v1.1.1 · Round-9 fixes: the first click, receipts that always verify, questions about itself

Round-9 fixes: the first click, receipts that always verify, questions about itself

Patch release on v1.1.0 with the fixes from the round-9 mock judges. Full list: CHANGELOG.md (v1.1.1).

What a judge will notice

• The first click buys one article and leads with the answer. On the live sandbox (live planner), the RLUSD escrow preset buys 1 article (S$0.40; v1.1.0 bought 3 for S$1.20) and leads with AllowTrustLineLocking. The x402 402 question is now the first preset, and the plan you confirm in the 5 s modal is the plan that runs.
• Every exported receipt verifies. Free articles without the writer's signed delivery proof are dropped like a failed read, so they never reach an answer or a receipt. Release admission now also requires the packed wheel and TS CLI to verify a hosted export that exercises this case.
• A paid run with no answer says so ("Bought 2 articles; neither gave a citable answer to this question."), with a notice naming the reason.
• It answers questions about itself: how it decides what to buy, how the writer gets paid, and what a writer earns each buy one team-written explainer (S$0.40) and lead with the answer. Off-topic chat gets one fixed line pointing back to what it researches.
• The sandbox no longer says "paused" when it is only slow, and uptime checks no longer fill its visitor-session table.
• MCP labels name who wrote the sources: team-authored explainers are labelled team-authored, fictional demo writers fictional.
• One very long word in a writer's passage no longer stalls answer writing (3.9 s → 8 ms, same decisions).

Try and verify (no clone, Node ≥ 22.13 / Python ≥ 3.10)

REL=https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.1.1
npx -y $REL/researchagent-mcp-1.1.1.tgz try --testnet
pip install $REL/researchagent-1.1.1-py3-none-any.whl
researchagent verify receipts.json --ledger          # Python, no Node
npx -y $REL/researchagent-mcp-1.1.1.tgz verify receipts.json --ledger   # TypeScript

XRPL Testnet with faucet funds or the sandbox's test-funds sponsor; nothing here uses real money. Not on npm, PyPI or the MCP Registry yet (owner step). Mac DMGs are unsigned.

Apple Silicon and Intel Mac apps attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

Verified locally with fixture providers and simulated settlement; native arm64 recovery passed. Intel DMG built; native Intel acceptance pending. Local acceptance and gate receipts attached.

v1.1.0 · Finals release: verify what the live sandbox exports

Finals release: verify what the live sandbox exports

The finals release for SingHacks at SFF 2026. Everything since v1.0.2, judge-facing changes first. Full list: CHANGELOG.md (v1.1.0).

What a judge will notice

• A live sandbox with nothing to install: https://ra-front-651013803640.asia-southeast1.run.app, open 13–19 Nov 2026 on XRPL Testnet (test funds, no real value). 5 questions per visitor per day, up to S$2.00 of test funds each.
• Receipts from the sandbox verify with these packages. v1.0.2's verifier reported FAIL anchor binding on hosted exports; this release's wheel and TS CLI pass them, and release admission now requires that.
• It buys the article titled with your question, buys a further article only when it adds part of the fact asked for, keeps the live plan when a question says "pay" or "buy", and leads the answer with the fact you asked for.
• Proof of refusal: a declined, anchored run becomes a public page you can check on XRPL in one click.
• Measured once on a blind bank: v5 held-out 16/20 bought and cited, 0/10 off-topic purchases (XRPL Testnet; fixture research and decision models with answers written offline; team-authored writers).
• Agent hires agent ships in the tgz: try --testnet --hire.

Try and verify (no clone, Node ≥ 22.13 / Python ≥ 3.10)

REL=https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.1.0
npx -y $REL/researchagent-mcp-1.1.0.tgz try --testnet
pip install $REL/researchagent-1.1.0-py3-none-any.whl
researchagent verify receipts.json --ledger          # Python, no Node
npx -y $REL/researchagent-mcp-1.1.0.tgz verify receipts.json --ledger   # TypeScript

Not on npm, PyPI or the MCP Registry yet (owner step). Mac DMGs are unsigned.

