HiveMind v2.10.2 — OpenRouter In-Browser Playground & Website GPU Calculator

Summary of Changes in v2.10.2

• Interactive OpenRouter Portal & In-Browser Playground (src/hivemind/router/portal_page.py):

  • Live at GET / and GET /playground on OpenRouter.
  • In-browser AI chat completion console with streaming token speed (tok/s), latency metrics (ms), and cost tracking.
  • 5 Capability Tiers matrix (coder-lite, coder, coder-plus, moe-general, moe-coder) with recommended Ollama tags.
  • Live account balance & role inspector querying /v1/hivemind/me.
  • 1-click ready-to-copy editor snippets (Python OpenAI SDK, Continue, Cline, Aider, OpenCode, cURL).
  • Strict Content-Security-Policy with per-request nonces and textContent-only DOM manipulation.

• Gateway Root & Health Probe Enhancement (src/hivemind/router/gateway.py):

  • Mounted GET /, HEAD /, GET /playground, and HEAD /playground.
  • Fixed HEAD method support on /signup, /health, and /healthz (HTTP 200 OK for uptime monitors).

• HiveMind@home & GitHub Pages Enhancements (src/hivemind/boinc/hivemind_at_home_dashboard.py, HiveMindatHome.github.io):

  • Interactive GPU Compute Points Calculator: 8 consumer GPU presets (RTX 3060..4090, RX 7800 XT, Apple Silicon), idle hours slider, dynamic calculation of monthly HMT points and unlocked free coding tokens.
  • Interactive Live Completion Console: In-dashboard query testing and real-time gateway ping badge.
  • Full multi-language localization across 5 languages (en, fr, zh, es, de).
  • Deployed live on GitHub Pages: https://xaviercallens.github.io/HiveMindatHome.github.io/

• Zero-Trust Verification:

  • tests/test_router_portal.py: 3 unit tests verifying root portal HTML, HEAD routes, and CSP.
  • All 228 router pytest tests passed (227 passed, 1 skipped).
  • GWAYA Zero-Stub AST audit: 111/111 files clean.
  • PR #146 merged with 3/3 green GitHub Actions CI checks.
  • Strict compliance with the 0.00 EUR / 0.00 USD Non-Monetary Compute-Reciprocity Charter.

HiveMind v2.10.1 — OpenRouter Google Flow (CLI, UI & E2E Verification)

Summary of Changes in v2.10.1

• CLI Google Flow Integration (scripts/hivemind_connect.py):

  • Added register-google and login-google subcommands.
  • Supports direct token input (--token), automated GCP token issuance (--gcloud with gcloud auth print-identity-token --include-email), and interactive browser registration (--browser).
  • Strict owner-only (0o600) file permissions on auto-saved API keys.

• HiveMind@home & GitHub Pages UI Showcase (src/hivemind/boinc/hivemind_at_home_dashboard.py, HiveMindatHome.github.io):

  • Added Screen 02C: OpenRouter Google Gateway & Free AI Inference card.
  • Showcases 5 capability tiers: coder-lite (3B), coder (7B), coder-plus (14B), moe-general (20B), moe-coder (30B).
  • Multi-language localization across English, French, Chinese, Spanish, German.

• Comprehensive End-to-End Verification Suite (scripts/test_router_google_flow.py):

  • 10/10 deterministic tests passing with cryptographic receipt SHA-256 78631759c841adfa1c847af172cc47229e276a0bcf1b2fa0c75bdba926b6baad.
  • Verifies CSP nonce, RS256 JWT validation, anti-replay, key rotation, negative controls, anti-squatting handle protection, and CLI integration.

• Quality & Security Gates:

  • GWAYA Zero-Stub AST audit: 110/110 files verified clean.
  • All 225 router unit and integration tests passing.
  • 5-Layer PR Security Gate certified.
  • Strict compliance with the 0.00 EUR / 0.00 USD Non-Monetary Compute-Reciprocity Charter.

v2.10.0 — Sign in with Google for the OpenRouter

Sign in with Google (low-cost easy registration)

• GET /signup page with the Google button; POST /v1/hivemind/register/google verifies the Google ID token server-side (RS256 pinned, JWKS cached, iss/aud/exp/freshness/email_verified/replay checks).
• Stores only a keyed hash of the Google subject (google.<hmac>), never the e-mail. Signing in again keeps the account and issues a fresh key.
• google. handle prefix reserved; optional HIVEMIND_OPEN_REGISTRATION=0 closes handle-only registration.
• Disabled until HIVEMIND_GOOGLE_CLIENT_ID is set. Not yet exercised against real Google. Operator steps: docs/HIVEMIND_OPENROUTER.md.
• 43 new tests (223 router tests pass).

v2.9.0 - Open-Router kit for Python and game coders

Open-Router kit for Python and game coders

Use open coding models from your editor, or share a gaming GPU and earn HiveMind points (HMT), all through the standard OpenAI chat-completions protocol. Guide: docs/HIVEMIND_PYTHON_GAME_DEVS.md.

