Changelog

Dated entries for changes that affect what you see, what you can connect, or what the evidence means. Entries marked pending deploy, when there are any, are complete in the source and not yet live on the platform.

2026-11-16

The evaluation itself is evidence: calibration, disputes and versioned rules. A result you can verify is not a result that is right, so a quality finding now says what it rests on. An evaluator reports how a version did on cases with a known answer (POST /v1/evaluator-calibrations), and the finding says whether that version is calibrated, weak, insufficient, stale or not calibrated; a calibration does not carry over to a new version. A person can dispute a result or a whole series with a reason (POST /v1/evaluation-disputes); the results are never edited, the series is read a second time with the disputed results set aside, and the evaluator's owner records how it was settled. The thresholds are now an organization setting with a version, who changed it and why (GET and PUT /v1/evaluation-rules); each finding names its rules version, and a follow-up is not comparable across versions. The agent's Quality tab shows all three. See Evaluated quality and the Evaluations API.

A rollback is recorded, and recovery is checked, for the same cohort. Your feature flag platform or any control system can report a configuration change (a rollback, a rollout) to POST /v1/interventions: the agent, the flag, the variants, the cohort, who or what made it and when. Provantix keeps it as that system's report and never makes or approves a change. On the agent's page, Changes checks whether the same cohort recovered, from the runs after the change: whether they are on the expected variant, whether the declared steps are back, whether the evaluator's results came back under the same definition, whether errors, run time, tokens and cost stayed where they were, and whether it held. The verdict is recovered, not recovered, not enough evidence yet, or not comparable, and a rollback event on its own is never read as recovery. The change also appears on an incident's timeline and beside a finding's follow-up. See Recovery after a change and the Interventions API.

2026-11-15

The app menu is seven entries, and related pages are tabs. The left menu is now Home, Agents, Spend, Investigate, Evidence, Connectors and Reports, with Setup, Docs, Settings and My account below. Pages that used to be separate menu items sit together as tabs: Agents (Inventory, Live state, Candidates), Spend (Cost, Providers), Investigate (Search, Findings, Incidents) and Evidence (Coverage, Decisions, Compare, Audit log). Reports keeps its own menu entry. Every earlier address still works. The menu collapses to icons on a desktop and opens as a drawer on a phone. The number on Investigate is the count of high or critical open findings; the Findings page now says how many of its open findings those are. See A tour of the workspace.

Home, Cost, Settings, Findings and Connectors are split into tabs. Home: Needs attention, Activity, Agents, Evidence, Cost, Pinned, under the Brief. Cost: Spend, Waste and savings, Statement, Seats, Budgets and detection, with a banner at the top when calls have no price. Settings: Evidence, Cost, People, Access, Delivery. Findings: Findings and Alerts and routing, with one Actions menu per finding, a severity filter, and the response-loop table folded away. Connectors: Sources, Destination systems, Enforcement tools. Pages that mention a Settings card now name its tab.

One look in light and dark. The dark theme uses the same palette as the website (cyan on navy-black), the charts use the same named colours, large counts on chart axes read as 14K instead of being clipped, small text is at least 11 pixels, and the pages fit a phone without scrolling sideways.

The Home source grid is faster. On a large organization the grid of sources by category took about eight seconds. It is now about one to two seconds when read fresh, and instant when kept (a minute, served as it was while it refreshes). Each figure is computed by the same code as before; a test compares every cell with the earlier route.

The open verifier repository matches the product. The verifier in the public agent-sdk repository now includes the cost-statement arithmetic checks, the decision, control, exact-change and reconciliation checks, the transparency-log mirror and its time-stamp check, the format history and every test vector. See The verification kit.

A cost statement you can check yourself. There is now a sample cost statement (synthetic data, signed with a sample key made for the kit, not a Provantix key) and the same statement with one figure changed and the manifest rebuilt, as two small downloads that can be dropped onto provantix.tech/verify, and a page that walks through checking both: Check a cost statement yourself. The altered one fails on the arithmetic and on the signature.

The published verifiers now check a cost statement's arithmetic. The Node and Python verifiers in the verification kit, and the browser verifier at provantix.tech/verify, now add up a statement bundle's figures from statement/cells.json and check the stated total, the totals by provider and model, each allocation, the cover total and the per-department files. Before, they checked only that the files were unchanged, so a changed figure with a rebuilt manifest passed. The kit has a sample statement and the same statement with one cell changed; the format specification and conformance vectors have the new checks. A verifier from before this does not add the figures up.

People using AI directly: seats and subscriptions. The Cost page now records the AI subscriptions your organization pays for (product, plan, seats, price per seat, billing period, renewal, owner), by hand or by upload, and shares each seat's cost with the person who holds it and so their department. A seat nobody holds is a cost line of its own. Where usage is known it shows seats paid for against seats used ("seat not used in the period", a statement about a seat and never about a person), cost per active user and overage, each with its source and date; a vendor with no usage interface shows seat cost only. GitHub Copilot can be read for seat activity (the seat list only; built from the documentation and not yet checked against a live account). The organization total is shown as three labelled parts: agent and API usage, seats and subscriptions, and not connected. Departments are the default; named people appear only when your organization has turned person level on and the viewer is an admin. No conversation content is read, and no name or email address is stored. This does not find AI tools you are paying for without knowing it. Read more.

The cost statement: a closed, signed account of a month. The Cost page now has a Cost statement. Closing a month freezes its figures (by day, agent, provider, model and the department, cost center and product tags, with the prices and provider figures as they stood). It shows totals by provider and model, reconciliation, allocation that adds up to the total with "(unallocated)" as its own line, budgets, avoidable cost, incidents with a cost, coverage, the method and what it does not show, and names the kind of every figure: estimated, provider-reported or reconciled. The first page lists what a reader must know (a provider not connected, untracked usage, days not settled, a silent source, calls with no price). Late records, re-tags, price corrections and provider restatements after a close are recorded as adjustments that point at what they correct and land in the next statement. Reopening needs a reason and keeps the earlier version as superseded. A month can close by itself a chosen number of days after it ends. Download a signed bundle with per-department chargeback files (the published Node, Python and browser verifiers check every file and the signature and add the figures up from the cells), a PDF or a CSV. It is cost evidence for AI and AI agents; it is not an audited financial statement. Read more.

Cost per outcome: what each run, action and confirmed result cost. The Cost page now shows unit costs for the organization, an agent, or a department, product or workflow: cost per run, per successful run (failures stay in the cost), per action that changed something in another system, per action the destination confirmed, per person served (counts only, hidden below the small-group minimum), per unit an owner defines ("a refund processed"), and the spend with no outcome. Each is shown against the previous period with the split by model and token kind, and a change in the cost of a successful run is taken apart into more tokens per run, a dearer price per token, or fewer runs ending without failure; a model or configuration change in the same period is named as a coincidence, not a cause. A unit cost needs at least 10 outcomes by default (the organization's to change); below that it says "too few to say" with the count. Where no destination check covers the actions it says "not checked against a destination", never zero and never 100%. Provantix reports cost per counted thing and does not put a money value on an outcome or say the spend was worth it. Owner-defined outcomes are kept with who and when, ended and never edited. Migration 067. See Cost per outcome.