Apple Silicon and Intel Mac apps attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

Verified locally with fixture providers and simulated settlement; native arm64 recovery passed. Intel DMG built; native Intel acceptance pending. Local acceptance and gate receipts attached.

v1.0.2 · ResearchAgent 1.0.2

ResearchAgent 1.0.2

ResearchAgent 1.0.2 is a patch release that fixes what the round-6 mock judges saw on v1.0.1. Final local acceptance passed on 2d3999f: the 2,000-run soak (loop p99 24.6 ms against a 50 ms limit), all release artifacts and clean consumers, the arm64 DMG, desktop recovery, the browser suite, Python, coverage and chaos. GitHub's release workflow builds both DMGs. Buying policy is unchanged. You can install everything straight from this release without cloning the repo. Publishing to npm, PyPI and the MCP Registry is up to the owner and comes later.

Prove it in one line (real XRPL Testnet, no clone)

npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.2/researchagent-mcp-1.0.2.tgz try --testnet

This line runs the full path:

1. Funds faucet wallets.
2. Buys an article by key-reveal escrow.
3. Catches a broken promise and gets a writer-signed refund.
4. Buys again.
5. Anchors the answer.
6. Verifies every transaction on the ledger.

It ends with RESULT PASS. Testnet has no real value. If the faucet fails, the run is labelled SIMULATED and gives the reason.

Ask your own question

# The full web app on Testnet (Node ≥ 22.13)
npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.2/researchagent-mcp-1.0.2.tgz app --testnet

# Any MCP client on Testnet. The operator sets the caps; the default budget is zero.
claude mcp add researchagent -e MCP_MAX_BUDGET_MINOR=200 -e MCP_DAILY_CAP_MINOR=500 -- npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.2/researchagent-mcp-1.0.2.tgz --testnet

Add --iou to either command to pay writers in TEST-SGD through XLS-85 token escrow. TEST-SGD is our own labelled stand-in for RLUSD.

Fixed since 1.0.1

• A validated anchor no longer shows as "unavailable" (#418).
  • In 1.0.1, every purchase run showed "Anchor unavailable" until you reloaded the page.
  • Now the screen says "Anchoring… awaiting validation" and switches to "Anchored on XRPL · ledger N" by itself.
  • A real failure still ends with a clear failure label.
• The app shows that a refusal is anchored too (#418). A question that buys nothing now shows "Anchored on XRPL · no purchase · ledger N" and the Export menu, so you can verify the abstention with researchagent verify --ledger.
• The plan says what policy does (#418).
  • The 5 s plan screen and the home page now state the real rule: at most one article per round, up to 3 rounds, within your budget.
  • The text is generated from the same constant the research loop uses, so the two cannot drift.
• Labels and numbers corrected (#418):
  • the app and landing page name the labelled team-authored notes alongside the fictional demo writers;
  • the ledger count reads 61/61;
  • the landing page links the SFF deck;
  • the pitch notes use the held-out numbers;
  • the Testnet app suggests three XRPL questions.

How well it answers (unchanged from 1.0.1, measured once, held out)

On the blind v4 judge bank, run once on bdc30f97:

• 7 of 20 questions are bought and cited;
• the key fact leads the answer in 2 of 20;
• 0 of 10 off-topic questions buy anything;
• there are 0 gate violations.

The run used the fixture decision model, team-authored writers (not independent) and faucet-funded Testnet. Evidence.

Merged since v1.0.1

• Feasibility: keep acceptance trees clean and run perf budgets alone in the local gate (#421)
• Feasibility: find the Settings window by URL in desktop smoke and recovery tests (#420)
• Reachability: prepare v1.0.2 (#419)
• UX / Technical depth: fix the anchor-pending regression and judge-visible honesty gaps (#418)
• Feasibility / Upkeep: local merge gate, slim hosted CI and parallel acceptance (#417)

The macOS DMGs are unsigned until the owner adds an Apple Developer ID. On first launch, right-click the app and choose Open.

Apple Silicon and Intel Mac apps attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

v1.0.1 · ResearchAgent 1.0.1

ResearchAgent 1.0.1

ResearchAgent 1.0.1 is a patch release. It contains the exact artifacts CI accepted for this commit. Final local acceptance passed on 369dd0e: the 2,000-run soak (loop p99 25.2 ms against a 50 ms limit), all release artifacts and clean consumers, the arm64 and x64 DMGs, desktop recovery, the browser suite, check:fast, Python, coverage and chaos. You can install everything straight from this release without cloning the repo. Publishing to npm, PyPI and the MCP Registry is up to the owner and comes later.

