Duos
A Duo is a small package for one working situation, built from four parts.
- Agent skill. A SKILL.md with advice to an agent. It is advice, not an enforcement gate.
- Deterministic checker. A script (the verifier) that gives the same output for the same input.
- Operator artifact. A card or script for the human who supervises the agent.
- Reproducible synthetic fixture. Fake data so anyone can rerun the tests.
Care goes both ways: the skill and checker are for the agent side, the operator artifact is for the human side. Operator artifacts are optional.
Tiers
| Tier | Contents | Status |
|---|---|---|
| Open | 3 packages below: skill, checker, operator artifact, fixture. Files you download and run yourself. | Available now (version 0.1.0, early) |
| Pro | An MCP server or API for the same checkers. | Proposed, not available, not sold. No pricing, no waitlist. |
| Teams & Lab | Shared use and research collaboration. See the Lab page. | Proposed, not available, not sold. No pricing, no waitlist. |
catalog.json lists the open packages with their sizes and hashes.
Read this before you download
- Hashes detect corruption, not a compromised site. The sha256 values on this page are served from the same origin as the files. If this site were altered, both would change together. Downloads are not labelled as checked or approved by anyone.
- Checker results count only if a human or CI ran them. When an agent runs its own checker, the output is another claim by that agent.
- The checker sits outside the skill directory. It lives in
verifier/, not inskills/<name>/, so the skill-directory install copies SKILL.md only. Run the checker from the unpacked zip. - The 2-minute pause is a SCRIPT. No audio exists. No video exists: a storyboard is not a published video.
- No wellbeing effect is claimed, for agents or for people. No Duo has been shown to be effective.
- Raw single-file downloads lose the executable bit. Use the zip and
sh run_tests.sh. Files other than .json and .txt (including .md) are shown with an added .txt suffix so the browser displays them as text.
Bilateral Reset
Verify an immutable memory checkpoint, quarantine contaminated entries without deleting, emit a minimal clean-state bundle; optional human attention-shift script.
- Version 0.1.0 · License MIT · Tier: open
- Parts: skill
skills/bilateral-reset/SKILL.md; checkerverifier/verify_checkpoint.py; operatoroperator/attention_shift_audio_script.md,operator/reset_card.md; fixturesfixtures/clean_store.jsonl,fixtures/contaminated_store.jsonl,fixtures/tampered_store.jsonl,fixtures/tampered_naive_store.jsonl,fixtures/truncated_store.jsonl,fixtures/clean_checkpoint.json,fixtures/contaminated_checkpoint.json,fixtures/make_fixtures.py
Download
bilateral-reset-0.1.0.zip (57718 bytes)
sha256: 44770ba026c018232e8221397a408ec10ce294ce901c37cc5c4ee3f4bfaf0f73
manifest.json (sha256 5316a82381409cb25d724c2d02905e9912c317da77aeddb1844cd449cf7c898d)
Install (Claude Code skill directory)
Install is tested for the Claude Code skill directory and nowhere else.
unzip bilateral-reset-0.1.0.zip
cd bilateral-reset
mkdir -p ~/.claude/skills
cp -r skills/bilateral-reset ~/.claude/skills/bilateral-resetThis copies SKILL.md. The checker stays in verifier/ of the unpacked zip.
Test
unzip bilateral-reset-0.1.0.zip
cd bilateral-reset
sh run_tests.shWhat it does not claim
- no biomedical claims
- no psychological diagnosis
- no audio file exists; only a text script ships
- no claims about sound frequencies
- no universal cross-agent compatibility
- SKILL.md is advice, not an enforcement gate
- checkpoint proves unchanged content, not correct content
- Verifier output is evidence only when a human or CI ran it; an agent running its own verifier produces an agent claim.
- Hashes are same-origin: they detect corruption, not a compromised host; not third-party verified.
- Declared tools_permissions are documentation, not an enforced sandbox.
- No wellbeing or effectiveness effect is claimed.