Cost incidents: security and cost in one account. An incident can now start from a cost trigger (a spend rate, a runaway run, a budget crossed or forecast, a new spender, a failure burn, a combined security and cost finding, or untracked usage), by hand or automatically above a severity the organization chooses, and holds the money (by model and token kind, with the excess as an estimate against a stated baseline, never a loss, and the provider's own figure by day), who and what, what it did, a timeline with the response-loop times, and the notes and decisions of the people who handled it, with a follow-up to check the cost stayed down. A security incident can have cost evidence added and a cost incident can have security findings added, each labelled. Untracked usage says plainly that Provantix has no records for it and cannot say who or what. Figures freeze at close with the prices and provider figures as they were; a later provider restatement is appended as a correction and the original stays. Each incident has a one-page summary and an evidence bundle whose cost section is covered by the signed manifest; the documentation says what the verifier does and does not check for cost. Migration 066. See Cost incidents.

Alert connectors: PagerDuty, Opsgenie, Slack, Teams, Jira and ServiceNow. Alerts can now reach the tools people already have open, as kinds of contact in a route. A pager connector opens, updates and resolves a page; Slack posts a message with Acknowledge and Open buttons and edits it in place with a line in its thread; Teams posts a card through a Workflows webhook (Microsoft retired the older connector webhooks in May 2026, and a Workflows webhook cannot edit a message, so each state is a new card); Jira and ServiceNow create a ticket with the figures and a link, comment on it, and read its status back. What a tool reports (acknowledged, resolved) is stored as that tool's report with who and when, stops the alert escalating, and never changes a finding. Credentials are sealed, shown once, rotatable and audited, each connector has a test send and a health state, and one that fails repeatedly opens a finding while alerts fall to the next step of their route. The webhook now receives a documented JSON payload that states the cost figures for cost events. Every connector is built from the tool's documentation, read 2026-11-15, and has not been checked against a live account; the product says so for each until an admin records a live check. Opsgenie is built as asked, but Atlassian ends its support on 2027-04-05. Migration 065. See Alert connectors.

Alert routing and escalation. Alerts can now go to the right person, be acknowledged, and step up if nobody answers. A route matches by kind (cost, security or both), severity and optionally a tag, and sends to the agent's owner, a budget's owner, a department's or product's named contact, a rota's on-call person, a named contact or the organization's default. A route has up to three steps; if nobody acknowledges within a step's wait, the next is notified; unacknowledged alerts are re-notified at an interval up to a cap; a resolved condition sends a cleared message to everyone who was alerted. An alert is acknowledged from a link in the message, from the product, or by the receiving system through the feed, recording who and when and acknowledging the finding so the response loop sees it. A delivery that fails goes to the next step at once, and an alert that cannot be delivered anywhere opens a finding rather than failing silently. Findings with one cause arrive as one message, notifications are deduplicated, quiet hours never apply to the highest severity, and each route has a test that sends a marked message through every step. Messages carry what, where, how much and a link, never record content or a personal handle, and the delivery log shows contact names, never addresses. The earlier alert rules are unchanged. Migration 064. See Alert routing and escalation.

Cost detection in minutes. A new check runs every minute over recent records only and opens, updates or resolves cost findings at once, emitting a feed event, so a runaway agent no longer waits for the six-hourly check. Seven rules: spend rate above the subject's own usual rate for this hour and weekday (agent, department, product or organization), a run far beyond the agent's usual that is still going, a daily or monthly budget at 80% or 100%, a budget forecast to be crossed within a set time, a model or agent that has never spent before, spend on failing calls, and a high-severity finding together with abnormal spend. A rule must hold for several minutes to fire and be clear for several to resolve, with one finding per subject. Where there is under 7 days of history a fixed threshold is used and the event says so. Late data is named in the event, and silence is never read as spend having stopped. Thresholds are set per organization with overrides per agent or tag; each firing keeps the numbers it saw. The Cost page shows the measured delay from a record being stored to its evaluation against a target of under 2 minutes, and each source's delay. The six-hourly check never resolves these findings. Migration 063. See Cost detection in minutes.

Savings register: from an insight to a decision to a checked result. Any avoidable-cost item can now become an entry on the Cost page: what was found, the estimated saving a week, an owner, and a decision to act, accept or dismiss (accept and dismiss need a reason). After the window Provantix reads the same measure again for the same agent over a period of the same length and records one result: fell, did not fall, not enough data, or not comparable (when the agent's volume, models or configuration changed so much that before and after are different populations; no saving is claimed). A checked saving is stated per run and per 1,000 model calls as well as in total, so a fall from lower traffic is not counted. Estimated and checked savings are separate totals and never added together, and every result says the cost fell after the decision, which does not show the decision caused it. An accepted insight is not raised again until it grows past the margin. Entries are never deleted; the register exports as CSV and is available as a section for a period's cost statement. Migration 062. See Savings register.

Avoidable cost: where the tokens go. A new section on the Cost page, and on an agent's Tools tab, lists eight suggestions about how calls are shaped, sorted by estimated saving: context that grows through a run, cost by step position, calls with unusually large output, reasoning tokens, small calls on a high-priced model, calls that could wait for a batch price, scheduled spend with no action elsewhere, and the most expensive runs. Each is a candidate or an estimate, states its assumption in a sentence, and is estimated from your own price table. None is a finding, none says money was wasted, and Provantix does not say a cheaper model would do the job; where the evaluator has reported results for both models they are shown beside it. An admin can dismiss a candidate for an agent with a reason, which stays recorded until restored. Every threshold is the organization's to change, audited and versioned. Where the data is too thin to say, it says so and why. Sizes and counts only: no prompt or output is read. Migration 061. See Avoidable cost.

Caching evidence: is the cache working, and what did it save. The Cost page now shows, per agent, model and workflow and per day, the share of input tokens read from a cache, the tokens written, the estimated saving and the extra paid to write a cache, from your own price table. A cache that costs more than it saves is marked with both sides of the sum. A source that does not report cache use shows "cache use not recorded", never 0%. A new finding, Cache hit rate fell, opens when an agent's share over the last 3 days is at least 20 points below its own share over the 14 before, on at least 30 calls that reported cache use on each side; it gives before and after, the calls behind each, an estimated extra cost a day, the first day and run where it changed, and any flag variant that appeared at the same time (a coincidence, not a cause). It cannot say why; the runbook lists the usual reasons. Where an application sends provantix.prompt.prefix_hash (a hex sha256 of the stable start of its prompt; SDK: model.promptPrefixHash), Provantix also lists beginnings sent repeatedly with no cache read, with an upper bound on the saving; without it, "cannot tell". Migration 060. See Caching.

Declared tools: calls outside what an agent is meant to use. An owner can now declare, on an agent's Tools tab, which tools the agent is meant to use (optionally on a system). It is kept with who declared it and when, ended and never edited, and nothing is blocked. Provantix then reports calls to a tool that is not on the declared list, with what they cost: a new finding, Tool not on the declared list, high where a call changes or sends data. It also shows, with cost, tool calls made after a run's last model call, the same tool with the same content more than once in a run, tools that mostly fail, and calls a control denied that ran anyway (that finding now says what they cost). The Tools tab gains a declared-list column and a cost column. Calls whose result no later model call used are "cannot tell": no source links a result to the call that used it. An agent with no list is not judged, and the words are "not on the declared list", never "unauthorised". See Declared tools.

The cost of failure and repetition. The Cost page has a new summary, "Spend on calls and runs that failed or repeated", with six lines: calls that failed, runs that ended in failure, retries, calls repeated in a row, loops (the same call with the same content more than five times in a run) and runs cut off by a token limit. Each shows an estimated cost from your own price table, its share of spend, the change against the period before, the split by department and agent, and the runs behind it. Every line states how it is added up and what it assumes, never says a cost was wasted, and shows retries that succeeded on their own because they were needed. "Same content" means the same content hash; Provantix does not read prompts. A model call with no price is counted and marked, not treated as free, and the lines overlap, so a total counts each call once. The repeated-call and failing-in-a-row findings now carry what their calls cost, and rise one step in severity when the estimate reaches 100 dollars. Records now keep a derived content hash and the stop reason of a model call (migration 058). See The cost of failure and repetition.