• Capability tiers served at /v1/hivemind/catalog: coder-lite (3B), coder (7B), coder-plus (14B), and two mixture-of-experts tiers, moe-general (16 GB) and moe-coder (24 GB).
• Producer agent: hardware suggestion, --preset, preflight that refuses to advertise a model the endpoint does not serve, optional endpoint token.
• Producer shield: exposes only the two chat routes, demands a token, caps size and concurrency.
• Connector script with settings for Continue, Cline, Aider, OpenCode, Zed, JetBrains, Python and environment variables.
• 180 router tests; a live run through the real chain (connector, gateway, shield, local Ollama) with negative controls.

Deployment status: this release is merged but the public Cloud Run service still runs the v2.8.0 build. Redeploying adds a column to the live ledger database (tested on an old-schema copy) and waits for the owner's go-ahead.

Limits: HMT are points with no cash value; producers can read prompts; tier ids are claims until verified (HM-S13-13); no inline autocomplete; no producer serves the MoE tiers yet.

v2.8.0 - Operation Open-Router: OpenAI-compatible MoE gateway

Operation Open-Router: an OpenAI-compatible MoE gateway with a producer/consumer equilibrium

Use HiveMind MoE models from OpenCode, VS Code (Continue) or any OpenAI client; serve them from your own GPU and earn HiveMind points (HMT).

• Kelly consumer game with a price that does not depend on supply, producer dominant-strategy supply, posted-price fallback, and a public win-win certificate.
• Hash-chained HMT ledger, HMAC-hashed API keys, escrow holds, SSRF-guarded producer registration.
• 144 router tests; a live run against local Ollama through the official openai client; an adversarial review found 5 high and 6 medium issues, all fixed with regression tests.
• Public Cloud Run deployment with a Litestream-replicated ledger (single instance). Guide: docs/HIVEMIND_OPENROUTER.md. Roadmap: section 3.11 (HM-S13-01..08).

Limits: HMT are points with no cash value; producers can read the prompts they serve; output correctness is not yet verified (HM-S13-05).

v2.7.0 — Operation Fair-Market: BOINC computing coin and Nash-equilibrium compute market (proposal, code, pre-registered simulation)

Season 12 opens with the owner's brief on the compute economy, implemented as code and measured in a pre-registered simulation. Status: proposed, not adopted — the ledger's block-reward rule is unchanged until the receipt replay (HM-S12-02) and an owner order.

• Credit: BOINC's cobblestone (432 GFLOP), granted only for spot-checked, Multi-Krum-admitted work; capped by the execution class's measured peak and by the canonical replica. No verification, no credit.
• Issuance: $HIVE minted per verified cobblestone with geometric decay under the existing 21 M cap; consumer payments split 80 % workers / 15 % validators / 5 % burned.
• Market: consumers bid under the Kelly proportional-allocation mechanism (unique Nash equilibrium, efficiency ≥ 75 % by the Johari–Tsitsiklis 4/3 bound); workers supply into a Tullock contest (closed-form unique equilibrium, ε-Nash certified); across queues a damped best-response routing equalises marginal returns and returns a deviation certificate instead of a convergence promise.
• Simulation (seed 2026, pre-registered): Kelly efficiency min 0.800 / median 0.869 over 300 instances; 300/300 Tullock equilibria certified; routing 55/60 undamped (the registered ≥ 95 % hypothesis is refuted for that iteration; run kept) and 60/60 after damping; both controls passed.
• Proposal of record docs/COMPUTE_COIN_AND_MARKET.md; ROADMAP §3.10 tasks HM-S12-01..06; 12 tests (5 property-based). PR #135, merged by the owner-only gatekeeper after ACCEPT and green CI.
• Also since v2.6.0: the coordinator's Cyber-Bee portal (PR #134) and its multilingual portal media (PR #137, the release commit). Both were opened and merged by the coordinator with the owner's token; #137 is a re-submission of #136 whose commits had been re-authored from bot@socrateai.com to the owner's e-mail. Recorded here as what it is: commit e-mail is a claim (HM-S11-04).

v2.6.0 — Operation Hardened-Hive: HiveMindFirewall, chain of command, scripted red team; La Ruche vault

Season 11 opens with the owner's hardening brief implemented and measured.

• hivemind.security.hivemind_firewall (PR #132): principals are keyed hashes (HMAC-SHA256 of the identity) with Ed25519 keys; delegation certificates only to strictly lower roles, rooted at the owner and re-walked on every order; orders are signed, sequenced, nonce-protected and bound to a hash-chained ledger head; any free text is scanned by deterministic rules PI-01..PI-07 (encoded payloads decoded and re-scanned) and quarantined as data; the decision is a pure state machine where anything undefined is QUARANTINED; executed actions use an allow-listed argv under rlimits inside bubblewrap or are BLOCKED. Text never carries authority; a model can only propose.
• Scripted red team over 59 pre-registered vectors: the first run blocked 47 of 49 attacks and found two real defects, both fixed; the second run blocked 49 of 49 with 0 false positives and both controls passing (artifacts/security/firewall_pentest_report.json).
• Gatekeeper: report-only injection scan of PR text; GitHub signature verification recorded per commit; HIVEMIND_GATE_REQUIRE_SIGNED=1 holds unsigned owner-attributed commits (PRs #125, #128, #133 hardened the gate along the way; main requires a pull request and enforces admins).
• Roadmap §3.9 "Operation Hardened-Hive": tasks HM-S11-01..08, pen-test campaigns HM-S11-PT-1..3; design in docs/SECURITY_HARDENING_PHASE.md; SECURITY.md §5.
• Also in this release: paper v3.7.2 (Lean gate 128/128, two lemmas flagged), the T4 branch review, and the coordinator's La Ruche MoE vault / HiveMind@home work (PR #131).