Prove it in one line (real XRPL Testnet, no clone)

npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.1/researchagent-mcp-1.0.1.tgz try --testnet

This line runs the full path:

1. Funds faucet wallets.
2. Buys an article by key-reveal escrow.
3. Catches a broken promise and gets a writer-signed refund.
4. Buys again.
5. Anchors the answer.
6. Verifies every transaction on the ledger.

It ends with RESULT PASS. Testnet has no real value. If the faucet fails, the run is labelled SIMULATED and gives the reason.

Ask your own question

# The full web app on Testnet (Node ≥ 22.13)
npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.1/researchagent-mcp-1.0.1.tgz app --testnet

# Any MCP client on Testnet. The operator sets the caps; the default budget is zero.
claude mcp add researchagent -e MCP_MAX_BUDGET_MINOR=200 -e MCP_DAILY_CAP_MINOR=500 -- npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.1/researchagent-mcp-1.0.1.tgz --testnet

Add --iou to either command to pay writers in TEST-SGD through XLS-85 token escrow. TEST-SGD is our own labelled stand-in for RLUSD.

How well it answers (measured once, held out)

We wrote the v4 judge bank blind and ran it exactly once, on main bdc30f97, against the packed MCP tarball:

• 7 of 20 questions bought and cited. Every purchase was verified on Testnet.
• In 2 of 20 the key fact leads the answer.
• 0 of 10 off-topic questions bought anything. All ten v4-negatives abstained.
• 0 gate violations.

The run used the fixture decision model at threshold 0.20, team-authored writers (not independent) and faucet-funded Testnet. v4 is now used, so any future held-out claim needs a new blind bank. On development banks (not held out), v1 scores 16/20 and v3 7/20. Evidence.

Highlights since 1.0.0

• Every DONE Testnet run verifies online (#403).
  • A run that buys nothing now carries the 12-drop audit anchor.
  • Online verification on the frozen banks went from 16/20 to 20/20 (dev) and from 8/20 to 20/20 (v2).
• Lead findings are drawn from purchased passages (#404, #410).
  • The lead finding cites delivered purchased passages. On the held-out v4 bank, the key fact leads the answer in 2 of 20.
  • Filler went from 50 to 0, and the premium-leak audit is 0.
• Declines explain themselves (#404). Only a question with no candidates says "No listed writer covers this". Every other decline names each policy reason, in the app and over MCP.
• 22 sourced XRPL explainers (#407). They come from a disclosed team-authored writer and cite primary references.
• UX polish (#408, #409).
  • No Activity step is ticked before it happens.
  • The payment card reads escrow → key → refund.
  • Repeated answer text is gone.
• Ledger page: 61/61 Testnet transactions validated read-only in 1.65 s, up from 45 rows (#402).
• Leaner tree. 1,012 lines of frozen proof tooling are retired (#412), the Cast QA page is removed (#411), and the packed-consumer and bank runners share one module (#415).
• No buy-threshold change (#413). A constrained sweep on the dev bank found no value that met every constraint, so the threshold stays at 0.20.

Merged since v1.0.0

• Feasibility: fix tsx registration in desktop smoke, recovery and recording scripts (#416)
• Feasibility / Technical depth: consolidate packed consumers and the bank runner (#415)
• Reachability: prepare v1.0.1 (#414)
• Feasibility: fixture threshold sweep; keep 0.20 (#413)
• Retire twelve historical proof generators (#412)
• Remove the retired Cast QA page (#411)
• Lead-fact selection from purchased passages (#410)
• UX follow-ups (#409)
• UX polish (#408)
• 22 sourced XRPL explainers (#407)
• Blind v4 judge banks (#406)
• Judge-bank measurement tooling (#405)
• Cited findings, policy-decline reasons and label-leak prevention (#404)
• Anchored DONE Testnet abstentions and IOU provisioning progress (#403)
• Honesty sweep and first-screen proof (#402)
• Blind v3 judge banks (#401)
• Deck crops (#400)
• Pixel agent v2 and LinkedIn exports (#399)
• Judge coverage: v2 bank and sourced team notes (#398)
• Portable startup and Testnet funding seam (#397)
• Parallel browser CI (#396)
• Scoped test and acceptance resources (#395)
• Project sidebar history (#394)

The macOS DMGs are unsigned until the owner adds an Apple Developer ID. On first launch, right-click the app and choose Open.