2026-11-14

Reconciliation: Provantix's records against what the provider reports. For each connected provider, model and day, Provantix now compares the tokens its records add up to with the provider's own figure, and for each day the cost estimate with the provider's cost. Each comparison has one state: matched (within 2% on tokens, 5% on cost, and a small floor so test calls are not a gap), untracked usage (the provider reports more than Provantix has records for), over-recorded, price difference (tokens match, cost does not, with the price ratio that would make them match, without changing your price table), or not comparable with the reason in words (still settling, a currency with no rate, Provantix calls with no price, Priority Tier, and more). A provider that is not connected is listed, not called matched. Only settled days are compared, and the page shows the days still settling. When a provider restates a day the comparison is redone and the earlier result is kept. Two new findings, Untracked AI usage and Cost estimate does not match the provider, state the size, the days, since when and the provider's own split by project or key, and never a cause. Providers shows, per provider and month, the share explained, untracked and unpriced, and the comparisons export as CSV with their figures. A match shows the records add up to the provider's figure, not that every call was necessary or correct. See Reconciliation.

Providers: Kimi and OpenRouter, every provider listed, and invoice totals by hand. Kimi (through its Admin API, a beta) and OpenRouter (through its activity report) join OpenAI and Anthropic as providers Provantix reads, on the same rules: a read-only administrative key shown once, hourly pulls, versioned restatements, figures labelled as the provider's. Kimi's documentation does not state the currency of its bills, so Provantix records none and says so rather than guessing. OpenRouter's spend paid with your own provider key is also in that provider's own report, so it is kept as a marked row and counted once, at the provider that billed it. Providers now lists every provider Provantix knows: connected, not connected, or offering no usage report (GLM and DeepSeek offer an organization nothing to read; Gemini bills through Google Cloud billing, which is not built yet), each with the date its documentation was read, and the provider total says when it is partial. For a provider with no report, an admin can enter an invoice total by hand, stored as entered by hand with who and when and never added to the reported total. Amounts stay in the provider's currency; a displayed conversion uses a rate an admin entered and shows its date. Built from published documentation, not yet checked against a live account. See Providers.

Providers: what OpenAI and Anthropic themselves report. Connect an OpenAI or Anthropic organization with a read-only administrative key and Provantix pulls, hourly and on demand, what the provider says you used (by model, project or workspace, key and hour, with the cache and service tier) and what it charged (by day). The key is shown once, stored encrypted, can be replaced or revoked, and every use is audited. A provider can restate a past figure, and the restatement is kept as a new version beside the old one, each with the time it was fetched. The Providers page shows these figures on their own, labelled as the provider's, with how late they are and what each report leaves out. A provider project or workspace can be mapped to a department, so provider-reported spend shows by department even where no telemetry exists, with unmapped spend its own line. A key the provider rejects, or a connection that has not pulled for six hours, raises a finding. Built from each provider's published documentation and tested with responses shaped from it; not yet checked against a live account. Comparing these figures with Provantix's estimate is the next step. See Providers.

Claude Code and Codex as sources. Provantix now takes the per-call log events of Claude Code and OpenAI Codex at /v1/logs, as well as Claude Code's beta traces. A model call becomes a record with its tokens, the cache read and creation counts (Claude Code's input_tokens leaves the cache out, so the parts are added back, which the trace adapter did not do before), Claude Code's own cost_usd kept as reported by Claude Code beside the estimate, the person (their email, kept only as a handle) and the tags a team sets once per machine in OTEL_RESOURCE_ATTRIBUTES. A tool call becomes a record with its outcome. Nothing with content is read: no prompts, responses, tool input or output. Metrics are not taken, and the reason is written down: they are totals with no call behind them and would count spend twice. The same request seen as an event and as a span is recorded once. A source set up that has never delivered says "nothing received from this source yet". A new page covers Codex, the Claude Code page is rewritten, and What each source can carry now lists how far each source is proven. Both assistants are built from published documentation and have not been run from a live session; Codex's attribute names come partly from third-party descriptions.

Spend by any tag, tied to who did what, and budgets by tag value. The Cost page can group spend by any cost tag, the agent, the model or the provider, and nest a second dimension. Every row opens: to its agents, then to the calls, and each call opens as the same sealed record the audit side shows. Unallocated spend is its own row, and for a tag filled from the people directory it says how many calls name a person who is not in the directory. Each row shows the previous period and the change, the trend is drawn per day, and what a source itself reported sits in its own column beside the estimate. A budget can be set on a tag value (department=Sales), names an owner, and shows a forecast that states its method (the average of the last 7 complete days) or says "not enough data". The first time a budget reaches 50%, 80% and 100% of its cap, or its forecast says it will pass it, an event is recorded with the owner to be told. Investigate filters by any tag, Fleet can show only agents carrying a tag value, and the same breakdown is available as CSV, from the API and as the MCP tool get_cost_breakdown. The rows add up to the total, and spend per person follows the people settings everywhere, including the CSV. See Spend by tag, and who did it.

Every source carries tags and token detail, and says what it cannot. A new page, What each source can carry, lists for the SDK, the ingest API, OpenTelemetry, the hosted intake, LangChain over OpenTelemetry, the Langfuse and LangSmith imports and the MCP server whether each can carry the five cost tags, cached and reasoning tokens, a cost of its own, and the person. Every yes is a test through the real ingest, and the table is generated from the list the tests check. Admins can list attribute names their spans already use for tags. A standard tag name set as run metadata by LangSmith's OpenTelemetry export, OpenInference or OpenLLMetry is read, and those frameworks' span and token names are read as GenAI ones. The Langfuse import now also brings each model call, with token kinds, and both imports map their tags and metadata through a mapping you give them. What a source says a call cost (Langfuse, LangSmith, or provantix.usage.cost_usd) is kept as reported by that source beside the estimate and never replaces it. Settings has a Source coverage card: per source, the share of calls with each tag and token kind. Written from each project's published conventions, not checked against live exports.

People and departments: a directory that places each call in the department its person was in. Upload a file saying which department and cost center each person belongs to, from when. Settings shows what the upload would change before anything is saved, rejects bad rows with the reason, and records who uploaded, when and how many rows changed. A person who moves department on a date has spend before it in the old one; a person who leaves keeps their history. Calls that name a person not in the directory are counted and listed for an admin, never dropped, and calls that name no person are their own line. A department, and a cost center, then reach the cost tags as the person source, below what the call itself or its agent says. Provantix stores a handle (a one-way code), the department, the cost center, the manager's handle and the dates. It stores no name, email address or employee number. Spend per named person is hidden until your organization turns it on, and then only admins see it; a department with fewer people than the minimum you set (5 by default) is folded into one shared line. The directory can also be filled from your single sign-on (department and cost center the identity provider sends when someone signs in) and from directory sync (SCIM 2.0, with a token shown once that you can rotate or turn off). Where the ways disagree, the most recent row wins and the disagreement is listed. See People and departments.

2026-11-11

Cost tags: department, cost center, product, project, environment and any tag your telemetry carries. A call can carry the labels finance asks by. Send them as span attributes (provantix.cost.department and the other standard names, or provantix.tag.<name> for any other), through the SDK (tags), or in the record's ext. A name Provantix has not seen before becomes a tag by itself. A tag a source sends wins over one you set on an agent, which wins over a rule you wrote; each call keeps which one it came from and any value it overruled. Spend with no tag is shown as unallocated and never spread. You can give a tag a list of recognised values, and a value on no list is kept as sent and counted. A change applies from its date, and earlier calls can be re-tagged, with what they were kept. What a source stated stays in its record unchanged; what Provantix made of it is kept apart so it can be corrected. Limits (20 tags per call, 50 tag names) are counted when reached, never dropped silently. Settings has a Cost tags card. Splitting the Cost page and budgets by a tag comes next. See Cost tags.

