diff --git a/docs/build/sprints.md b/docs/build/sprints.md index ea5645b..e010de2 100644 --- a/docs/build/sprints.md +++ b/docs/build/sprints.md @@ -67,7 +67,7 @@ Status: **Done** (merged and signed off by Jim), **In progress** (work merged or | C4 | Full eval run | paticolcol | **Done** (PR #21; signed off by Jim 2026-11-12) | `73a0e0d` (seven presets, call counts, cost); `docs/results/results.json` (`78e3fcb`, total 300); `classifier_only` not run (classifier model download failed); MELON not built | | C5 | Conformal hold threshold + unseen-family rotation | paticolcol | **Done** (PR #21; signed off by Jim 2026-11-12) | q̂ frozen at 0.84 (`61e9f52`), 150 benign, ε = 5%; realised false-hold 4.67% (7/150); `calibration.py` strict `>` fix `28a33e9` | | C6 | README results and limits; freeze thresholds | — | **Done** (PR #21; signed off by Jim 2026-11-12) | `README.md` Results and Limits from `docs/results/results.json` (`78e3fcb`); `FUSION_HOLD_THRESHOLD` 0.84 | -| D1 | Landing page skeleton | Helena | **In progress** | Merged in PR #14 (`site/`); not deployed yet (done when a deployed URL exists) | +| D1 | Landing page skeleton | Helena | **Done** | Merged in PR #14 (`site/`); deployed 2026-11-13 to https://eclipse-20v.pages.dev/ (Cloudflare Pages project `eclipse`, from main `38aeaf0`) | | D4 | Results section on the real `results.json` | Helena | **In progress** | PR #24: E3's `ResultsSection` reads `docs/results/results.json` (P6, classifier on) and `redteam-results.json`; per-family from `by_family`; `cd site && npm test` 28/28, `npm run build` pass | | E1 | Live receipts panel | Ace | **In progress** | PR #11 review changes: panel integrated into `site/` using the shared React 18.3.1 / Vite 5.4.21 stack and brand; standalone app removed. 8/8 unit tests, production build and 7 Edge browser checks pass (390px, no horizontal scroll); Helena's re-review pending; see [site/README.md](../../site/README.md) | | P1 | Classifier fails loudly; `firewall-eval --preset`, scenario sets `external` and `paysim` | Jim | **Done** (signed off by Jim 2026-11-13) | `805a573`, follow-ups `73cb901`, `457d727`; signed off by Jim 2026-11-13 | @@ -102,7 +102,7 @@ Sprints 2–3: planned (owners and due dates below), not started. Sprints 4–5: | B2 | B | Backend | S | Nightly end-to-end script: one benign and one attack payment on testnet (human runs it). **Done, merged in PR #9 (`agent/src/firewall_agent/nightly_e2e.py` (`firewall-nightly-e2e`), `scripts/nightly-e2e.*`; unit tests in `agent/tests/test_nightly_e2e.py`; runbook & expected output in [nightly-e2e.md](nightly-e2e.md)); signed off by Jim 2026-11-11** | Script plus expected output | — | Rc | Done | | C1 | C | Backend | S | Freeze `results.json` (schema + example under `schemas/`) and add fixture DecisionRecords with filled signals for fusion tests. **Done, merged in PR #7; signed off by Jim 2026-11-11** | Schema and example committed; E3 and A3 can use them | — | paticolcol | Done | | C2 | C | Backend | M | Injection classifier: host `protectai/deberta-v3-base-prompt-injection-v2`, checker client scores each context item into `signals.injection`, measure latency. **Done, merged in PR #7; signed off by Jim 2026-11-11** | Score per item in the record; latency recorded | frozen `signals.injection` | paticolcol | Done | -| D1 | D | Frontend | M | Landing page skeleton: choose a static stack and host (must stay up through 2026-11-18, ideally 2026-11-25); sections: problem, how it works, NEAR's role, Jev's role, safety layers, demo, results, links | Deployed URL with draft copy | — | Helena | Not started | +| D1 | D | Frontend | M | Landing page skeleton: choose a static stack and host (must stay up through 2026-11-18, ideally 2026-11-25); sections: problem, how it works, NEAR's role, Jev's role, safety layers, demo, results, links | Deployed URL with draft copy | — | Helena | Done (deployed to https://eclipse-20v.pages.dev/ on 2026-11-13) | | E1 | E | Frontend | M | Live receipts panel: read-only calls to Guardian views over a public testnet RPC, explorer links, no wallet. **2026-11-11: integrated into `site/`, with filtering, pagination, refresh and exact token amounts; Helena's re-review pending.