v0.5.48 — Scanner Cluster Block

v0.5.48 — Scanner Cluster Block

Security-hygiene patch: blocks a coordinated cluster of config/secret scanners plus a combined config/RCE scanner, all confirmed via GCP forensic analysis. The application-layer IP blocklist grows to 26 entries; one new rotation-proof UA-substring block added. Zero sensitive-path data was served to any blocked actor — existing defense layers (rate limiting, HTTP→HTTPS redirects, 404 handling) already fully deflected them.

• Security: config/secret scanner cluster (2 IPs, same subnet as a prior probe) + a combined config/RCE scanner (1 IP, with active injection probing that our Python/FastAPI server ignored — zero execution, zero leak) blocked at the application layer
• Security: rotation-proof UA-substring block added for a self-identifying scanner tool
• Monitoring: two crawler user-agents flagged for a 14-day watch; no IP block warranted at current volume

No functional changes — reasoning visibility, the BYOK frontier-model menu, and the free default path are all unchanged. Existing tests pass.

🤖 Generated with Claude Code

v0.5.47 — Model Currency

v0.5.47 — Model Currency

Keeps the BYOK frontier-model menu current and migrates the Gemini integration to the supported SDK. Every model ID was live-verified against its provider before listing. The default free-tier path (Gemini 2.5 Flash + Groq) is unchanged — cost and stability preserved.

What changed

• Gemini SDK migration — moved to the current google-genai client SDK (required for Gemini 3.x). gemini-3.1-pro-preview and gemini-3.5-flash added as BYOK/frontier options; default stays gemini-2.5-flash.
• OpenAI → GPT-5.5 (BYOK default) + a gpt-5.x request-contract fix.
• Mistral → Medium 3 (BYOK default).
• Probe #23 — config/secret scanner added to the IP blocklist (zero leak).

Verified

699 tests pass; pre-deploy live Trinity smoke green on both the default and a frontier BYOK config; post-deploy live validation on production confirms real X/Z/CS inference on the migrated SDK.

Qwen deferred to v0.5.48 (pending an API key).

v0.5.46 — BYOK Robustness

v0.5.46 — BYOK Robustness

A production-hardening release for Bring-Your-Own-Key (BYOK) users, surfaced by dogfooding our own multi-model evaluation. Strictly additive — no change to the default free-tier path. 698 tests pass.

What changed

• provider/model shorthand accepted — BYOK configs like anthropic/claude-opus-4-8 are now parsed correctly server-side, with an actionable error on an unknown provider. Fixes a recent rejection of valid BYOK model configs.
• Higher output ceiling — Z/CS agents raised from 4096 → 8192 max tokens (per-provider clamps where needed) so verbose models no longer truncate mid-response.
• Token-monitor repair — the internal output-token monitor is reading real values again (it had silently read 0 since v0.5.3).
• Provider schema robustness — structured evidence returned by some providers is now normalized, fixing multi-provider schema-conformance failures.

Why

Exercising non-default model providers for the first time during our critique-quality evaluation surfaced latent gaps that never affected the default path. Fixing them ships real value to BYOK users and unblocked a clean measurement run.

PR: #262

v0.5.45 — Probe Blocklist

v0.5.45 — Probe Blocklist

Security + metrics-integrity patch. Blocks an MCP-endpoint scanner family that was polluting engagement metrics, plus a self-declared MCP scanner caught via GCP forensic analysis.

What changed

• Blocked an MCP-endpoint scanner family at the application layer — one member was generating cron-driven traffic that would otherwise register as verified engagement. Total 12 IPs added to the blocklist.
• Added a self-declared MCP endpoint scanner (identified via Sentinel GCP forensic analysis): a single high-rate dictionary sweep of MCP transport paths. Defense stack held — zero content served, zero leak.
• Added rotation-proof User-Agent blocks for the scanner families (defense-in-depth vs IP rotation).

Why

Without these blocks, upcoming metrics reports would have inherited automated scanner activity as if it were genuine engagement. Our honest-metrics gate held the next analytics report until this shipped. No real handler was reached by any blocked actor.

Tests: 684 passing. No change to the 13 free tools.

PR: #260

v0.5.44 — Reasoning Visible

v0.5.44 — Reasoning Visible

Makes the auditable reasoning layer — the core of the product — visible in tool responses by default. The X/Z/CS agents already generated per-step reasoning, framework citations, an ethics scoring breakdown, and Socratic questions on every call; the flagship run_full_trinity previously trimmed all of it and returned only scores. This is serialization, not new inference — no added cost or latency.

What changed

• detail parameter on run_full_trinity + a new additive reasoning block:
  • standard (new default) — per-agent reasoning steps, Z's 5-dimension ethics scoring breakdown + framework citations + detected jurisdictions, CS threat level + Socratic questions + attack vectors, and per-agent inference quality (incl. the v0.5.43 degraded-inference warning).
  • full — adds per-step evidence and the heaviest structured fields (12-dimension security matrix, 6-stage record, MACP assessment, market competition).
  • summary — omits the block, reproducing the exact prior response shape (opt-out).
• Strictly additive — no existing response field changed at any level; summary is byte-compatible with prior releases.
• consult_agent_x/z/cs gain the same detail ladder (they previously dropped per-step evidence + structured fields).
• Markdown report now renders the ethics scoring breakdown, jurisdictions, CS threat level, and the degraded-inference warning.