Cost now counts cached, reasoning and batch tokens at their own prices. The estimate used to be input times one price plus output times one price, which was wrong for any call that used a cache. Provantix now keeps the cache read, cache write, reasoning and tier a source reports, makes input and output mean the same for every provider (Anthropic reports input without the cache, so the parts are added up), and prices each kind at its own rate. A price you have not set is never zero: that part of the call is shown as unpriced and the call is counted with the unpriced calls. The Cost page shows tokens by kind per model, the cache share over the calls that reported one (not recorded where none did), and an estimated cache saving. Model prices accept a cached input price, a cache write price and a batch multiplier. Records already stored are not changed; new Anthropic records show a higher input count than before because it is now the full total. Written from the providers' published usage fields, not yet checked against live responses. See Token kinds and how each is priced.

Provantix no longer ships any model prices. An earlier version put 11 list prices into every organization, and their per-million figures sat in a per-thousand column, so they were 1,000 times too high. They were invented figures, and an evidence layer does not invent a price, so they are removed. A new organization has an empty price table. Models you priced yourself are unchanged. A model with no price shows as unpriced, never as a guess. No call on the live platform had been priced from the removed rows, so no figure you have seen changes.

A Tools tab on the agent page. For each tool and model an agent called, across all its runs in the last 7 or 30 days: calls, failure rate, calls with no recorded outcome (counted apart), and the median and slowest-5% duration, each with the period before beside it. A tool whose calls carry no duration reads not recorded, never zero. Sub-agents' calls appear under their own name. A row is marked when its failure rate or slowest calls moved past a threshold; the marker says where to look, not what is wrong. See Tools and models.

A response loop on the Findings page, and follow-ups that say what they compared. For findings that first appeared in the last 30 days, per kind: the median time from the first record that showed the condition to the finding opening, to the first response (a person's action or an outside system's acknowledgement), and to resolution. Findings with no response recorded are counted apart, never as zero minutes, and nothing is scored. Follow-ups on quality findings now record the evaluation definition they compared, and read not comparable instead of improved when the model, dataset, rubric, evaluator version or configuration changed in between. Earlier follow-ups read as before. See Response loop.

The flag variant a step ran under is kept, and an evaluation is not compared across variants. If your spans carry OpenTelemetry feature flag attributes or feature_flag.evaluation events, the record keeps each flag and variant as your instrumentation reported it, and the run view shows them beside the step. Evaluation results accept a configuration, and a cohort whose recent window ran under a different variant than the one before reads not comparable instead of showing a fall or a rise. Existing records and results are unchanged. Provantix does not say that a flag caused anything. See Configuration.

Expected steps: say which tool a run should call, and see the runs that skipped it. An agent can answer fluently while a step it should have taken never ran, and nothing else looks wrong. On the agent's page, declare the tool its runs should call (and optionally the system). Provantix counts the ended runs since then that did not include it, shows them beside the runs that did, and raises Expected step missing, with example runs. A call that failed still counts as included, a run still in progress is not counted, and it cannot say why a step is missing. It does not stop an agent or require the step. See Expected steps.

Budgets for team and organization, a daily cap, and a roll-up. Budgets were monthly caps per agent or owner, set only through the API. You can now also set a budget for a team (a label on agents) or the whole organization, add a daily cap beside the monthly one, and see the organization, teams and agents as a roll-up on the Cost page with spend against each cap. Budgets and teams are set on the Cost page. Crossing a daily cap raises the same Spend over budget finding. A budget is a statement compared with recorded spend, to watch the cost of your agents; nothing is blocked. Existing agent and owner budgets read exactly as before. See Cost.

A Cost page. One place for what your agents are costing: the estimate for this month, the last 7 days or the last 30 days, the cost per minute right now, a breakdown by agent and owner, system, model, workflow and requester, each budget against its cap with the open spend findings, and the models with no price (the usual reason cost reads as not recorded). It reads the same figures as the Model usage and cost report and calculates nothing of its own. Cost is still an estimate, not reconciled to provider billing, and calls with no price are counted and never estimated. See Cost.

A failed span from a standard OpenTelemetry exporter is now recorded as a failure. OTLP's JSON encoding writes an error status as the number 2 or as "STATUS_CODE_ERROR", and the intake only recognised the word "ERROR", so spans that failed at the source were recorded as successes. New spans are read correctly. Records already stored keep what they were recorded as, so a source that sent failures before this change shows fewer failures in the past than it had. See OpenTelemetry exporter.

A Fleet page: what each agent is doing now, and what is being spent now. Every agent is a card with a state derived from records (waiting on a person, failing, working, idle, quiet, or its lifecycle), with the rule and the evidence beside it. "Waiting" comes only from an enforcement tool's recorded ask. Above the cards, the estimated cost per minute and output tokens per second, with the last hour as a bar per minute. Provantix reads records and answers nothing, controls nothing. See Fleet.

Claude Code traces become records. Claude Code's trace export (a beta) can be pointed at Provantix with a few settings: its API requests and tool calls appear as model and tool records, grouped by prompt, with a tool's success or failure carried from its execution span. It is built on the span names Claude Code documents and has not yet been run from a live session. Content is never taken, and what a person answered at a permission prompt is not recorded. See Claude Code.

Calls failing in a row, as a finding. When an agent's latest three or more calls in a trace or session all failed, with the first in the last 30 minutes, Provantix raises Calls failing in a row, with the tools, the error classes the records carry, and examples. It closes when a later call succeeds, and it says it does not know why. See An agent's calls keep failing.

The same call repeated, as a finding. When one agent makes the same call five or more times in a row in a trace (same tool and target, each within five minutes of the last, nothing different in between), Provantix raises Same call repeated, with the count, the span, the outcomes and examples. It says "repeated", not "stuck": a retry policy, a poll and a loop look the same in the records. An agent whose owner declared how often it runs is not reported for a run a minute or more apart. See An agent made the same call again and again.

A timeline on every run. The run view now places every step on one time axis, with a bar to scale where the source recorded how long the step took and a point labelled "not recorded" where it did not. Idle gaps are shown only after a step that recorded a duration, and the note under the timeline says when clocks disagree. See One run, as a chain.

Cost by workflow and by requester. The model usage report gains two tables: cost per trace and cost per requester handle. Unpriced calls are counted beside the priced cost and never estimated, calls with no trace are counted apart, and a requester is a keyed handle, never a name. See Reports.

New pages for findings you may meet. What to do when quality falls, when an enforcement tool denied an action that has a record, or when actions ran with no verdict, plus an Evaluations API reference and definitions for the new figures in Metrics definitions.

What an agent was delegated to do on its own, as of the time of each action. You can now record an agent's autonomy envelope: the mode (observe only, recommend, approval required, independently authorized), the environments it may change, the actions and resources it may act on, any emergency exceptions, and who states it. It is kept in the same append-only history as the owner and the reach, so an earlier action shows the envelope it ran under and says it has changed since; a later grant or revocation never rewrites it. A claim by the agent or the owner is shown as not corroborated, and "independently authorized" is accepted only from a named, corroborated source. An action from before any envelope was recorded reads as not recorded, which is unknown, not permitted or forbidden. Provantix records the delegation; it grants and revokes nothing. See The run view.

