Unisona 1.8 — the personal reasoning cockpit goes public

What's new, by loop stage

The system is one loop with four core objects (Memory, Task, Tool, Convergence Record). 1.8 doesn't add a stage — it gives each existing stage a surface you can see and drive.

🔭 Observe — the Explore feed

A single ranked feed (/explore.html) of things worth your attention — games, reading, builds, docs, beliefs — every card carrying a why and a source per the Verify rule. The 1.8 release itself appears here as a card.

🧠 Remember — confidence-decay memory (#1422)

Memory that forgets gracefully. A fact's confidence decays over time unless it's reinforced, so stale claims fade instead of being asserted forever. Persistence is still append-only JSONL + CSF archive — nothing is deleted, confidence just shifts.

💭 Reason — the personal cockpits

• Financial reasoning cockpit (#1434) — reasons over your own numbers and

shows the evidence behind every answer.

• Personal preference model (#1426) — retrieval-based: grounds

recommendations in what you've actually said before, instead of guessing.

• Learn-anything tutor (#1438) — an interactive tutor for any topic that

tracks what you actually retain over time.

⚙️ Act — autowork you can steer

In-chat Approve / Rework / Discard for autowork draft PRs (#1503), with the Σ₀ council wired into self-coding runs so convergence Δ accrues on real decisions (#1598). The harness is the product — you stay in the loop.

✅ Verify — grounding you can read

• Fact-check button (#1430) — paste a claim, get a grounded verdict with

sources.

• Grounding Diff viewer (#1420) — see claim · evidence · source · confidence

side by side.

• Drift canaries (#1428) — passive observability that watches for behavioral

drift and surfaces it as a signal.

• Symptom journal (#1435) — calibrated, cite-or-abstain honesty about health:

it cites a source or abstains rather than guessing.

• The Σ₀ council exec-verify now defaults ON (#1640) and grounds its verdict

on a real execution check, closing the refuted→retry loop in chat (#1568): a run test is ground truth, and execution overrides text.

🎯 Converge — calibration and replay

• Decision journal + calibration (#1436) — log a decision with your predicted

confidence, record the outcome later, and get scored on whether your gut calls are actually good.

• Convergence replay / time-travel debugger (#1419) — step back through the

convergence records to replay exactly how a conclusion was reached.

• Verified-patch / honesty / route-quality metrics (#1411) are now recorded

as first-class convergence signals.

Under the hood

• Brand — home rebranded with canonical + social meta (#1657); every page's

chrome, nav, and footer rebranded to unisona.ai (#1661); the Triagon low-poly signal mascot added (#1604).

• CSF — ships the CSF-Col transform and corrects compression on the live

paths (#1601); the own-the-model PLT training package lands with ADR-0011 (Proposed) (#1645).

• Stock trader — fee-aware trade-EV gate (#1648), surfaced market news

(#1582), and a global header with a working light/dark toggle (#1579).

• Chat — production-grade, continuous voice-to-text dictation (#1607).
• Security & hygiene — automation git helpers and self-edit/auto-version

steps routed off the shell through lib/safe-exec.js; the consolidation linter made multiline-aware to catch interpolated execSync sites.

Point releases since 1.8.0

1.8 shipped on 2026-06-30 and kept moving. The 1.8.1–1.8.7 point releases added real user-facing surface and tightened the dev loop underneath:

• Cloud-free model crystallization (1.8.7) — the Σ₀ continual-training flywheel can now

improve the local looped coder without any cloud teacher. The verified-capable local Qwen2.5-Coder-7B acts as the teacher: it solves coding tasks under the Σ₀ prompt, each solution is executed against its asserts, and only the green ones are distilled into the small Ouro-1.4B student — then eval-gated on HumanEval before the live adapter swaps. One box, no key, no rate limit; teacher quality is self-limiting because only code that runs trains. See the crystallization writeup.

• Three Doors, regenerated (1.8.6) — the Kingdome of Hearts Explore game is

reworked end to end: every scene keeps a fixed theme + meta-lesson while the prose is regenerated per visit, each turn offers exactly three randomized doors, and scenes render real art (curated Kingdome concept art with a gpt-image-2 fallback) instead of placeholders. The guide is Lantern throughout.

• Kalshi terminal reskin (1.8.5) — the swipe deck adopts the shared

unisona.ai chrome, a polished device-console frame, and a collapsible account · profile · positions HUD showing the live Σ₀ council verdict; it now passes the WCAG a11y audit.

• The grounded Kalshi trader (1.8.4) — since momentum has no edge after fees,

the profitable arm chases information the thin market hasn't priced: it web-grounds a cited P(YES) on near-term event markets (weather first) and takes a position only when the edge clears the fee hurdle, graded forward by Brier score rather than asserted.

• One brand, cleaner home (1.8.3) — shared chrome, the Create page, and stale

body copy all read unisona.ai; the home demotes its heavy panels to a compact link row; and a WCAG font/style contract runs in CI so contrast regressions can't ship.

• Workflow hardening (1.8.1–1.8.2) — dynamic per-contributor PR lanes and an

assigned-issue convergence merge gate (#1755), with the dual-boot quickstart now installing the workstream hooks automatically.

The principle behind 1.8

Nothing is accepted without evidence. Every important claim carries [claim, evidence, confidence, source].

Every cockpit in this release ships under that External Reality Rule. None of them assert — they cite, or they abstain. That's the whole product: a loop you can watch ground itself.

Observe. Remember. Reason. Act. Verify. Converge.

See also: Changelog · What's New · Σ₀ Briefing