** | Shows the existing testnet receipts | live testnet Guardian | Ace | In progress | **Decision at the end of Sprint 1:** if A1 is not working, lane A replaces A4 with **A0: HTTPS + signed-verdict fallback mode** ([idea.md Risks](idea.md#risks-and-fallbacks)) in Sprint 2 and A4 moves to Sprint 3. diff --git a/site/README.md b/site/README.md index 6da91df..ae52d3d 100644 --- a/site/README.md +++ b/site/README.md @@ -56,9 +56,11 @@ The section shows three headline numbers for the conformal row against no defens ## Deploy (human, needs a Cloudflare account) +Live at https://eclipse-20v.pages.dev/ (Cloudflare Pages project `eclipse`, first deployed 2026-11-13 from main `38aeaf0`). The project is not connected to GitHub: redeploy from `site/` on an up-to-date `main` after each merge that touches the site. + ``` npm run build -npx wrangler pages deploy dist --project-name +npx wrangler pages deploy dist --project-name eclipse --branch main ``` The first run asks for a Cloudflare login and creates the project. Any static host works: `dist/` uses relative paths. diff --git a/site/src/components/ResultsSection.jsx b/site/src/components/ResultsSection.jsx index fbd4380..71aa5c4 100644 --- a/site/src/components/ResultsSection.jsx +++ b/site/src/components/ResultsSection.jsx @@ -135,7 +135,7 @@ export function ResultsSection({ rows = resultRows }) { n/a {formatResultNumber(redteamResult.missedPct)} ({redteamResult.held} held, {redteamResult.blocked} blocked) {formatResultNumber(redteamResult.medianMs)} - 1 Jev call + 1 classifier call + {formatCalls(redteamResult)} melonMasked re-run (MELON) only @@ -165,7 +165,8 @@ export function ResultsSection({ rows = resultRows }) {
  • Detection covers the six evaluated attack families. Against new or adaptive attacks, the loss is bounded by the - hard caps in contract code: {PER_PAYMENT_CAP} tUSDC per payment, {DAILY_CAP} tUSDC per day. + hard caps in contract code: {PER_PAYMENT_CAP} tUSDC per payment, and {DAILY_CAP} tUSDC committed per UTC request + day. Payouts on one day can also include holds committed the day before.
  • Latency includes the injection classifier running on a local CPU. Cost is unpriced, so the table shows calls per diff --git a/site/src/results.js b/site/src/results.js index bfd6f7b..f3183a6 100644 --- a/site/src/results.js +++ b/site/src/results.js @@ -68,6 +68,7 @@ export function headline(results) { export function redteamRow(redteam) { validateResultsFormat(redteam); const s = redteam.summary; + const [config] = redteam.configurations ?? []; return { total: redteam.attack_total, missedPct: s.missed_attacks_pct, @@ -75,6 +76,8 @@ export function redteamRow(redteam) { held: s.held_attacks, blocked: s.blocked_attacks, medianMs: redteam.latency_ms.p50_ms, + jevCalls: config?.jev_calls_per_payment ?? 0, + classifierCalls: config?.classifier_calls_per_payment ?? 0, }; } diff --git a/site/src/results.test.js b/site/src/results.test.js index 083b3d0..6c271dc 100644 --- a/site/src/results.test.js +++ b/site/src/results.test.js @@ -42,7 +42,9 @@ test("maps every real configuration in source order with a README label", () => test("headline, red-team and family numbers match the README Results section", () => { const { base, best } = headline(results); assert.deepEqual([base.missed_attacks_pct, best.missed_attacks_pct, best.false_blocks_pct, best.false_holds_pct], [100, 19.33, 0, 5.33]); - assert.deepEqual(redteamRow(redteam), { total: 17, missedPct: 41.18, missed: 7, held: 7, blocked: 3, medianMs: 1137.9 }); + assert.deepEqual(redteamRow(redteam), { total: 17, missedPct: 41.18, missed: 7, held: 7, blocked: 3, medianMs: 1137.9, + jevCalls: redteam.configurations[0].jev_calls_per_payment, classifierCalls: redteam.configurations[0].classifier_calls_per_payment }); + assert.equal(formatCalls(redteamRow(redteam)), "1 Jev call + 1 classifier call"); const missed = Object.fromEntries(familyRows(results).map((f) => [f.family, f.total - f.caught])); assert.deepEqual(missed, { direct_injection: 1, encoded_injection: 2, indirect_injection: 1, lookalike_receiver: 2, fake_invoice: 12, memory_poisoning: 12 }); }); @@ -57,7 +59,7 @@ test("renders the real results with all rows, units and caveats, and no sample l const html = renderToStaticMarkup(React.createElement(ResultsSection)); for (const entry of results.configurations) assert.ok(html.includes(entry.configuration), entry.configuration); for (const label of ["False holds (%)", "False blocks (%)", "Missed attacks (%)", "Median latency (ms)", "Per payment", - "19.33%", "exchangeability", "in-sample", "six evaluated attack families", "Not built (cut list)", "docs/results/results.json"]) { + "19.33%", "exchangeability", "in-sample", "six evaluated attack families", "committed per UTC request", "Not built (cut list)", "docs/results/results.json"]) { assert.ok(html.includes(label), `Missing ${label}`); } assert.doesNotMatch(html, /Sample data|Cost per payment|Numbers land after/);