Approvals for an exact production change. An approval can now say the environment, the resource, the operation and a digest of the change it was given for, and the action says what it changed. A change to production with an approval for staging reads as not covered; the approval being given after the action ran does too; a value the action does not record reads as could not be fully checked, never as covered. An approval that came through an emergency (break-glass) route is recorded as one and is never shown as verified. The offline verifiers recompute all of it, with five new conformance vectors, so neither can be reported falsely. Nothing blocks an action. A verifier that predates this will not apply the new rules. See Approvals.

Evaluated quality: send your evaluator's results, and see what a model change broke. If you run your own evaluator or benchmark harness, you can send its results to Provantix. They are kept as its report, with the model version, dataset and rubric, and never with a prompt, an answer or data. A finding, Evaluated quality fell, is raised only when the same cases under the same rules show a fall of at least 5 points with ranges that do not overlap; otherwise the page says not enough evidence, source quiet or not comparable. For a model change, a case-by-case comparison lists the cases that passed before and fail now beside the ones that were fixed, so a better overall figure cannot hide a regression. Reviewer detection and operational impact are shown as not measured. Provantix does not detect hallucination, block an upgrade or certify the scores. See Evaluated quality.

Connected tools can report what they did about a finding. A gateway, policy engine or identity system that reads the findings feed can send back an acknowledgement: observed, escalated, approval requested, denied or credential revoked. It is kept as that tool's own report, beside the finding, and the Findings page shows it as "reported by" the tool. It is signed when the subscription has a webhook secret, safe to retry, and never changes the finding. Provantix did not do any of it and does not show that it happened; a finding no tool has reported on shows as unknown. See The findings feed.

The offline verifiers check control decisions. An export that holds an enforcement tool's verdicts, or actions that point at one, now gets two more checks in both the Node and the Python verifier: that each verdict agrees with the action that cites it, and that verdicts are well formed. A pointer swapped to another action, or a verdict id claimed by two records, fails even when every hash was recomputed. A missing verdict or one that names no action is reported and does not fail the export. The checks show that the records agree with each other; they do not show the enforcement tool was in the path or that its verdict was obeyed. Six conformance vectors cover them. See Control decisions.

Actions that ran with no verdict from an enforcement tool. You can now declare, on the Connectors page, that an enforcement tool covers changing actions on a system from a date. Provantix shows how many changing actions since then have a verdict from it joined to them (out of how many), and raises a finding for the ones that do not. It is your statement and Provantix does not check the deployment; it cannot say why a verdict is missing; the last two hours are left out so a verdict still arriving is not called missing. See Control decisions.

A decision can say which instruction it answers and which model made it. Decision evidence gains two optional references: the instruction or ticket the decision answers, and the model invocation behind it (model, version, invocation id). The run page shows them, or says they were not recorded. They are references only: the instruction and the prompt are never kept. Both verifiers check them when present, as part of whether a decision is well formed. See Decision evidence.

Control decisions: what an enforcement tool decided, checked against what happened. An agent gateway, a runtime security product or a policy engine can record its verdict (allow, ask, deny or steer) with the policy and version behind it, using the SDK. Provantix compares each verdict with the records that followed. A deny that still has a record of the action raises a new finding, Denied by a control, but it happened: high when the same action id is on both records, medium when the join is only probable. The run page shows the verdict beside the action, the incident timeline has a control decision stage, and the autonomy assessment counts verdicts apart from the agent's own actions. Provantix does not block anything, does not say the control failed, and does not report an action that ran with no verdict. See Control decisions.

2026-11-10

Autonomy assessment: evidence for a review of what an agent may do. A new Autonomy tab on each agent shows, for a window you choose, what the agent was granted, what it attempted, what a destination system confirmed (with counts out of how many, and a 95% interval) and what is not known, written from the data. It recommends one of four plain statements to the agent's human owner, with its reasons and the thresholds behind it. There is no score, and it changes nothing: Provantix grants, revokes and blocks no access. An owner can record a review, which stores the assessment as it was measured. See Autonomy assessment.

An approval is now judged against the policy version in force when the action ran. Before, an action's approval was compared with the policy version in force today, so setting a new version could flip the approval state of earlier actions from verified to not covered. Each action now reads the same after later policy changes. An action that ran before any version was registered reads as unchecked, with that reason.

Incidents: one timeline across an operational incident. An incident is a task id an agent put on its decisions, and the records that share a key with them: other records in the same trace, another agent's records in the same session, the destination system's own log about the same entity, and later reversals. The new Incidents page works it out when you open it (nothing is stored), shows each entry with how it was joined and how sure that is, says what is out of order, repeated, missing or contradictory, lists records that might be related but are not included, and downloads a bundle of exactly the incident's records that verifies offline. A person can add or remove a record with a reason; each correction is kept and counted. The page never says an action prevented or resolved anything. Also fixed: the run export download left out three files its own manifest lists, so a downloaded run bundle failed the offline file check; it now holds them all. See Incidents.

Why an agent chose an action, kept as evidence. The SDK can record a decision before the agent acts: the hypothesis and the evidence it cites, the candidate actions, the one selected, what was rejected and why. The action's own record points back at it. The run view has a new card, "Why the agent chose this", and badges that say whether the decision and the action agree. The offline verifier checks the same thing from the records alone, with two new checks. A decision is what the agent said it considered; it does not show the reasoning was sound or that the action happened as chosen, and anything the agent did not say reads as not recorded. See Decision evidence.

2026-11-09

Each transparency log head is time-stamped by an outside authority. The hash of every signed head is sent to a public RFC 3161 time-stamp authority and the signed answer is published beside the head. It shows when the head existed, from a party we do not control; it does not show the log is complete or true, and it is not a copy of the head. If the authority is down the head is still signed and says the time-stamp is pending, and the mirror tool and the export verifier report that as not anchored, never as failed. The page now has a step-by-step check with openssl ts -verify, and the verification kit includes a real token to practise on. See The transparency log.

A feed of finding events for the tools that enforce. Subscribe by pull (a cursor and a credential) or by signed webhook, and receive an event when a finding opens, rises, is acknowledged, resolves or is followed up. Events name the finding and link back to it; they carry no record content. Provantix reports; your gateway, policy engine or identity system decides. See The findings feed.

Release 1 documentation. A new page, Destination confirmation, explains the five states, how an action is matched, what makes it pending or unverified, later outcomes, approvals, and what none of it shows. The self-hosted pages cover the new LEDGER_CREDENTIAL_KEY setting, the outbound access a destination connection needs, and backing the key up. Screenshots of the run view, Coverage and Connectors are refreshed.

The two caveats sit beside every verified result. Every verifier (Node, Python and the browser page), the run view's verify panel, the export's new README.txt and manifest, the verification kit's guide and the Trust page now say the same two things in the same words: a valid signature shows who signed, not that what is signed is true; and integrity is not completeness. A test fails if any copy drifts, and another fails if a blended trust score, grade or evidence-quality number is ever added to the app, API, MCP, Brief or exports. See Trust.

Completeness is always scoped. The Brief no longer makes a flat claim about reach or about the evidence being whole. It says what was checked and over which period, and that a pass shows what was delivered was not altered, not that every action was captured. The coverage and integrity statement gains a section naming what was reconciled against a destination system and for which period. An export's manifest has a scope_statement. The AI Insider refuses a completeness claim with no scope, and a test fails if a catch-all completeness phrase appears anywhere the product speaks. See Trust.

A reconciliation scope can name a tool. A refund is recorded as a tool call (call_tool) with the tool issue_refund. A scope now matches either the tool name or the action type, so it finds the records real agents write. A stand-in Razorpay test agent (razorpay-refund-agent.mjs) is added so the whole path can be tried on a test payment. See Connect Razorpay as a destination.