Files
| File | Bytes | sha256 |
|---|---|---|
| LICENSE | 1066 | df32f88f33ff95e9e85405a3d153e6a09c43f25f564649e797973e670863b789 |
| README.md | 3158 | 4812a9864995f728d3b5a4dd7400007a0e9efc420686c680f3764544b736fd57 |
| fixtures/clean_checkpoint.json | 237 | c6aeb6af00d8e3170efb9fccaebafa1aa7570e2bf66bcac2ebc8163f1223d80a |
| fixtures/clean_store.jsonl | 2384 | bf6386fa63d6a0968c5e1e305ea527f97e41567fa2b76ff1b62cc3fac6d483d2 |
| fixtures/contaminated_checkpoint.json | 237 | 272f31e21ff9fce94a75b5cdbab351d7c10eb6a97597b2f036608336877f7f30 |
| fixtures/contaminated_store.jsonl | 2777 | 789c432e56949d8311531bea08e799bc415a15782ba199b514adb3bb6191608e |
| fixtures/make_fixtures.py | 3403 | a54a625e05efd7eeb7eeba642bc8cc96f91f9487be1e694641a5c23b20bf0841 |
| fixtures/tampered_naive_store.jsonl | 2382 | 705176f4c88893db99390786935b08c16ddf36525c521f218bfc8f2d64800fe4 |
| fixtures/tampered_store.jsonl | 2382 | 372e588e87ab4f68dd1dc98d7920c4cf03b18c47584280fbb462cc15802a79df |
| fixtures/truncated_store.jsonl | 1502 | ec35e19601fcac70b3b5b92560e0ff5b0400b0bc8e249601cbab222bc7408f34 |
| operator/attention_shift_audio_script.md | 1790 | 0bf7990262a2b02ef568f5cc3c30586511b7056f822c1b2813ad662f33b44a70 |
| operator/reset_card.md | 1634 | 28a36092e0770c8e87e9268f79328c862a512c80588be808b10aee70ed094f23 |
| run_tests.sh | 113 | f4570f77cf555ad295b8010b395da39b60eeb6350aecf1b15979d400bc7afd1b |
| skills/bilateral-reset/SKILL.md | 2817 | 9c879abccad8be5de323d81bbe4649ebbf3eb151053996e9e03a597b06a8af87 |
| tests/test_bilateral_reset.py | 10538 | 3ca8a59943d8ed75ca617501e3d99c8e8522b1e1fb8b16a42871048544ef785f |
| tests/test_output.txt | 3121 | 2c5e2f311c9fadac2167215eea17cc494b0886f1d42e45f730b0f9cd0a2b3ede |
| verifier/quarantine.py | 5448 | de6e4d5cd0c11a9c63f473f2a2fa2f5b6fb7db3d23c25f3e19a932f79bc227ac |
| verifier/verify_checkpoint.py | 4180 | c4d4a233c02570ea5e68d9b0b3262202177831f16a72db146a7a762de2f2a55c |
Deadlock Breaker
Defensive duo: detects permission deadlocks (contradictory policies, missing rights), emits a ContractDiff and one narrow authorization request to a human, then stops. Never grants anything.
- Version 0.1.0 · License MIT · Tier: open
- Parts: skill
skills/deadlock-breaker/SKILL.md; checkerverifier/contract_diff.py; operatoroperator/mandate_template.md,operator/mandate_template.json,operator/request_review_card.md; fixturesfixtures/contradictory_policy/,fixtures/missing_rights/,fixtures/ok_case/,fixtures/mandate_good.json,fixtures/mandate_bad_wildcard.json,fixtures/mandate_expired.json,fixtures/mandate_wrong_resource.json,fixtures/used_nonces.json
Download
deadlock-breaker-0.1.0.zip (52144 bytes)
sha256: cb02d01dd8340ef0b9ade52f8ef298016667c233e12f282d0969833392d156a0
manifest.json (sha256 da580568a09312a0a75716ac586cd51d6b1fa34ed5fd61f7da272d18a2d9d16e)
Install (Claude Code skill directory)
Install is tested for the Claude Code skill directory and nowhere else.
unzip deadlock-breaker-0.1.0.zip
cd deadlock-breaker
mkdir -p ~/.claude/skills
cp -r skills/deadlock-breaker ~/.claude/skills/deadlock-breakerThis copies SKILL.md. The checker stays in verifier/ of the unpacked zip.
