session string | date timestamp[s] | title string | summary string | messages int64 | findings int64 | models list | example string | files list |
|---|---|---|---|---|---|---|---|---|
2026-08-30-forgedesk-2 | 2026-08-30T00:00:00 | ForgeDesk, part 2: inherit a partial product, verify it, complete it | A session that inherits the partial ForgeDesk from part 1, refuses to trust its documentation, verifies the code, finds and fixes two high-severity authorisation holes, and closes with a final audit that traces every claim to the code: 123 messages, 1h31, 27,800 lines of Python, 174 tests in 13 files, no third-party pa... | 60 | 15 | [
"Claude Haiku 4.5",
"GLM 4.7",
"GPT-5.6",
"Gemini 3.7 Flash"
] | examples/forgedesk-2 | [
"00-brief.md",
"00-summary.md",
"01-plan.md",
"02-record.md",
"03-findings.md",
"04-closeout.md",
"05-positive-memory.md"
] |
2026-09-05-shiftboard | 2026-09-05T00:00:00 | ShiftBoard: a visual production planner, seven roles, two correction cycles | The most demanding brief of the set: an interactive scheduler for a small workshop, with constraints, conflicts, undo, KPIs, scenarios and documentation. Seven roles, seven findings in two correction cycles, one word from the person after the brief. 48 messages, 0h50, 2.49M tokens, no flagship model. | 25 | 7 | [
"Claude Haiku 4.5",
"GLM 5.3 Flash",
"GPT-5.6",
"Gemini 3.6 Flash",
"Gemini 3.7 Flash"
] | examples/shiftboard | [
"00-brief.md",
"00-summary.md",
"01-plan.md",
"02-record.md",
"03-findings.md",
"04-closeout.md",
"05-positive-memory.md"
] |
2026-09-06-queuelens | 2026-09-06T00:00:00 | QueueLens: a first version that looked complete, and what independent checks found | A person writes a nine-requirement brief, approves the plan with one word, and comes back forty minutes later to a finished application: six problems found by two independent roles, corrected, re-verified, and traced back to the brief by a final audit. Six roles on four models, none of them a flagship: 44 messages, 0h4... | 23 | 6 | [
"Claude Haiku 4.5",
"GLM 5.3 Flash",
"GPT-5.6",
"Gemini 3.7 Flash"
] | examples/queuelens | [
"00-brief.md",
"00-summary.md",
"01-plan.md",
"02-record.md",
"03-findings.md",
"04-closeout.md",
"05-positive-memory.md"
] |
2026-09-07-api-contract-explorer | 2026-09-07T00:00:00 | API Contract Explorer: two views that broke on first use, three security holes, one reviewer swapped mid-session | A developer tool from a ten-requirement brief: import an OpenAPI document, browse, inspect, search, compare two operations, export Markdown. The first version passed its own checks; the Reviewer broke the Compare and Export views on the first ordinary click, Security found three holes, all six findings were fixed and r... | 27 | 6 | [
"Claude Haiku 4.5",
"Claude Sonnet 5",
"GLM 5.3 Flash",
"GPT-5.6",
"Gemini 3.7 Flash"
] | null | [
"00-brief.md",
"00-summary.md",
"01-plan.md",
"02-record.md",
"03-findings.md",
"04-closeout.md",
"05-positive-memory.md"
] |
2026-09-07-shipgate | 2026-09-07T00:00:00 | ShipGate: fifteen findings, one reviewer timeout, no one asked | A release-readiness tool for a small software team, from a six-sentence brief. Fifteen findings across four correction cycles, all closed; one Reviewer attempt hit the thirty-minute limit and Tandem reassigned the work by itself; the person typed one word after the brief. 69 messages, 1h50, 4.67M tokens, no flagship mo... | 33 | 15 | [
"Claude Haiku 4.5",
"GLM 5.3 Flash",
"GPT-5.6",
"Gemini 3.7 Flash"
] | null | [
"00-brief.md",
"00-summary.md",
"01-plan.md",
"02-record.md",
"03-findings.md",
"04-closeout.md",
"05-positive-memory.md"
] |
2026-09-23-pricelist-import | 2026-09-23T00:00:00 | Supplier price-list import tool | A command-line importer for supplier price lists, built from a twelve-line brief: six roles from five model families, two words from the person, seven findings raised by three independent roles and all closed, and a final audit that found a gap in the review's coverage rather than in the product; 91 messages, 1h34, 78 ... | 48 | 7 | [