Changes nobody claimed. Provantix now also looks from the destination's side. A refund the payment system holds, in a period it was collected for completely, that no agent claimed is listed on Coverage under Changes nobody claimed, with who the destination says made it (if it says) and the agents active then, and raises the finding unclaimed_destination_change. The cause is not guessed. A period that was not collected completely is reported as incomplete and the finding is withheld, never "none". See A change nobody claimed.

Connect Razorpay as a destination. On the Connectors page, Destination systems lets you connect a Razorpay test-mode key pair. Provantix then reads Razorpay's own refund records every five minutes and compares them with what the agent claimed, so each refund becomes destination-confirmed, conflicting, pending or unverified. A Razorpay key cannot be limited to read-only, so only test-mode keys are accepted for now (Provantix itself only ever reads). The secret is stored encrypted and shown only once. A destination that is down or throttled shows as health, never as a failure of the agent, and the period is recorded as not collected. Revoking the key (in Provantix or in Razorpay) turns the affected actions unverified, with the reason. An earlier version of this entry described a connection to a different payment provider, which is not part of Provantix and has been removed. See Connect Razorpay as a destination.

Exports carry reconciliation results, approval binding and later outcomes. Export format 1.1 adds an optional reconciliation.json: for each reconciled action, the state, method, confidence, the destination record it rests on and where it came from, the collection window and the scope statement; whether each approval covers its action; and later reversals, retries, disputes, reopenings and overrides, kept apart from the action. It also adds authority.jsonl, what each agent was allowed to do at the time of its action. The Node and Python verifiers check the new section (structure, consistency with the destination record carried, approval coverage) and agree on every test vector; altering a result breaks verification even when every hash is recomputed. The verification kit is rebuilt with a reconciled sample and a tampered one, and its checksum is republished. A 1.0 export still verifies. The verifier does not show that a destination's record is true. See How verification works and the verification kit.

The run view, Coverage, the Brief and Home show what the destination system recorded. For actions in a reconciled scope, the run view shows each action's state (destination-confirmed, conflicting, pending, unverified or agent-reported) with what it was checked against, the period, how it matched and what would change it. Coverage has a new block, Confirmed by the destination system, scoped to one destination and period, with the counts and the three unresolved groups. The Brief counts actions the destination contradicts (attention) or has not confirmed (watch), and Home's heat map has two new rows. Investigate has a new field, Destination check. Pending and unverified are never treated as a failed action, and nothing is combined into one score. See One run, as a chain and Coverage.

Later outcomes are linked to the action. A reversal, a dispute, a reopened ticket, a human override or a retry can now be attached to the action it is about. It shows on the action in the run view, in time order, and Later outcome in Investigate finds the actions that were later reversed or disputed. The action's record, its hashes and its reconciliation result do not change. A link records that something happened later, not that the agent was wrong. Retries with the same action id count as one action. See One run, as a chain.

A claim that names its own refund is no longer matched to another one. When the agent's claim cites a refund id, or carries an action id, a different refund for the same payment and amount is never taken as its match.

Authority at the time of each action. The run view and the record drawer now show what the agent was allowed to do when the action happened: owner, lifecycle, risk tier, approval-required systems, declared and granted reach, and identity mappings, as of that moment, with what has changed since. Changing an owner, a reach or a policy today no longer changes what earlier actions show. An action from before the history was kept says "authority at that time not recorded". The run export carries it in authority.jsonl. See One run, as a chain.

Approvals can say what they cover. An approval being present no longer reads as the approval covering the action. If your application sends what an approval was given for (the target, an amount limit, an expiry, the policy version), the run view and the Home approvals card show approval present and authority verified separately, with the reason when the approval does not cover the action, and a new finding, approval_does_not_cover_action, is raised. Approvals that say nothing about what they cover keep working and read as "what it covers is not recorded", never as verified. See Approvals.

Plainer wording about what a check shows, and about LangSmith. The FAQ and the Trust page now say the two limits of a passing check side by side: a valid signature proves who signed, not that the content is true, and integrity is not completeness. The FAQ describes observability tools as recording traces and administrative activity, not as a lesser record. LangSmith import is described the same way everywhere on the site and in the docs: rolling out, enabled on request, while Langfuse import is available.

2026-11-06

A missing batch says whose it is. The Missing batches drill-down no longer shows only the end of an id. The source layer is named by the name you gave the emitter or collector ("Cruz"), and a batch's detail has a new Whose line: the emitter or collector, and the agents it delivers for, most records first, with a count of records that name no agent. Inside a finding, where the source layer is skipped, the top tile names the source. The agents are the ones the source has delivered for, which is where the missing batch would have gone. See Batches never arrived.

Every finding says when it was reported. On the Findings page the status now has the date and time the finding was first reported, with how long ago, beside it: "reported 27 Oct 2026, 10:20 AM · 10d ago". Hover for when it was last seen, and when it was resolved. On Home, each row of the Brief's "Read these first" list shows the same, and the attention queue and What to do next say "reported 10d ago" with the exact time on hover. The Brief over MCP carries first_seen_at on its data gaps and next actions.

Search in the documentation. The portal has a search box in the header: press / or Ctrl/Cmd+K, type a topic, a setting name or a question, and it lists the matching pages and the exact headings, with the matching words marked. Arrow keys move, Enter opens, Esc closes. It searches every page, runs in your browser, and sends nothing anywhere. Every heading now also has an address of its own, so a link ending in #heading opens the page at that heading.

The "Connect" page for an AI assistant shows the Provantix logo again. When the page was reached through the MCP address, its logo was a path that address does not serve, so it showed a broken image. The logo now loads from the app's own address.

The transparency log signs a head every time it grows. Until now a head was signed only when someone asked for the latest one, so a size nobody looked at had none. The log now checks every hour, and signs a head whenever it has grown since the last one, so /heads lists a head for every size the log has been at, to within an hour. See The transparency log.

Follow up on a finding: did the decision work? Under any finding, Follow up on this records who is responsible, what was decided or done, and how long to watch (3, 7, 14 or 30 days), together with the latest human decision on the finding and the condition as it was. When the window has passed, Provantix reads the checks it already ran and records Improved (cleared, stayed clear, did not come back, over enough healthy checks), Not improved (still there, or it came back; an acknowledgement is not a remedy) or Not enough evidence (too few healthy checks, a source in error, or cleared too late to know). An outage is never read as an improvement. The finding's own lifecycle is unchanged. An outcome says what happened after the decision, not that the decision caused it, and every follow-up says so. Setting or cancelling one appears on Decisions. See Findings and what to do about them.

2026-11-05

AI Insider wording is checked against the insight it is about. A rewrite by the model must now rest on that insight's own figures, names, time window and explanation, not only on the figures of the whole analysis. A real number attached to the wrong agent, a name that belongs to another insight, a time window the plain wording never used and a cause the insight does not give are each refused, and the insight keeps its built-in wording with the reason recorded. The model is also given each insight's explanation so a cause can come from it. The same checks apply to Explain on an insight. See AI Insider.

The log mirror keeps signed heads, and says what is and is not independent. verify-log.mjs now keeps every tree head it sees exactly as signed, checks on each run that every head it kept is signed by the log's key and is still the start of the log, and, if history was rewritten, writes the signed heads that no longer match to a .conflict.json file you can hand to anyone. The log also lists every head it has ever signed at /.well-known/provantix-log/heads, and the verification kit includes a GitHub Actions workflow that runs the mirror daily from a repository you control. The transparency log page now separates publication, independent retention and verification, and says plainly that no party outside Provantix is known to be retaining heads yet.

