Second Opinion v0.5.0

Second Opinion v0.5.0 - read-only Windows crash triage. It changes nothing on the machine.

Download SecondOpinion-v0.5.0.zip below, Extract All, then double-click Run Second Opinion.cmd.
The ZIP contains that launcher, src/Invoke-SecondOpinion.ps1, and the README.

Verify your download (optional). SHA256SUMS.txt holds the SHA-256 of the ZIP. In PowerShell:

Get-FileHash SecondOpinion-v0.5.0.zip -Algorithm SHA256

and confirm the printed hash matches the line in SHA256SUMS.txt. The README's
"Verify your download" section has a copy-paste one-liner that does the check for you.

Second Opinion v0.4.1

Second Opinion v0.4.1 - read-only Windows crash triage. It changes nothing on the machine.

Download SecondOpinion-v0.4.1.zip below, Extract All, then double-click Run Second Opinion.cmd.
The ZIP contains that launcher, src/Invoke-SecondOpinion.ps1, and the README.

Verify your download (optional). SHA256SUMS.txt holds the SHA-256 of the ZIP. In PowerShell:

Get-FileHash SecondOpinion-v0.4.1.zip -Algorithm SHA256

and confirm the printed hash matches the line in SHA256SUMS.txt. The README's
"Verify your download" section has a copy-paste one-liner that does the check for you.

Second Opinion v0.4.0

Second Opinion v0.4.0 - read-only Windows crash triage. It changes nothing on the machine.

Download SecondOpinion-v0.4.0.zip below, Extract All, then double-click Run Second Opinion.cmd.
The ZIP contains that launcher, src/Invoke-SecondOpinion.ps1, and the README.

Verify your download (optional). SHA256SUMS.txt holds the SHA-256 of the ZIP. In PowerShell:

Get-FileHash SecondOpinion-v0.4.0.zip -Algorithm SHA256

and confirm the printed hash matches the line in SHA256SUMS.txt. The README's
"Verify your download" section has a copy-paste one-liner that does the check for you.

Second Opinion v0.3.2

A maintenance release that hardens the v0.3.x optional features, from an internal adversarial trust-audit. No change to the deterministic scorer or rankings, and read-only as always.

Fixed

• Honest abstention (deep-dump): a malformed or foreign-architecture crash dump could cause -DeepDump to error out and leave a misleading "clean" result. It now abstains honestly — "not usable... not clean".
• Honest abstention (diff): with -Baseline, a signal that was unreadable in the current run could produce a false "activity dropped" line versus a readable baseline. Unreadable signals are now reported as "not comparable", never as good news.
• Redaction: very short (1–2 character) account names could slip through into the share-safe Helper Packet. Two-character names are now masked (ordinary prose is preserved by token-boundary matching; 1-character names stay excluded so they can't shred prose).
• Minor: stricter crash-dump header validation, and a cosmetic fix to the free-space comparison line.

Trust

Guarded by 33 snapshot fixtures + 135 guardrail assertions on Windows PowerShell 5.1 and PowerShell 7. The deterministic scorer still ranks (the AI never does); the read-only and never-changes-a-tier invariants were independently confirmed by the audit.

Second Opinion v0.3.1

What's new in v0.3.1

• Diff / trend mode (-Baseline) — compare a run against a saved redacted-evidence.json from a prior run. It emits a "what changed since the baseline" section as notes only: new/resolved crashes, WHEA, TDR, and restart volume; OS-build, dump-policy, and free-space changes; readability transitions; and whether the top hypothesis changed. It never moves a tier or confidence, and it stays honest — a missing, unreadable, or wrong-schema baseline reads "no usable baseline (not a clean comparison)", never "no change", and a no-delta diff says so explicitly. With -HelperPacket it also writes a redacted baseline-diff.md.

This completes the v0.3.x feature set: the honest-evidence layer, the correctness pass, the deep-dump bridge, the Helper Packet, and now run-to-run diff.

Trust is unchanged: a deterministic scorer (the AI never ranks), read-only by design, and guarded by 33 snapshot fixtures + 131 guardrail assertions on Windows PowerShell 5.1 and PowerShell 7.

MIT licensed. See the v0.3.0 notes for the full feature list. report.html is local and unredacted; ai-prompt.txt and the Helper Packet are the share-safe artifacts (best-effort redaction).

Second Opinion v0.3.0

A read-only Windows crash-triage tool that ranks likely culprits with honest abstention and hands you a report plus a redacted, paste-safe AI prompt. Highlights since v0.1.0:

New

• Helper Packet (-HelperPacket) — write a shareable, always-redacted case bundle to out\packet\: a forum/Discord-safe summary, a schema-versioned evidence JSON, a redaction audit (counts only, never the masked values), and an unreadable-signals list. Each artifact is stamped with the tool version + KB hash.
• Deep-dump bridge (-DeepDump) — if a crash dump exists, parse it read-only with locally-installed debugging tools (no downloads, no symbols) to name the faulting module as weak corroborating evidence. It never changes the deterministic ranking; a generic OS module stays inconclusive.
• Single-file / terminal run — src/Invoke-SecondOpinion.ps1 is self-contained (it embeds the bugcheck knowledge base), so you can download one file and run it from any terminal. Robust output-path handling for elevated or relocated shells.

Improved

• Honest evidence, made visible — an "observed but below threshold" channel so corrected-WHEA and sub-threshold signals never read as "clean"; a graded headline that never says "your PC is healthy"; and a signal-readability matrix showing what each run could and couldn't read.
• Correctness — fixed rapid same-code crash over-merging (which undercounted a recurring GPU bugcheck), XML-first bugcheck parsing (robust to localized event text), and stronger redaction (IPv6 + whole-word short-username masking).
• Safer AI prompt — machine-derived values (device / GPU / module names) are flattened to inert data so a hostile string can't forge prompt instructions.

Trust

• Deterministic scorer; the AI never ranks. Read-only — it makes no changes to the machine.
• Guarded by 33 snapshot fixtures + 119 guardrail assertions, on Windows PowerShell 5.1 and PowerShell 7.

MIT licensed. report.html is local and unredacted; ai-prompt.txt and the Helper Packet are the share-safe artifacts (best-effort redaction, not a guarantee).

Second Opinion v0.1.0

First public release.

A read-only Windows 11 crash-triage tool: it ranks the likely culprits for crashes, random restarts, WHEA errors, GPU TDRs, and drive/device issues into confidence tiers with evidence for and against each, then emits an unredacted report.html plus a best-effort-redacted ai-prompt.txt for your own AI. No system changes, no network.

Independently reviewed (presentation + security): no XSS or code-execution; AI-prompt hardened against injection. See the README and docs/DESIGN.md.