Apple Silicon and Intel Mac apps attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

v1.0.0 · ResearchAgent 1.0.0

ResearchAgent 1.0.0

ResearchAgent 1.0.0 contains the exact artifacts CI accepted for this commit. Final local acceptance passed on 467c653: the 2,000-run soak (loop p99 40.3 ms against a 50 ms limit), all release artifacts and clean consumers, the arm64 and x64 DMGs, desktop recovery, check:fast, Python, coverage and chaos. The browser suite was then re-run on this SHA and passed 22 of 22; the only difference is two updated test assertions (#393). You can install them straight from this release without cloning the repo. Publishing to npm, PyPI and the MCP Registry is up to the owner and comes later.

Prove it in one line (real XRPL Testnet, no clone)

npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0/researchagent-mcp-1.0.0.tgz try --testnet

This line runs the full path:

1. Funds faucet wallets.
2. Buys an article by key-reveal escrow.
3. Catches a broken promise and gets a writer-signed refund.
4. Buys again.
5. Anchors the answer.
6. Verifies every transaction on the ledger.

It ends with RESULT PASS. A cold run takes about 84 s: two of four cold runs finished under 90 s, and the other two are kept in the evidence. Testnet has no real value. If the faucet fails, the run is labelled SIMULATED and gives the reason.

Ask your own question

# The full web app on Testnet (Node ≥ 22.13)
npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0/researchagent-mcp-1.0.0.tgz app --testnet

# Any MCP client on Testnet. The operator sets the caps; the default budget is zero.
claude mcp add researchagent -e MCP_MAX_BUDGET_MINOR=200 -e MCP_DAILY_CAP_MINOR=500 -- npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0/researchagent-mcp-1.0.0.tgz --testnet

Add --iou to either command to pay writers in TEST-SGD through XLS-85 token escrow. TEST-SGD is our own labelled stand-in for RLUSD.

Become a writer

npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0/researchagent-mcp-1.0.0.tgz init my-notes
npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0/researchagent-mcp-1.0.0.tgz publish ./my-notes --testnet

From install to a validated XRPL Credential takes 33.9 s. A malformed note is rejected in 1.5 s, with hints for each field.

Install the libraries

npm i https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0/researchagent-sdk-1.0.0.tgz
pip install https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0/researchagent-1.0.0-py3-none-any.whl
researchagent verify receipts.json --ledger   # no Node needed: checks payments, key reveal, refunds and the anchor on Testnet

Highlights since rc.2

• One-line judge path: try --testnet (#387).
  • The MCP package shrank from 21.4 MB to 3.4 MB.
  • Cold MCP start dropped from 56.5 s to 6.3 s, and the app is ready in 27 s instead of 134 s.
• Team-authored notes can supply cited XRPL evidence (#390).
  • Team-authored builder notes are published as a writer and clearly labelled "team-authored · not an independent writer".
  • dev 15/20 at the time; see #398 for held-out. Post-release local candidate: held-out 8/20, dev 16/20 buy-and-cite, fixture models on Testnet. #398
  • All measured runs returned a response; abstentions are not cited answers and no-purchase exports lacked anchors.
  • The flagship answer card is repaired.
• A stranger becomes a writer on the first try (#389): init and publish --testnet validate notes before provisioning and use an ephemeral port. A running app discovers local writers.
• Stablecoin key reveal on the zero-clone surfaces (#384): --iou, with both verifiers reporting PASS.
• Settings (#386): one page for keys (masked, in the Keychain on macOS), network, budget caps, payments, wallets, writers, agent connection, data and about.
  • Only a person can change the caps, and each change is audited.
  • The desktop app gains onboarding, ⌘, for Settings, and saved window state.
• Historical 7 Nov ledger measurement: 45/45 transactions validated live in about 2 s, down from 27.5 s (#388).
  • Currency codes are decoded.
  • The /demo/ proof strip is collapsed.
• A pixel-art research agent (#385) in the README, on the site and in the app's loading state.