2026-11-04

Landing pages with a "Get an invite" form, and cookie-less site counts. Four pages on provantix.tech (audit trail, who authorized this agent, compliance evidence, and for auditors) each carry an invite form: name, an email confirmed with a one-time code, whether you want a demo, and the latest time you can take one. Requests land in the Applications queue. The marketing site and this portal now count page views, kit downloads and clicks in aggregate, with no cookie, IP address or identifier, and nothing is sent if your browser sends Do Not Track or Global Privacy Control. See the Privacy Policy. Customer-run deployments send nothing.

A public transparency log for daily roots and export manifests. Each closed day's root and each signed export's manifest hash is now committed to in a public, append-only log (a Merkle tree with signed tree heads), under an opaque organization reference. Exports carry proofs in transparency.json, and the verifier checks that the manifest hash and every root are what Provantix logged (pass the log's published keys with --log-keys). A new tool, verify-log.mjs, recomputes the whole public log and fails if history was rewritten since your last run. This is a commitment by Provantix until others mirror it, and heads are not yet anchored outside Provantix; the page says so. See The transparency log and the updated Trust page. The verification kit now includes a sample with a log proof and the mirror tool.

Reports export as PDF (live). Export PDF on any report page downloads it as a PDF: the headline numbers, the summary, the over-time table and every table (up to 200 rows each; the CSV still has them all). The last section says the PDF is a readable copy, not signed evidence, and gives a data fingerprint (SHA-256 of the report's figures) so it can be matched to a CSV or a later run of the same period. See Reports and dashboards.

A run exports as a one-page PDF summary (live). PDF summary on the run page downloads a readable summary of the run: who requested it, the agents that acted, each step in order with its outcome and any approval, and what is and is not known about the destination. It is not signed evidence and says so; it carries a fingerprint of the run's record hashes that the run's export bundle also holds. See One run, as a chain.

Approvals: who approved what (live). Home has a new card, Who approved the sensitive actions?, with the approved and unapproved counts, the approvers (as handles) and the agents with the most actions without an approval. Investigate has an Approval column and Approval and Approver in Value counts, so you can filter to actions with no approval or to one approver. The AI Insider reports sensitive actions with no approval on record, and says so plainly when no approval has ever arrived, because Provantix stores an approval only when your application sends it with the action. A new report, Approvals and authorization, lists approvers, a per-agent share and the actions with and without an approval. Setup shows how to send approvals from the Node SDK and from OpenTelemetry. See Approvals: who approved what.

2026-11-03

Single sign-on with any OpenID Connect provider (live). The Single sign-on card now offers Other OpenID Connect beside Microsoft Entra and Google, for Okta, Keycloak, Microsoft AD FS, Ping Identity, Auth0 and similar. Enter the provider's issuer URL; the provider needs the authorization code flow, RS256-signed id tokens and the email address in the token (in email, or preferred_username or upn). The same domain rule applies as for the other providers. See Sign-in and SSO.

Node SDK 0.3.0 is published, and can lift the queue limit (live). new AgentEmitter({ spoolMaxBytes }) sets how much the on-disk queue may hold while Provantix is unreachable; the default stays 2 GiB, and null removes the limit.

Coverage charts and status colours are readable in both themes (live). The status chart and the status badges on Coverage had drawn in black on the dark theme; they now use the same colours as the key, and the counts in the headline carry the colour of the state they open. See Coverage.

The Enterprise plan lists customer-run deployment as planned. The pricing page now says that on-premises or private-cloud deployment, and dedicated hosting, are planned for Enterprise, and no longer shows a "From" price for it. Nothing about the hosted service changes. Customer-run deployment is not yet available; if you need it, talk to us before relying on a date. What is decided so far, with the planned steps, is in the Self-hosted (planned) section.

Behind the scenes (no change to the hosted service). The platform now has a deployment mode and a single place that decides whether telemetry is accepted: on the hosted service, as before, a lapsed or unpaid subscription is refused with HTTP 402 and using more than your plan includes is billed, never refused. This is groundwork for customer-run installs and changes nothing you see.

Coverage speaks plainly and links to the evidence (live). The page is headed Evidence coverage. Attest is now Mark as expected and Exclude is Exclude from coverage, with a note that excluded activity stays visible in Investigate. The Seen by column is Source: the agent's own report, a gateway or intake, or an independent source. Each pair has View records, which opens Investigate for that agent and system, and an agent's Reach tab links to Coverage filtered to that agent. The header's "coverage gaps open" now reads "delivery gaps open" and links to Findings, because it counts missed batches. See Coverage.

The coverage and integrity statement carries the conclusion (live). The report now lists each expected connection with who expected it and since when, what observed it, the attributed count, and the known delivery gaps of its sources within the period ("unknown" where no source can be tied to it), and lists run-id system names separately. Coverage links to it. "Coverage gaps" in the report are now called delivery gaps, and "attested" is now "marked as expected" / "decisions". See Coverage.

Declaring reach and marking a pair as expected are one act (live). Mark as expected on Coverage now adds the system to the agent's declared reach, so the agent's Reach tab lists it with who declared it and why; declaring a system on the Reach tab (with an optional reason) marks the pair as expected on Coverage with the same origin, "declared by the owner". Removing a declaration removes it from both, and declaring a pair again lifts an exclusion. Both routes write the same entries to the audit log. Pairs already marked on Coverage were copied into declared reach.

Each Coverage row shows four facts (live). A pair now shows how it came to be expected (and since when, by whom for pairs marked on the page), what observed it, how many of its records are attributed, and the known delivery gaps of the sources that supplied it, or "unknown impact" where no source can be tied to it. See Coverage.

Coverage opens with a question and a sentence (live). With no expected access confirmed, the page asks which systems you expect your agents to use, says how much activity was observed, and offers Review observed access and Add expected access; the add form moved up from the bottom. Once access is confirmed, the headline is one sentence ("Of your N expected connections: X have independent evidence, Y rely on the agent's own report, Z have no evidence") and each count opens exactly those pairs below. The definitions and the 0.8 threshold sit under How coverage is calculated. With an agent selected, the headline and chart follow it. See Coverage.

Coverage leaves run-id system names out of the count (live). When an agent sends a different system name for every run, those names are no longer listed one by one for you to confirm and no longer enter the denominator. A Naming problem card on Coverage shows, per agent, how many distinct names there are and points to the fix (a stable action.targetSystem); the headline says how many were left out. Records are unchanged. See Coverage.

2026-11-01

One single sign-on button (live). With an email typed, Continue with single sign-on tries that organization's own provider first and falls back to Provantix's Google; with none, it goes to Provantix's Google. An organization that has switched on its own single sign-on is refused through Provantix's Google, with a message telling its people to use their own. See Sign-in and SSO.

Continue with Google, managed by Provantix (live once the Google client is configured). The sign-in screen offers Continue with Google, with no company setup and no email to type first. An admin adds a person from Members with a role; Google proves who they are; a person who already has a password account is signed into that same account. Anyone no admin has added is turned away. See Sign-in and SSO.

Single sign-on is invite-only by default, and a staff domain can have an owner (live). Each organization sets up SSO once with its own identity provider; by default only people an admin has invited can sign in through it, and an admin can switch that off. A Provantix staff domain can now name an owning organization that alone may use it for its own SSO, so the operator's own organization is no longer blocked by its own staff domain. The sign-in screen now says SSO needs the work email first. See Sign-in and SSO.

Coverage explains itself (live). The top of the Coverage page now says what it tells you, uses one of your own pairs as the example, and gives a one-line key to the five states.

