Upload 2-README.md
Browse files- 2-README.md +40 -0
2-README.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
pretty_name: Observatory Lab - first batch experiment evidence
|
| 4 |
+
tags:
|
| 5 |
+
- multi-agent-systems
|
| 6 |
+
- llm-agents
|
| 7 |
+
- coordination
|
| 8 |
+
- error-propagation
|
| 9 |
+
- evaluation
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# Observatory Lab - experiment evidence (first batch, 2026-10-27)
|
| 13 |
+
|
| 14 |
+
Raw evidence for the Observatory Lab first experiment batch: per-call
|
| 15 |
+
transcripts (verbatim model responses), run logs, and aggregates for five
|
| 16 |
+
pilot experiments on multi-agent coordination, consensus, error propagation,
|
| 17 |
+
instruction conflict, and supervision cadence.
|
| 18 |
+
|
| 19 |
+
- Model under study: gemini-3.1-flash-lite (Google AI Studio free tier).
|
| 20 |
+
- Evidence class: exploratory (n=3 seeds per condition; B01: 1 seed x 8
|
| 21 |
+
items). Not a benchmark; no statistical or generalization claims.
|
| 22 |
+
- Code, methodology and report: https://github.com/olserra/observatory-lab
|
| 23 |
+
(report file reports/first-batch-report.md is mirrored here).
|
| 24 |
+
- Landing: https://observatory-lab.vercel.app/
|
| 25 |
+
- Evaluators are deterministic and included in the repo; every number in the
|
| 26 |
+
report traces to the JSONL files here.
|
| 27 |
+
|
| 28 |
+
## Contents
|
| 29 |
+
- data/exp-*/transcripts/*.jsonl - one record per logical model call.
|
| 30 |
+
- data/exp-*/summary*.json - per-condition aggregates.
|
| 31 |
+
- reports/first-batch-report.md - exploratory report (final).
|
| 32 |
+
|
| 33 |
+
## Known limitations (read before reusing)
|
| 34 |
+
- First pass of A01/C01/D01 persisted aggregates only (labeled preliminary);
|
| 35 |
+
rerun 1 and later runs persist full transcripts.
|
| 36 |
+
- Prompts in pre-2026-10-27 records are stored hashed, not verbatim;
|
| 37 |
+
responses are verbatim. Later runs log prompts verbatim.
|
| 38 |
+
- B01's mediator arm is incomplete (daily quota); completed-item accuracy is
|
| 39 |
+
a descriptive subset.
|
| 40 |
+
- All tasks are synthetic with fictional agents; no personal data.
|