Merged since v1.0.0-rc.2

• UX: update two acceptance specs to the current writer provenance label (#393)
• Reachability: prepare v1.0.0 release (#392)
• Feasibility / Reachability: first-try local writer onboarding (#389)
• All criteria via the submission: refresh SFF submission pack (#391)
• Reachability / Feasibility / UX: cited answers for judge questions and repaired flagship card (#390)
• UX: shared owner Settings and macOS desktop integration (#386)
• Reachability / Feasibility: one-shot Testnet judge path and slim MCP (#387)
• UX / Builder feedback / Feasibility: fast ledger proof and fresh judge-path numbers (#388)
• Pixel-art agent character: sprite sheet, README hero GIF, site loop, app loading sprite (#385)
• Technical depth / Reachability: TEST-SGD key reveal on zero-clone MCP and app (#384)

The macOS DMGs are unsigned until the owner adds an Apple Developer ID. On first launch, right-click the app and choose Open.

Apple Silicon and Intel Mac apps attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

Unversioned DMG aliases were removed on 7 Nov. Use the versioned DMGs; the historical desktop-manifest records these as sourceFile (identical bytes and hashes).

v1.0.0-rc.2 · Release candidate 2 · Real Testnet without cloning

Release candidate 2 · Real Testnet without cloning

This is a pre-release containing the exact artifacts CI accepted for this commit. You can install them straight from this release without cloning the repo. Nothing is on npm, PyPI or the MCP Registry yet; 1.0.0 follows final acceptance on 9 Nov.

Ask your own question on real XRPL Testnet with no clone

# The full web app on Testnet (Node 22 only): faucet wallets, key-reveal escrow, an anchored answer
npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0-rc.2/researchagent-mcp-1.0.0-rc.2.tgz app --testnet

# Any MCP client on Testnet. The operator sets the caps; the default budget is zero
claude mcp add researchagent -e MCP_MAX_BUDGET_MINOR=200 -e MCP_DAILY_CAP_MINOR=500 -- npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0-rc.2/researchagent-mcp-1.0.0-rc.2.tgz --testnet

Testnet has no real value. If the faucet fails, the app falls back to a run labelled SIMULATED and states the reason.

Install the libraries

npm i https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0-rc.2/researchagent-sdk-1.0.0-rc.2.tgz
pip install https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0-rc.2/researchagent-1.0.0rc2-py3-none-any.whl
researchagent verify receipts.json --ledger   # no Node: checks payments, key reveal, refunds and the anchor on Testnet

Highlights since rc.1

• Zero-clone real Testnet: app --testnet (#380) and MCP --testnet (#376). Wallets are ready in about 17 s on a cold start.
• The verifier passes our own runs (#377). TS and Python --ledger check every payment, key-reveal proof, refund and anchor on Testnet. Anchoring is the default on Testnet.
• The default /demo/ replay is the live key-reveal UC3 (#379): key published, broken promise, refund, re-buy. Refunds after a key reveal are labelled honestly: writer-signed, the fixture writer cooperates, and a fraud proof is attached.
• Stablecoin key reveal (#381): XLS-85 token escrow with a TEST-SGD IOU from our own Testnet issuer, a labelled stand-in for RLUSD. RLUSD needs only its issuer's AllowTrustLineLocking flag.
• Writer cold start (#378): publish ./notes --testnet gets a validated XRPL Credential in about 44 s. The first sale arrives as an EscrowFinish.
• Ledger page: 45/45 Testnet transactions validated live on load.
• Quiet-host soak passes all five gates (#374). Cost card: 435 drops of fees per key-reveal purchase.

Merged since v1.0.0-rc.1

• Reuse decisions harness scenarios across tests (#383)
• Reachability: prepare rc.2 installs, live anchors and ledger evidence (#382)
• Feasibility / UX: writer cold start (publish --testnet), real-writers list, first-touch polish, judge path (#378)
• Technical depth: TEST-SGD key-reveal token escrow on Testnet (#381)
• Reachability: zero-clone web app (npx app --testnet) (#380)
• Default replay is the live key-reveal UC3; honest refund semantics after key reveal (#379)
• Technical depth: the verifier passes our own runs (key-reveal proof, Python --ledger, anchor by default) (#377)
• Zero-clone Testnet MCP (--testnet), devcontainer Testnet CI, cost card (#376)
• Reachability: RC install links, MCP version and evidence polish (#375)
• docs: establish quiet-host A/B/C and full-soak evidence (#374)

Apple Silicon and Intel Mac apps attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

v1.0.0-rc.1 · Release candidate · Install without cloning

Release candidate · Install without cloning

Pre-release. These are the exact artifacts CI accepted for this commit. You can install them straight from this release without cloning the repo. Nothing has been published to npm, PyPI or the MCP Registry. Version 1.0.0 follows after final acceptance on 9 Nov.

