Since SingHacks21 Sep 2026 → 13 Nov 2026
At SingHacks, ResearchAgent was one web page where a language model proposed each purchase and a person approved every quote. Since then it has become released software: a web app, a Mac app, an MCP server and two SDKs. Policy code buys evidence from writers within your budget, and every payment leaves a receipt on the XRP Ledger.
Facts at 484856c · generated 13 Nov 2026 · how we count · data.json
01Every day since SingHacks
347 pull requests and 403 commits since 21 Sep 2026. The busiest day was 12 Nov, with 66 merges. Numbered marks open the stop on the rail below.
Pull requests merged into main per day, Singapore time. Hover or tap a day for commits. The October build has its own archive: October build log.
| Day | PRs merged | Commits |
|---|---|---|
| 19 Sep | 1 | 47 |
| 20 Sep | 0 | 4 |
| 21 Sep | 0 | 1 |
| 22 Sep | 0 | 2 |
| 2 Oct | 0 | 1 |
| 4 Oct | 17 | 46 |
| 5 Oct | 6 | 6 |
| 6 Oct | 4 | 5 |
| 7 Oct | 26 | 35 |
| 8 Oct | 4 | 5 |
| 9 Oct | 11 | 15 |
| 10 Oct | 2 | 2 |
| 11 Oct | 3 | 3 |
| 2 Nov | 6 | 6 |
| 3 Nov | 54 | 60 |
| 4 Nov | 40 | 40 |
| 5 Nov | 3 | 3 |
| 7 Nov | 34 | 34 |
| 8 Nov | 4 | 4 |
| 10 Nov | 9 | 10 |
| 11 Nov | 51 | 52 |
| 12 Nov | 66 | 66 |
| 13 Nov | 7 | 7 |
02The release rail
4 milestones and 15 releases. Open a stop for what shipped, a screenshot or terminal output, and the proof: the pull request, the release notes or a ledger transaction.
The prototype that took us to the finals: one web page, synthetic sources, a language model proposing what to buy, and a person approving every quote.
What shipped
Writers were paid at their own development-network wallets, and the run showed the ledger entry.
What shipped
Purchases moved from the language model to deterministic policy code, inside a budget set once.
What shipped
An on-stage demo of the October build for AI Tinkerers Singapore, a separate chapter from the finals.
What shipped
What shipped
claude mcp add researchagent -e MCP_MAX_BUDGET_MINOR=200 -e MCP_DAILY_CAP_MINOR=500 -- npx -y …/releases/download/v1.1.1/researchagent-mcp-1.1.1.tgzWhat shipped
What shipped
What shipped
What shipped
COMPLIANCE.md.
What shipped
What shipped
npm run verify:answer checks an exported answer: 40/40 citations and 4/4 Testnet checks in the recorded run; a tampered file exits 1.PASS checks: 57 PASS, 0 FAIL, 2 SKIP
PASS audit anchor: Validated XRPL Testnet ledger 22057921
PASS testnet 12/12 (read-only, s.altnet)
RESULT PASSWhat shipped
verify_answer and hire tools.
What shipped
What shipped
What shipped
--testnet --iou.try --testnet judge path: UC3 with a S$2 cap, six numbered ledger links, --ledger verification and RESULT PASS.init <dir> scaffolds two commented example notes.Testnet · no real value · Fixture models and fictional writers
1. Create: testnet.xrpl.org/…/9126E19D…
2. Finish · sha256(K) = Condition ✓: testnet.xrpl.org/…/805AA52F…
3. Refund · writer-signed · fixture writer cooperates (a real writer could refuse): testnet.xrpl.org/…/BFAC5EF0…
4. Re-buy create: testnet.xrpl.org/…/278087AB…
5. Re-buy finish · sha256(K) = Condition ✓: testnet.xrpl.org/…/484A818A…
6. Anchor: testnet.xrpl.org/…/22081E19…
…
RESULT PASSWhat shipped
What shipped
What shipped
What shipped
PASS checks: 46 passed, 0 failed, 2 skipped
PASS audit anchor: Validated XRPL Testnet ledger 22228575
PASS testnet 11/11 (read-only, s.altnet)
Skipped is not failed. PASS does not prove who a writer is.