Home shows a heat map of sources by category (live). The source-by-category grid is now a heat map on one scale: the darkest cell in a row is that row's highest count, none is flat grey and not measured is hatched. See Reading the Home page.

Investigate shows value counts and opens records in place (live). Beside the timeline, Value counts lists the values of each field for your window and filters, with a count and a share bar; a click on a value filters by it, excludes it, or redraws the new chart above the table split by it. Filters are chips carried in the URL. A row opens in place to its stored properties, with content never stored shown as not stored and erased fields shown as erased. See Records and Investigate.

Location on records (live). Records can say where the agent ran and what it acted on, as short codes labelled declared or derived, never verified, and counted in their own group in Investigate. The person's country is kept only where an admin turns on Settings, Where people are, and goes with the person's other data when they are erased. Nothing finer than a country and region is kept. See the record schema. Node SDK 0.3.0, now published, adds a location option and reads the cloud region it runs in; with older SDKs, location arrives through OpenTelemetry attributes and the Cloudflare path.

Cloudflare Workers can send into Provantix (live). A Worker's outbound calls arrive through the hosted intake as agent actions, from a source marked as an agent. See Sending data to Provantix.

The sealing notice sends itself (live). Provantix emails each organization's admins the notice, starts the 90-day clock, and sends reminders at 30 and 7 days and on the day refusal begins. Nobody sets the date by hand. See The sealing deadline.

Node SDK 0.2.0 seals every batch (live). @provantix/agent-sdk 0.2.0 is on npm. The emitter signs the hash of every record in each batch and adds it to the envelope, so an auditor can check, with the emitter's public key and without trusting Provantix, that your emitter vouched for each record. It still verifies after a person's data is erased. Sealing is on by default and batches queued by an older version are sealed when 0.2.0 sends them. Upgrade with npm install @provantix/agent-sdk@latest. See The emitter proof and Node SDK.

Per-record proof (live). The platform accepts sealed batches, keeps each seal, and exports them with the evidence. Every verifier (the standalone Node and Python scripts, the MCP verify_export tool and the browser Verify page) now reports whether a record was vouched for by its emitter, and says whose key it was checked against. Setup shows each emitter's proof status and has Download public keys. A key Provantix generated is labelled as a Provantix attestation, which suits a pilot or trial; register your own key for production.

Sealing deadline (live). From a date Provantix sets and announces, a batch that is not sealed is refused and the source is paused until it is upgraded. Pilot and trial sources are exempt. Two findings, Sealing deadline approaching and Telemetry paused (SDK upgrade needed), say which emitters are affected. See The sealing deadline.

The AI Insider reports agent state and proof (live). Insights now say whether an agent is healthy, degraded, uncertain, unknown or inactive, and say what is missing when the data cannot show it. A new proof insight reports how much of your telemetry is sealed, whose key signed it, which emitters have not upgraded, and how long remains before the deadline.

Documentation: Sending data to Provantix. One page now says which way in fits what you have (the SDK, the OpenTelemetry exporter, direct HTTP, the hosted intake, connectors or imports), what each proves, and covers the hosted OTLP intake in full: the endpoint, creating a token, what is kept, which agent a span belongs to, the limits and the responses, with an OpenTelemetry Collector example. Cloudflare Workers export is listed as planned.

2026-10-27

AI Insider insights name their emitter. Each insight now says which emitter its figures come from, with each emitter's share where several send data, and Show me sets the Source filter to it. A new data-quality insight names an emitter whose records carry no agent. A change on last week is no longer claimed when the week before is only partly there, insights are ordered by importance, and the AI wording is requested with a lower reasoning effort so it is no longer empty.

The AI Insider gives insights. The panel now leads with what your data says: a short overview, then up to six insights across activity, reliability, cost, data quality, governance and change, each with its figures, a Do next line, and Show me, Explain and Dismiss. The decisions that change your records stay below, and the outside alerts that duplicate the Findings page fold into one line. An AI model rewrites the insights once per analysis, checked so it cannot add or change a number, and the built-in wording is used when it is off or the budget is used. See The AI Insider.

The AI Insider's schedule and time zone (live). Scheduled analyses now run at 10:00 and 16:00 in the organization's time zone, instead of every twelve hours from the first run. The zone is taken from an admin's browser the first time they sign in, an admin signing in from somewhere else is asked whether to follow it, and it can be changed under My account. The daily allowance of manual analyses resets at midnight in that zone. See The AI Insider.

The AI Insider (live). A light bulb on the right edge of every page shows recommendations waiting for you. Open it to approve a change inside Provantix, acknowledge advice about your own systems, deny, or ask for an explanation. It can be pinned to dock beside the page. Analyses run twice a day, and Run AI Analysis starts one on demand, with a daily allowance by plan. Items are drafts, never part of the audit record. See The AI Insider.

Early warning: missed runs (live). An owner can now say how often an agent should run, under Expected to run on the agent's page. When the agent shows no activity for longer than that, a Missed run finding opens, and closes when it runs again. See An agent missed its run.

Early warning: trends. Two findings compare an agent's last 24 hours with its own previous 14 days: Error rate rising and Latency rising. Home's source grid gains a Workflows at risk row that counts these and missed runs. An email alert rule can now go to the owner of the agent a finding is about. See A workflow is failing more or running slower.

Settings (live). Two cards are new, for organization admins: Model prices (add a model's per-million-token rates; past records for that model are repriced on save) and Business hours (days, start, end and time zone for the off-hours finding). Both were API-only before. On Setup, Delete now appears for a retiring emitter with no records, not only a revoked one.

Documentation. New pages for situations seen in real use: a new source's records have no agent, signal never observed, off-hours activity, a registered emitter that never shipped, an export that is consistent but not authentic and your AI assistant sees no data. Sources, filters and comparing covers naming sources, the Source filter, Compare and snapshots. A new Worked cases section walks through two real readings. The batch gaps and cost pages are expanded.

MCP access for admins (live). An admin has all access inside their own organization over MCP: their connection includes cost:read automatically, marked "included for admins" on the access page, whether or not Allow cost data over MCP is on. That setting now governs members only. Roles are read on every call, so demoting an admin removes cost from their connections on the next call. The access page also now names the organization it will read, which it printed blank before.

MCP server (live). Nine tools are new: list_sources, compare_sources, get_run, get_cost, list_candidates, list_systems, list_decisions, run_report and verify_export. Every other tool except whoami accepts a source argument to read one source at a time. Standing resources (provantix://brief, coverage, open-findings, sources, systems and a per-source Brief) and five ready-made prompts are available. The server stays read-only.

2026-10-25

MCP server (live).

  • The connection requests three scopes by default: evidence, findings and coverage. Cost is a fourth scope that a client must ask for, and an organization admin must allow it under Settings, Connected AI clients.
  • Two tools are new: get_brief (the organization's Brief) and list_approvals (approval references, filterable by record, agent, system or a range of up to 90 days).
  • summarize_activity adds a metrics block, field completeness, and a comparison with the previous period of the same length. It can also return one agent's hourly baselines.
  • A client that asks for more scopes than you approved before shows the access page again, with the new scopes marked.

Sealing time (live). A day is sealed 72 hours after it starts, which is 48 hours after it closes. The earlier code sealed a day 24 hours after it closed. Days that are already sealed keep their original creation times.

Findings (live). The findings page now lists every finding kind the platform produces, grouped by family. Two kinds that were missing a label now have one: "Target system named by a GUID" and "Sensitive action without approval".

Wording. The docs and site no longer describe full prompt capture as a tenant option. The hosted intake stores only a hash and the byte length of prompt and completion text. Cost is described as an estimate that changes when the price table changes. Attribution is described as method plus confidence, with unresolved records kept for review.