Install without cloning

npm i https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0-rc.1/researchagent-sdk-1.0.0-rc.1.tgz
pip install https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0-rc.1/researchagent-1.0.0rc1-py3-none-any.whl
claude mcp add researchagent -e MCP_MAX_BUDGET_MINOR=200 -e MCP_DAILY_CAP_MINOR=500 -- npx -y https://github.com/Collaboration95/theFastandtheFungible/releases/download/v1.0.0-rc.1/researchagent-mcp-1.0.0-rc.1.tgz

The MCP server spends only up to the caps set by the operator. Its default budget is zero.

Highlights since v0.8.0

• Key-reveal escrow on XRPL Testnet. The writer is paid only by publishing the article key in an EscrowFinish. The buyer checks sha256(K) = Condition before granting. Replay
• Answers anchored on XRPL. researchagent verify --ledger checks the audit-head anchor against Testnet.
• Writer identity through XRPL Credentials.
• Multi-origin writers with source-bound receipts.
• Agent spending kill switch across TypeScript, Python and MCP.
• BYOK, cooperative Stop and bounded active work.
• Key reveal is the default in make testnet. The live app shows the escrow being created, the key published in ledger N, sha256(K) = Condition ✓, decryption ✓, and a CancelAfter countdown. ESCROW_MODE=payment opts out.
• make testnet works from a cold start. Faucet wallets give a first validated hash in about 60–75 s.
• Live ledger page. It validates every committed Testnet hash when it loads.

Soak status

The same-engine soak passes all five gates on a quiet host. Over 2,000 asks, the worst-interval event-loop p99 was 25.8 ms (limit < 50 ms) and peak RSS was 919 MiB. On a heavily loaded shared host the loop gate had failed; a controlled A/B found no regression. Evidence: docs/evidence/soak-nov4-ab/ (#374, merged after this tag).

Merged since v0.8.0

• ci(desktop): download the Electron binary before building the DMG (#373)
• Creativity / Technical depth: key reveal in the live Testnet product (#372)
• Attribute engine loop stalls, reduce Store work, and clean consumer scratch (#371)
• Reachability: prepare v1.0.0-rc.1 artifacts (#370)
• Fix Testnet reachability and first-impression money consistency (#367)
• Creativity / UX: key-reveal escrow on first touch (#369)
• Reduce engine nursery footprint with retained-heap and full-soak evidence (#368)
• Anchor answer audit heads on XRPL Testnet (#366)
• Reachability: zero-clone MCP npm artifact and install evidence (#364)
• Diagnose full-soak memory and make quick RSS slope diagnostic (#365)
• Deliver article keys through writer-paid XRPL escrow (#363)
• test: complete H2 rehearsal evidence and preserve acceptance misses (#360)
• Make UC3 first touch real on XRPL Testnet (#362)
• Make the fixture first run buy, cite and abstain; add Codespaces (#361)
• Make XRPL settlement verify itself on first touch (#359)
• Bound engine verification work and split synchronous research batches (#358)
• feat: verify writer identity with trusted XRPL Credentials (#357)
• Remove retained-history accounting scans with exact spending equivalence (#356)
• docs: link the live Testnet ledger page (#355)
• Docs front door: sharp README, audience-based docs index, CHANGELOG (#354)
• Measure same-engine workload and retain shared-host baseline misses (#353)
• Verify receipts independently in Python without Node (#352)
• Polish receipt contributions, Mac Settings and recovery screen (#351)
• Explain what each purchase contributed with verifiable receipt links (#350)
• Agent spending kill switch across TypeScript, Python and MCP (#349)
• Search independent local writers with source-bound purchases and receipts (#348)
• Make SDK, Python and MCP artifacts independently installable; land browser replay and truthful controls (#332)
• Require coverage, browser and Python gates; recover oversized browser streams (#344)
• Engine BYOK, cooperative Stop and bounded active work (#347)
• Recover Mac startup, BYOK Settings and durable pause (C2) (#346)
• Reliable history projections, bounded streams and late refund recovery (#345)
• feat: ranked XRPL feedback and verifiable Testnet ledger (#343)
• fix: diagnose soak stalls and bound provider waits (#342)
• feat: add typed standard-library Python SDK and fixture CI (#341)

