# ThunderGate 1.3.0 — What's New for You

*August 2026 — "The Source of Truth Release"*

ThunderGate 1.3.0 is about one thing: **the difference between "I checked" and "I
remember."** Every assistant ever built says both in the same confident voice. This release
makes them structurally different — and puts that difference in front of your agent before
it speaks. Here's what that means for you, in plain language.

---

## The Sources of Truth Map

Your agent now keeps a register of **where truth lives** for the subjects it speaks about —
which file, which system, which person is the authority on each thing, and crucially,
**your ruling on which source wins when copies disagree.** That ruling is yours alone: the
agent cannot write it, and it renders in your own words every time the subject comes up.

When a registered subject enters the conversation, the agent sees a compact card — not the
answer, but the **address** of the answer: where to read it, when it was last actually
read, and which source is entitled to control the claim. Then it goes and reads before it
asserts. A fact it verified this minute and a fact it remembers from last month no longer
sound the same, because they no longer *are* the same.

- **Freshness is a word, not arithmetic** — FRESH, AGING, STALE, or NEVER-READ, right on
  the card. And for facts about the world that change without warning (a person moved, a
  policy changed in a meeting), the card says the honest thing instead: *observed on this
  date — change may be unobserved.*
- **Copies don't count as evidence.** A database, its dashboard, and a report exported from
  that dashboard are one witness, not three. The map counts independent origins only.
- **Nothing becomes permanent unasked.** New entries wait for your review, and a large
  batch of new entries deliberately holds for your explicit go — a mass-written register
  cannot bless itself overnight. Retired sources are marked dead with a reason, never
  silently deleted: *"we used to trust this, and here's when and why we stopped"* stays on
  the record.

## Calendar answers that say how far they looked

Ask about your schedule and the agent can now look as far ahead as you need — weeks or
months, across **every** calendar on your account, not just the primary one. And every
calendar answer now states the window it actually searched, so a quiet week in the reply
means a quiet week on the calendar — not a question that quietly stopped looking next
Tuesday.

## Honesty, everywhere we could put it

A theme of this release: wherever the runtime reports on itself, the report has to be
earned.

- **Green means ran-and-passed.** Health checks and test summaries can no longer show green
  for something that was skipped. If a check didn't execute, the summary says so.
- **Self-grades need evidence.** A background component can no longer rate itself healthy
  just because it was busy. Volume is not quality, and the reports now know the difference.
- **Your numbers are yours.** The safeguard that corrects the agent's own cost and usage
  figures now cleanly distinguishes those from numbers *you're* discussing — budgets,
  limits, plans — and leaves yours alone.
- **The audit trail doesn't cry wolf.** Routine nightly bookkeeping no longer logs itself
  in language that sounds like data loss.

## Quality of life

- A reviewed batch of register entries can be approved with a word — your say-so promotes
  them, and only your say-so.
- The agent's session can be reset on your instruction without waiting for the nightly
  cycle.
- Long notes in the learning ledger are stored whole instead of clipped.

---

*As always: your hardware, your data, your rules. The full operator's manual — every
command, every setting, and an honest section on what is not built — is at
[/thundergate/README.md](/thundergate/README.md).*