RESULT PASS03Then and now
Each row starts at the SingHacks prototype and ends at what ships today, with the pull request that first delivered it.
| Capability | At SingHacks | Now | First shipped |
|---|---|---|---|
| 01Payment | A simulated by default; development-network optional | Each writer paid at its own development-network wallet, one charge per intent | 5 Octreleased in v0.1.0#101 #103 |
| 02Who decides to buy | A language model proposed; a person approved every quote | Policy code buys within the budget you set, after one plan confirmation; a decision model only scores | 8 OctCheckpoint tag |
| 03Agents | One web page | An MCP server any agent can reach by URL, an x402 paid endpoint and TypeScript and Python SDKs | 3 Novreleased in v0.1.0#266 #333 #523 #538 |
| 04Refunds | None | An opt-in holdback returns a share of the price without the writer; the rest pays on key reveal; no-show refunds use EscrowCancel after the deadline | 3 Novreleased in v0.5.0#308 #540 |
| 05Mac app | None | An installable Mac app with keys in the Keychain (unsigned) | 3 Novreleased in v0.4.0#303 |
| 06Proof | Citations checked inside the app | An offline verifier anyone can run, a protocol model, simulated crash schedules, and a large batch of real purchases each re-read from a public node | 3 Novreleased in v0.7.0#322 #530 #533 |
| 07Writers | Team-run only | Anyone can publish a folder of notes, or a note from a phone through moderation, as a paid writer | 3 Novreleased in v0.3.0#300 #389 #532 |
| 08Fair exchange | Pay first, then read | Key-reveal escrow: the writer is paid only by publishing the key | 4 Novreleased in v1.0.0-rc.1#369 |
| 09Stablecoin path | XRP only | Token escrow in TEST-SGD, our stand-in: official RLUSD issuers have escrow locking off, so RLUSD cannot be escrowed yet | 4 Novreleased in v1.0.0-rc.2#381 #384 #525 |
| 10Audit trail | None | Validated answer anchors are recorded on the ledger, even when nothing is bought; pending or unavailable anchors are labelled | 4 Novreleased in v1.0.0-rc.1#366 #403 |
| 11First try | Clone, install, add a model key | One npx command ends in RESULT PASS with six ledger links | 7 Novreleased in v1.0.0#387 |
| 12Sources | Synthetic sources only | The real XRPL and x402 documentation, plus ledger reports signed to a validated ledger | 12 Nov#527 #536 #537 |
04The numbers, honestly
Every chart names its source file. Development and blind measurements are never mixed, and nothing was rerun until it passed.
Measured with an empty npm cache. Each row is on its own scale.
MCP server starts
56.5 → 6.3 s −89%
Testnet app ready
134.0 → 27.0 s −80%
Download size
21.4 → 3.4 MB −84%
release candidate 2 today
Source: round4-j2
| Measure | Before | Now |
|---|---|---|
| MCP server starts | 56.5 s | 6.29 s |
| Testnet app ready | 134.0 s | 26.98 s |
| Download size | 21.4 MB | 3.42 MB |
2 of 4 cold runs of try --testnet finished inside 90 s. The misses stay on the record.
Source: round4-j2
| Run | Seconds | Result | Evidence |
|---|---|---|---|
| Run 1 | 73.76 | passed | try-initial-timing.json |
| Run 2 | 90.17 | timed out | try-timeout.json |
| Run 3 | 90.00 | deadline exceeded | try-exact-timeout.json |
| Run 4 | 84.19 | passed | try-timing.json |
Development banks guided the work; the blind bank is the number we report. Off-topic purchases: 0/10.
Fixture research and decision models · offline answers · XRP Ledger · 11 Nov 2026 · leading fact 7/20.
development blind held-out not run yet
Source: v5-heldout
| Bank | Bought and cited | Source |
|---|---|---|
| v1 · development | 16/20 | r5-p0/summary.json |
| v3 · development | 7/20 | p1-threshold/sweep-summary.json |
| v4 · blind held-out | 7/20 | v4-heldout/README.md |
| v5 · blind held-out | 16/20 | v5-heldout/README.md |
Curated ledger transactions on the ledger page grew from 9 to 61, each tied to a committed source file.