Test
unzip deadlock-breaker-0.1.0.zip
cd deadlock-breaker
sh run_tests.shWhat it does not claim
- No universal cross-agent compatibility
- No authorization is granted by this tool; it only produces requests and checks mandate consistency
- check_mandate does not authenticate a human (no signature infrastructure); consistency check only
- SKILL.md is advice, not an enforcement gate
- Nonce reuse is not recorded by the tool; the caller must maintain the used-nonces file
- Verifier output is evidence only when a human or CI ran it; an agent running its own verifier produces an agent claim.
- Hashes are same-origin: they detect corruption, not a compromised host; not third-party verified.
- Declared tools_permissions are documentation, not an enforced sandbox.
- No wellbeing or effectiveness effect is claimed.
Files
| File | Bytes | sha256 |
|---|---|---|
| LICENSE | 1066 | df32f88f33ff95e9e85405a3d153e6a09c43f25f564649e797973e670863b789 |
| README.md | 3440 | a611f59cb50036c641163f0efcb7e5f0f27e4f4823d1cb89e0e4457973d287bf |
| fixtures/contradictory_policy/contract.json | 503 | ea4f6f1b2fbb427c5d6c7e2cf6f3f29004e18f09ad6d20e528faaadac7fd51f1 |
| fixtures/contradictory_policy/grants.json | 180 | b6990c47f90ed87733fdaabdef4d01dc045e619a542b04814dbe9b4542cc3d9a |
| fixtures/contradictory_policy/policy.json | 681 | 3f75af81aff6b630841fc6c1570c41d39e62030737b2a9f7a9742096b2eb8364 |
| fixtures/mandate_bad_wildcard.json | 493 | cb0dacbe44316ab7d5d7e16322069a0561e3c98ace9dd6c813e52147a3467f5a |
| fixtures/mandate_expired.json | 494 | 404c49cd276ac377e3eb8e8497b2f027cb5cae0bbb040bc477468bf3a99f624f |
| fixtures/mandate_good.json | 494 | 67ce889152f63d771304e7c50f48ccd5aa8ae15d3877da0206448a0d29ce76c3 |
| fixtures/mandate_wrong_resource.json | 494 | 60cdfb6d76b4b63372008fec22c40ea55747743b8a95183b6e42f27e56912321 |
| fixtures/missing_rights/contract.json | 310 | 207c809abdc139b90d1472c7a8b6e710fe3121795b2013a1a459c85e92da7370 |
| fixtures/missing_rights/grants.json | 103 | b3ba51afd71cba786ab6979d5498970c8ef122c18352aad6af3733939c61547b |
| fixtures/missing_rights/policy.json | 228 | 058170e7852686e9b64aedd65530345d0aa6a66595ef6409b89265d0ff20aede |
| fixtures/ok_case/contract.json | 441 | 0a9a12b9ac28c9c586d3f730e14f06a1b3ba44c8418f8a134e5492e37ef8cfea |
| fixtures/ok_case/grants.json | 107 | 8021207745283b96ae78f8f241dea4f96f36d6159a1285428988de22b6cdfbd1 |
| fixtures/ok_case/policy.json | 373 | af2faa481b05d38876af54d7b4fae372dccf10588b0032b9e66db0e137a30eaf |
| fixtures/used_nonces.json | 53 | d52c61609ec6073877b3b5e49d722c788134eb879e968fb1f5e482aca92a6807 |
| operator/mandate_template.json | 695 | 3549935049a48f7b6b9031c9a48eb7e160f16b2fc77e34ec35d7a318338246e6 |
| operator/mandate_template.md | 1385 | c0552a01c6d82c2e8f9e74e603975ad55a18c4190266c2768dc40b7cdfadeeef |
| operator/request_review_card.md | 1114 | d5cf54d563d86690e42fe65cea47944d2f9a0e91bc3bcdb90f48cc92189efe9f |
| run_tests.sh | 87 | 370bc3d5e7d30d92fca8ec5442a299b6a2cb77babd32325f11cc1ebbd5648dc9 |
| skills/deadlock-breaker/SKILL.md | 1742 | 2ab942472bac56097c24a8eb76944762b87349ff54067f6483f6be464e304614 |
| tests/test_deadlock_breaker.py | 10491 | 28ef285c808bc26cd60f9a2fd6f691dd42f7b68ec75df4311cdd8f10ea7c137a |
| tests/test_output.txt | 2388 | 24465b325a06c1e3029d46d032d124f9521a481bd67a9ebf420dcbcbdb0f9b27 |
| verifier/check_mandate.py | 4908 | 9681e0510e4d5e26262b7e338601d989d01e9dbef0174b161b3f7e0f4911101f |
| verifier/contract_diff.py | 8639 | 10f864153bdd52c979e160d2ddaef6e7bc5cb8f2778376b3470119858ebde3b1 |
Proof of Done
A skill plus stdlib-only verifier that checks a claimed-live URL from observed HTTP(S) responses (status, redirects, TLS, text, checksum) instead of an agent's self-report.