Apple Silicon and Intel Mac apps attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

v0.8.0 · Day 7 · Any agent, one wallet

Day 7 · Any agent, one wallet

Any agent, one wallet

• #333: SDK v1 provides typed asks, streaming, verification and x402 payments. @researchagent/sdk is packed, not published; examples use fixtures.
• #334: The Mac app hosts MCP over loopback HTTP, with per-agent Keychain tokens and allowances, plus verify_answer and hire tools. Its Testnet proof used fixture research and decisions.
• #335: Crash-safe engine restart, notify-only update checks and separate arm64 and x64 DMGs. A packaged app made a Keychain-backed Testnet purchase (transaction A2B630B93E0BAC46D3EA4888E9F6E107715E81DF4B78DF4205CEE35378FEEA5A) using an earlier artifact; crash recovery was tested with fixtures. Builds are unsigned.
• #336: Fully live operational successes improved from 14/30 to 27/30 on a fixed sequence with a synthetic corpus and XRPL Testnet. One lexical audit flag remains; see the PR for its qualification. This is not an unqualified audit pass.
• #337: Starter Kit x402 interoperability in both directions with native XRP on Testnet and synthetic content. Validated transactions: FE6EF0F8785B4FC776551B378B62B8E9BA480D17FA375A3BB637FCE310AD1373 and 570DE6EB58157FB8C5D1F8E23ECAF0A8340750ADFCB6A238CB8E70E2FD4B5588.
• #338: One-click Claude Desktop MCPB and registry-ready server.json; the registry entry is not published. The extension defaults to fixture providers and simulated payments.
• #340: Official Testnet RLUSD: UC2 paid 0.90 RLUSD and UC3 refunded 0.30 RLUSD, with fixture research and synthetic corpus. RLUSD escrow is unavailable because the issuer locking flag is unset; the refund was cooperative.

Apple Silicon and Intel Mac apps attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

v0.7.0 · Day 6 · Proof you can check

Day 6 · Proof you can check

Proof you can check

• #322: npm run verify:answer checks the evidence for an answer. The recorded verification covered 40/40 citations and 4/4 Testnet checks; a tampered file exits 1. #314 adds automatic recovery of verified runs.
• #317: chaos testing recorded 1,729 fault episodes, 431 SIGKILLs and 0 double charges. This was fixture-only testing, with 0 XRP spent.
• #326: refreshed README and site header.

Also on main since Day 3

• Escrow (#308)
• Hire the agent (#306)
• Test USD stand-in (#315)
• Audit export (#309)
• One wallet, many agents (#319)