Why

VerifiMind's differentiator is auditable, multi-model verification reasoning — and it was being computed, then discarded before reaching the caller. Returning it by default is the proof surface for the epistemic-verification positioning. 16 new tests; 682 passing.

PR: #256 (code) · #257 (docs)

v0.5.43 — Foundation Integrity

v0.5.43 — Foundation Integrity

Patch release hardening the verification core and the MCP resource surface, from a foundation-effectiveness audit of the flagship Trinity tool.

What changed

• Ethics-veto fail-safe: if the Z Guardian's inference is degraded (truncated and schema-repaired), the verdict is now capped at REVISE with an explicit inference_warning requiring human review — instead of potentially surfacing a clean recommendation on synthesized defaults. Genuine inference is unaffected.
• Validation-history privacy: save_to_history now defaults to off, and the history resources (genesis://history/*) return non-identifying summaries / aggregate statistics only — a caller's concept text is never exposed to others on a shared instance.
• Methodology resource accuracy: genesis://config/master_prompt is now generated from the live agent configuration, always matching the exact X / Z / CS prompts in production.
• Current-date awareness: agent prompts are anchored to the real current date, so market-recency and regulatory-deadline reasoning stay current.
• Version currency: genesis://state/project_info reports the live server version.

Why

A live-tested audit confirmed the core is effective (the ethics veto fires; production runs genuinely multi-model; scoring is proportional) and surfaced that the verdict could rest on synthesized defaults in the degraded path, and that the resource surface had drifted from production. Fail-safe-first behavior and resource accuracy are foundational to an epistemic-verification product. 10 new regression tests; 666 passing.

PR: #254 (code) · #255 (docs)

v0.5.42 — Server-Card Description Refresh

v0.5.42 — Server-Card Description Refresh

Refreshes the /.well-known/mcp/server-card.json MCP discovery surface.

What changed

• Description refreshed — removed stale legacy marketing copy; now reflects the current feature set: 13 free MCP tools (Trinity validation + prompt-template library + cross-session coordination), BYOK across 6 providers (Gemini, Anthropic, OpenAI, Groq, Cerebras, Mistral), free tier on Gemini 2.5 Flash.
• Discovery card de-coupled from Smithery — the Smithery listing sunset 2026-03-01; this endpoint remains a valid MCP discovery card served directly at the deployment URL.

Why

Pre-submission accuracy pass: the public MCP discovery surface now matches the live v0.5.41+ product.

PR: #251

v0.5.41 — Register Page Dead-Link Fix

v0.5.41 — Register Page Dead-Link Fix

A patch release that removes the last dead link in the registration funnel.

What changed

• The /register page's "Already registered? check your status" link pointed to a bare, incomplete URL that returned 404. It now points to /whoami — the v0.5.40 self-serve status endpoint (returns your tier and registration state; accepts your UUID via query param or MCP client header).

Why

v0.5.40 fixed the registration funnel's structural 404; this removes the last UI link that still pointed users at a dead end. The registration flow is now clean end-to-end.

PR: #248 · #249

v0.5.40 — Registration Funnel Fix + /whoami + Model Currency

v0.5.40 — Registration Funnel Fix + /whoami + Model Currency

Closes the Scholar UUID registration funnel gap, adds a self-serve tier status endpoint, and brings the Anthropic model list current.

What changed

Registration funnel fix

• Rate limiter now reads the correct collection for tier resolution — Scholar UUID holders no longer incorrectly tier-resolved as anonymous
• /early-adopters/status/{uuid} returns a helpful 200 response with a register CTA for valid unregistered UUIDs (previously returned 404, blocking users who had UUIDs from checking their status)

New: /whoami endpoint

• Self-serve tier and registration status check via X-VerifiMind-UUID header or ?uuid= query param
• Returns tier, rate limit, status, and next-step guidance — no auth required
• Rate-limit exempt so users can always check their own status

Model currency

• Anthropic provider model list updated to current Opus version
• AI Council skill Z-Agent updated to current Claude Sonnet model (unblocks 4-agent Council sessions)

Why

AY/AZ forensic analysis found 89 distinct interested visitors over 30 days with only 2.2% completing registration — caused by a structural collection mismatch that sent UUID holders to dead ends instead of the registration flow. This fix converts that lost demand into identified Early Adopters we can talk to.

PRs: #245 · #246 · #247

v0.5.39 — Registry Scanner Block + P2 Batch-2

v0.5.39 — Registry Scanner Block + P2 Batch-2

Combines a security-hygiene block of a registry scanner (10th application-layer block, forensics from AY+AZ) with the second batch of SonarCloud P2 dup-literal constant extractions. Defense outcome: zero handler reached on the new actor (400 prior requests, zero leak).

What changed

• Tenth scanner blocked — an automated MCP/OAuth surface-enumeration scanner with self-identified registry UA; 5-day cron-like pattern; never reached a real handler. BLOCKED_IPS: 10 entries total (was 9).
• SonarCloud P2 batch-2 — 8 module-level constants extracted from 25 duplicated string-literal occurrences across the LLM provider config and the agent dispatch surface (provider default models + agent role names + master prompt filename). Behavior-identical refactor; identical wire output.

Why

A routine forensic sweep surfaced the registry scanner; the application-layer block trims upstream noise without changing the legitimate path. The constant extractions continue the credibility-through-clean-internals work started in v0.5.32 / v0.5.39's batch-1 (#238) — every extracted literal is a smaller surface for drift.

PR: #239