4 deprecated relation types
vague_about, related, topical, contextual. They were shortcuts that hurt precision. Maps will rebuild on these edges automatically over the next 48h.
fixedSlack threads losing reply context after 100 messages
Connector was paginating but not stitching. Long threads now read end-to-end.
34 commits9 contributors1 rollback (re-shipped same week)
v1.9.0major surface
MCP 1.0: the open wire is stable
Released Apr 10 · 2026Our Model Context Protocol implementation hits 1.0. Frozen wire format, signed releases, semver guarantees. If you’re building an editor, agent, or chat client that wants to talk to Pulse, this is the contract, and we won’t break it without a major version bump.
addedpulseteam/mcp v1.0 release
Homebrew, npm, Docker, standalone binary. Signed releases. Reference clients in 5 languages.
addedStreaming retrieval over MCP
Sources arrive as they rank; the synthesis tool can start drafting before retrieval completes. Cuts perceived latency on long answers by ~40%.
changedDefault transport: stdio (was: stdio + websocket)
Websocket transport is still supported but no longer the default, too many editor sandboxes refuse to open arbitrary outbound sockets. Stdio “just works”.
securityCapability tokens are now per-session, not per-install
An MCP client gets a fresh capability token at session start. Compromise of a long-lived install token can’t replay across sessions. No action needed for hosted users.
62 commits21 contributors0 rollbacks
v1.8.0
The calibration loop, turned on for everyone
Released Apr 3 · 2026For the last quarter, every confidence number Pulse showed was verified against ground-truth feedback in a held-out cohort. The loop’s been quietly improving for ten weeks; this release flips it on tenant-wide. Brier score on the production traffic dropped from 0.14 to 0.09, calibration is now actually calibrated.
improvedCalibration loop default-on
Confidence numbers shown to users now reflect post-hoc calibration. Loop is per-tenant, your feedback only tunes your numbers, never anyone else’s.
addedCalibration dashboard for admins
Reliability diagram, ECE, Brier score, all stratified by query type. Useful for figuring out where Pulse is overconfident in your specific data.
fixed“Confident-and-wrong” cluster on legal queries
Pulse was over-trusting policy docs that hadn’t been updated in 18+ months. Recency now factors into confidence.
28 commits10 contributors0 rollbacks