Apple Silicon Mac app attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

v0.6.0 · Day 5 · Answers, made clear

Day 5 · Answers, made clear

What shipped

• UX lane #312: answer and claims, article-specific Activity, supporting tables/palette/settings, mobile Spend sheet and neutral site. UC3 PDF 50 → 4 pages on the same v0.5.0 snapshot; citation-backed dated lead replaces the author introduction. Rebased onto Lane B #320; UI makes no spending-policy change.
• UX lane #311: shell, environment popover, plan, payments/verification/refund, net Spend and receipt sheets; four allowed keyframes, no in-app cast/odometer. Recovery state clarified; #312 retains answer/supporting-screen/PDF/site work.
• Lane B #276: value-of-information veto, conservative signed-menu duplicate checks, keyed receipt refit and exact intent/answer labels; held-out waste 39.39% → 35.48% (target missed), flow F1 0.588 → 0.625, uplift +15.38 pts unchanged; dev waste 51.79% → 44.19%, flow F1 0.308 → 0.296. Zero gate violations; scorecard/README refreshed.
• Lane B #260: signed corpus fact menus, both decision providers and local scorecards; live F1 0.235 → 0.419, waste 36.19% → 47.19% (both targets missed), zero gate violations; README regenerated, fixture quality unchanged.
• Lane A #259: installed unsigned Apple Silicon DMG, 136,597,522 bytes (130.27 MiB); launch to interactive 2.070 s with fresh fixture data. UC1–UC3, refund, report, Keychain round trip and hardening checks pass; no live payment/signing claim. Evidence.
• Published the branded explainers and shared header to Pages, preserving the evolution dataset and branch-only content; added a manual publish script.
• UI polish: use the brand palette's purse icon in the budget card and correct singular count labels.
• Lane light #272 (pass 1): startup-style README (hero, Claude Code MCP in one command, local run, how a run works, Testnet proof, fixture-labelled scorecard table generated by scripts/readme-scorecard.mjs); MCP and make-target detail moved to mcp.md and running.md; one header across the explainers pages via scripts/site-header.mjs (checked in npm run hygiene; slides and notes keep their own presentation chrome). App GIF and Mac download link wait for pass 2.
• Lane C #271: shipped the open-index brand kit, deterministic SVG/PNG/ICO/ICNS exports, app and explainers-source favicons/OG cards, naming audit, and two restrained cast refinements.
• Lane D #285: typed TypeScript SDK, SDK-backed MCP tools, validated config and mechanical API routers; fixture delivery/idempotency, route parity and unchanged MCP e2e verified.
• Architecture review (REVIEW-NOV-2026.md): four do-now seams filed as #285 (SDK v0, config, route split), #286 (one spend gate, auth, injection hardening), #287 (live reconciler, health, metrics), #288 (audit trail, export, COMPLIANCE.md); fold-ins commented on #259, #273–#275, #277–#281.
• Lane light #269: repo front door. Description, topics and homepage set; MIT licence (revertible until the owner's veto on #265), CONTRIBUTING, SECURITY, Code of Conduct, issue templates, CODEOWNERS; npm run hygiene in check:fast; scripts/release.mjs and a release workflow for one tag per feature day.
• Lane A #270: keyboard-first Run/Writers, one event-derived state model, visible simulation rails, stage-review fixes #225–#231 and isolated recording/18 s clips; 11 production-browser checks pass, UC2 CLS 0.0138.
• Planned the feature-a-day run (9 days, 5–13 Nov, slack of 4), repo identity and the post calendar; filed #269–#281; owner-only items moved to #265 (calendar).
• Lane A #255: reusable isolated engine + four stdio MCP tools; human caps and durable retries; fixture search → simulated purchase → delivered citations proven end to end.

Merged since v0.5.0

• UX-2: answer the question and finish the redesign (#323)
• Reduce wrong-domain spend and retain short fact identifiers (Refs #276) (#324)
• docs(submission): SingHacks finals pack: descriptions, deck PDF from code, video script, pitch and Q&A (#321)
• Preserve purchased facts and report gross/net waste with CIs (#320)
• Give each agent a scoped allowance from one wallet (Refs #280) (#319)

Apple Silicon Mac app attached. Unsigned: move to Applications, then right-click → Open (Privacy & Security → Open Anyway if blocked). See desktop/SIGNING.md.

UX-2: answers, made clear

PR #323, merged at 2e415d4caf227c1125c4d101945495329dc0e767.

The lead finding ranks exact sentences from validated, delivered claims against the question, rather than taking the author's first sentence. UC3 now leads with the dated 26 → 18 week comparison and its supporting citation. Activity names each writer and amount, and collapses after completion. Supporting tables, mobile Spend, the palette and site source follow UX-DIRECTION.

The same v0.5.0 UC3 snapshot exports in 4 pages instead of 50: two pages of content plus two pages of unique cited passages. Access and citation gates remain in force. Four allowed keyframes remain; one environment indicator, with mandatory refund/receipt/export labels.

Local verification on the rebased code: 695 fast checks, 14 production browser tests, UC1–UC4, mobile, axe, PDF and performance. Mac application staging builds. Record mode and the new 20-second UC3 clip work. Clip, screenshots and PDF use an offline demo, fictional writers and simulated payments; they are not evidence of live model or Testnet settlement acceptance.

Watch the 20-second UC3 refund clip · Four-page sample PDF · Measurements · Asset provenance

Site source is aligned; gh-pages publishing remains the separate owner step specified in #312. Writers' article counts are scoped to the current run. Engine stance classifications are preserved.

Original 50-page PDF, same UC3 snapshot is also attached so the reduction can be checked directly.