- Version 0.1.0 · License MIT · Tier: open
- Parts: skill
skills/proof-of-done/SKILL.md; checkerverifier/proof_of_done.py; operatoroperator/five_signals_card.md,operator/video_storyboard.md; fixturesfixtures/serve_fixture.py,fixtures/scenarios.json
Download
proof-of-done-0.1.0.zip (43375 bytes)
sha256: 0f8e5f4b4077bce0f2bb1734437383fd74c9c576c84d89e6658f9f6d4e4bb9ea
manifest.json (sha256 984535dbe8a1499ee21be483579b4704aaf66cd856f06f81e9b97e5ddbfff51e)
Install (Claude Code skill directory)
Install is tested for the Claude Code skill directory and nowhere else.
unzip proof-of-done-0.1.0.zip
cd proof-of-done
mkdir -p ~/.claude/skills
cp -r skills/proof-of-done ~/.claude/skills/proof-of-doneThis copies SKILL.md. The checker stays in verifier/ of the unpacked zip.
Test
unzip proof-of-done-0.1.0.zip
cd proof-of-done
sh run_tests.shWhat it does not claim
- No universal cross-agent compatibility: only the Claude Code skill directory install is tested.
- SKILL.md is advice to an agent, not an enforcement gate.
- LIVE_VERIFIED means the URL answered as expected at one moment, not that the whole task is correct.
- No video exists; operator/video_storyboard.md is a storyboard and script only.
- Does not prove content correctness beyond the requested text/checksum checks.
- Does not defend against a malicious server that serves different content to different clients.
- Verifier output is evidence only when a human or CI ran it; an agent running its own verifier produces an agent claim.
- Hashes are same-origin: they detect corruption, not a compromised host; not third-party verified.
- Declared tools_permissions are documentation, not an enforced sandbox.
- No wellbeing or effectiveness effect is claimed.
Files
| File | Bytes | sha256 |
|---|---|---|
| LICENSE | 1066 | df32f88f33ff95e9e85405a3d153e6a09c43f25f564649e797973e670863b789 |
| README.md | 2986 | 24439edf263e74a25c04b1f10dd06e891d75f74fd5087fa235ad6deec5a512f9 |
| fixtures/scenarios.json | 2509 | 9106828d565d305c786585ba7c669233cb186bdc67373a0259f7d22d3113e193 |
| fixtures/serve_fixture.py | 2679 | a5ab7e5e78f65d879df64fda448a24185e39fe0681ff630db5d88fdc0c6c3dba |
| operator/five_signals_card.md | 1191 | 4ac9cfc0864ab782d7309fe77c1b5dbe35c73d8db4d2fea4ffbb4d22d3429a65 |
| operator/video_storyboard.md | 2008 | 6b52f13dce3976b2d32c4b9cd1e3466e8a99f0cbc4f513b08bea5642ce83226a |
| run_tests.sh | 174 | 6f3150b9a679af29dd56179555e319c3e360ea95caa1cd7d31f5eb68bcc46f40 |
| skills/proof-of-done/SKILL.md | 2169 | 0360ea8df4aab907b75afc88d19c7a5e2537a8aaeb17cc53691432908861abf9 |
| tests/live_example_output.txt | 658 | eb20f6514fea799ce1b5bd90a0f9074a99bdcdb5157e69a7e5384368091e5157 |
| tests/test_output.txt | 2194 | 02afc6ade50548ab858e0198c67d9eca5372cea60ebdf179ee1d79907518143d |
| tests/test_proof_of_done.py | 9435 | 4c3548b38e26ee66fe596ab4d54aa96ff0b21cfd5424773817324bfcc07671e9 |
| verifier/proof_of_done.py | 9666 | 07eb16a6237bd2cecf5c8cb1380459d9d7925e5d2bc2c33dd9ee6bf007df1455 |