"Claude Opus 5.5",
"Claude Sonnet 5",
"GLM 5.3 Flash",
"GPT-6 Luna",
"GPT-6 Sol",
"Kimi K3"
] | examples/pricelist-import | [
"00-brief.md",
"00-summary.md",
"01-plan.md",
"02-record.md",
"03-findings.md",
"04-closeout.md",
"05-positive-memory.md"
] |
2026-09-24-fx-paper-trading | 2026-09-24T00:00:00 | FX Paper Trading and Options Risk Terminal | A dealing and risk terminal for simulated FX trading on a live public feed, with vanilla options priced by Garman–Kohlhagen and delta hedging from the risk view: seven roles from four model families, one word from the person, nine findings raised by three independent roles and all closed, the main one held open by the ... | 45 | 9 | [
"Claude Haiku 4.5",
"Claude Opus 5.5",
"GPT-6 Luna",
"GPT-6 Sol",
"Kimi K3"
] | examples/fx-paper-trading | [
"00-brief.md",
"00-summary.md",
"01-plan.md",
"02-record.md",
"03-findings.md",
"04-closeout.md"
] |
Tandem session records
Complete records of seven real sessions in which long delegated software work was run under a written plan, in bounded steps, with separate roles for planning, production, review, security and audit, often bound to different model families. Every message is included, from the brief to the closeout. Five of the applications the sessions produced are included as delivered, under MIT, next to their records.
The same material is published at https://github.com/jaryn-io/tandem. This dataset exists so that the records can be found and used as data: long-horizon multi-role traces with independently verified findings and the resulting code.
What a record contains
| File | Content |
|---|---|
00-brief.md |
The assignment as written by the person |
00-summary.md |
What happened, in one page |
01-plan.md |
The plan approved by the person before any work ran |
02-record.md |
Every message in order: role outputs verbatim, system messages to the person summarised in one line |
03-findings.md |
Every defect raised by review, security or audit, who raised it, and its status at closeout |
04-closeout.md |
The final report and the accepted deliverables |
05-positive-memory.md |
Lessons validated at closeout, where present |
Redaction: internal paths, identifiers, hashes, transport and sentences about the system's own mechanics are removed and logged; nothing else is rewritten. The models' prompts and the mechanics of the system are not part of the records.
Index
sessions.jsonl has one row per session: session, date, title, summary, messages (role outputs), findings (raised and closed), models (families and versions that appeared), example (path of the application, if published), files.
| Session | Date | Findings | Application |
|---|---|---|---|
| ForgeDesk 2, makerspace management system | 30 Aug 2026 | 15 | examples/forgedesk-2 |
| ShiftBoard, workshop production planner | 5 Sep 2026 | 7 | examples/shiftboard |
| QueueLens, support backlog explorer | 6 Sep 2026 | 6 | examples/queuelens |
| API Contract Explorer | 7 Sep 2026 | 6 | — |
| ShipGate, release readiness | 7 Sep 2026 | 15 | — |
| Supplier price-list importer | 23 Sep 2026 | 7 | examples/pricelist-import |
| FX paper trading and options risk terminal | 24 Sep 2026 | 9 | examples/fx-paper-trading |
Licence
Records and this card: CC BY 4.0. Each application in examples/ carries its own MIT licence. Jaryn, Tandem and AWOS names and logos are the property of Jaryn, all rights reserved (see TRADEMARKS.md).
Citation
@misc{jaryn2026tandemsessions,
title = {Tandem session records},
author = {di Castri, Francesco},
year = {2026},
url = {https://huggingface.co/datasets/fdicastri/tandem-sessions}
}
Contact: tandem@jaryn.io
- Downloads last month
- 86