From 38 on 20 Sep to 2,127 today, counted the same way at every checkpoint. 1,369 ran in the fast check on 7 Nov.
Static count of test declarations at each tag (definition) · executed count
| Checkpoint | TypeScript | Python | Total |
|---|---|---|---|
| 7 Sep · prototype | 8 | 0 | 8 |
| 20 Sep · SingHacks | 38 | 0 | 38 |
| 8 Oct · October | 302 | 0 | 302 |
| 3 Nov · v0.1.0 | 468 | 0 | 468 |
| 7 Nov · v1.0.0 | 892 | 21 | 913 |
| 8 Nov · v1.0.2 | 987 | 21 | 1008 |
| 13 Nov · today | 2093 | 34 | 2127 |
3,828 lines on 20 Sep, 90,733 today. Tests and tooling included; evidence files and generated JSON are not.
Nonblank lines in tracked TypeScript, JavaScript, Python and shell files (definition)
| Checkpoint | Lines |
|---|---|
| 7 Sep · prototype | 970 |
| 20 Sep · SingHacks | 3,828 |
| 8 Oct · October | 12,837 |
| 3 Nov · v0.1.0 | 20,738 |
| 7 Nov · v1.0.0 | 50,752 |
| 8 Nov · v1.0.2 | 52,836 |
| 13 Nov · today | 90,733 |
61 transactions across 3 days. Each square opens the transaction on the public XRP Ledger explorer.
Source: the ledger page, which re-checks every hash live · transactions.json
05Misses we kept
A progress page that only shows wins is marketing. These are measurements that missed a target, regressed, or replaced a better-looking number of our own. Each links to the evidence.
What we did Kept the timeouts beside the passes; the 90 s limit stayed unchanged.
What we did Reported the blind number instead of the dev bank, then froze and ran a new blind bank (v5, 11 Nov).
What we did The v2 held-out score was written after the team notes; a blind v4 bank replaced it everywhere.
What we did Kept the failure; a quiet-host rerun passed all five gates. No threshold changed.
What we did Show provisioning progress in the app; the target stays.
What we did Kept the default 0.20 and recorded that no value passes every constraint.
What we did Shipped the better selection, logged the regression and opened a follow-up.
06Builder feedback
Every friction point with XRPL, x402 and the development network goes into a dated log with evidence, a workaround and the mainnet impact.
07Next, and what we don't claim
—Method
Counts are dated and query-defined: facts at 484856c (13 Nov 2026). SingHacks ran 19–20 Sep 2026; the prototype was frozen earlier as tag hackathon-baseline (commit 935532f, 7 Sep), and the closest snapshot to the event's end is e16435a (20 Sep). Both are shown rather than picking one.
| Metric | Value | Definition |
|---|---|---|
| PRs merged | 347 | gh pr list --state merged --base main, merged on or after 21 Sep 2026 (SGT), merge commit reachable from 484856c2 |
| Commits | 403 | git rev-list --count e16435a3..484856c2 (after the last commit before 21 Sep 2026) |
| Releases | 15 | gh release list, published, tag reachable from the counted commit (prereleases included) |
| Testnet transactions | 61 | rows in site/ledger/transactions.json; the ledger page re-checks each one live |
| Held-out | 16/20 | v5 blind bank, bought and cited, run once on 11 Nov (docs/evidence/v5-heldout) |
| Tests in the repo | 2,127 | static count of it()/test() declarations in *.test.* / *.spec.* files plus Python def test_ (not executions) |
| Lines of code | 90,733 | physical nonblank lines in tracked TS/JS/Python/Shell files, tests and tooling included |
Rebuilt with npm run site:progress from git, GitHub, CHANGELOG.md, docs/SFF-PUSH.md (latest entry 12 Nov) and the evidence files linked above. Generated 13 Nov 2026, 10:41 SGT. Missing data stops the build instead of showing a stale number.