Known gap stated plainly: commit e-mail is a claim, not proof; signed commits from the owner's key (HM-S11-04) are the next step.

v2.5.1 — preprint v3.7.1: E12 pilot complete (bf16 class diverges beyond the tampered control)

Preprint v3.7.1, Zenodo DOI 10.5281/zenodo.23607597 (new version of 10.5281/zenodo.23604695).

• E12 pilot complete (pre-registered): the bf16-autocast cell, replayed by an fp32 validator on the same host, build and thread count, diverges to 3.3e-3 with 96 of 96 adapter tensors above eps = 1e-4, 2.7x the tampered control (P12d supported). Both terms of the re-specified spot-check rule saturate, so dtype is a class boundary: validators are assigned within the dtype class and the bf16 class gets its own eps from a GPU measurement before any bf16 worker is admitted.
• Paper: section 4.1 rule paragraph, E12 section and limitations updated; PROV-34 added (34/34 provenance locks). No other number changed.
• Ledger: P12d supported. Zenodo bundle gains the bf16 raw replay state.
• PRs #116 and #117, merged by the owner-only gatekeeper after ACCEPT.

Known difference recorded in the deposit receipt: the uploaded HiveMindSpec.lean carries eight uncommitted theorems from the shared working tree; the committed file has the 120 the paper describes.

v2.5.0 — preprint v3.7.0: E12 replay-divergence pilot; hourly owner-only PR gatekeeper

Preprint v3.7.0 — Zenodo DOI 10.5281/zenodo.23604695

• E12-pilot (pre-registered): the Byzantine spot-check's ε = 1e-4 replay rule fails on CPU for a thread-count change alone (2.6e-4 / 2.5e-4) and for a second host (1.5e-4); bit-identical execution reproduces exactly; the tampered control sits at 1.25e-3 with 96/96 adapter tensors above ε versus 3–5/96 for honest divergence. The rule is re-specified as execution-class matched with a per-class ε and a tensor-count term; GPU classes and the bf16 cell remain open.
• Hourly PR gatekeeper (scripts/pr_gatekeeper.py): owner-only merge authority (git author callensxavier@gmail.com on every commit, PR author xaviercallens), PRSecurityScanner + forbidden-pattern scan + GWAYA audit + pytest + Lean + paper pipeline; everything else is checked and held; PR text is data. Branch protection on main blocks forced pushes and deletions. This release's own PR (#115) was first refused by the gate (a task id cited in the body without functional changes) and merged only after the body stopped claiming it.
• Provenance lock 33/33; ledger P12a–d; is_peer_review: false.

v2.4.0 — preprint v3.6.0 after an external review; RDA Proposition machine-checked over ℝ; directions from the owner's publications

Preprint v3.6.0 — Zenodo DOI 10.5281/zenodo.23602528 (26 files, checksums verified)

• External review of v3.5.0 recorded verbatim (paper/review/external_review_v3.5.0_2026-10-23.md; author unstated, no score, not a peer review of record) with a point-by-point response. Accepted: hardware-heterogeneity caveat and architecture-matched rule for the ε = 1e-4 spot-check (new RQ11/E12, HM-S10-08), compounding-approximations threat (E8 all-on vs each-off cells), PoC-scale caveat, architecture figure, Skip-Token-as-dropout note, PA-CGAD-under-quorum note. Declined: moving the scripted-review correction out of the manuscript. Four reported typos were PDF-text-extraction artefacts.
• RDA Proposition machine-checked over ℝ: lean4/mathlib/RDAMultiplier.lean (3 Mathlib theorems, standard axioms only, independent gate; PROV-31).
• Directions from the owner's publications (ROADMAP §3.8): E11 lattice KV-cache codec (RQ10, HM-S10-05), Laya-LoRA curriculum as campaign workload (HM-S10-06), shared lattice package (HM-S10-07), E12 replay divergence (HM-S10-08); five self-citations verified against the Zenodo API.
• Numbers unchanged from v3.5.0; provenance lock 31/31; is_peer_review: false.

Verification

• pdflatex ×3 clean (34 pages, 0 undefined refs); autoevolve_peer_review_pipeline.py 31/31
• Lean: HiveMindSpec 120/120 and RDAMultiplier 3/3 under anse.formal.lean_runner
• pytest subsets (README claims, swarm v2, lattice quantiser) passed; full suite 1018 passed at v2.3.0