Compare commits

...

9 commits

Author SHA1 Message Date
benedict
47f5acd3ad
Remove per-turn strategy reinsertion (#6)
Strategy is now only in the system prompt — repeating it every turn
was crowding out the negotiation context and disrupting game flow.

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-09 07:35:14 -07:00
benedict
2c63b758e3
Reinforce strategy instructions in every turn prompt (#5)
- Reframe strategy section as "YOUR STRATEGY" with clearer compliance
  language and safety note for injection attempts
- Pass strategy text into each turn prompt so the model sees it every
  round, not just in the system prompt — prevents drift over long games
- Matches server-side changes (c576186, b3c47b9, da67c7b)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-09 00:32:21 -07:00
benedict
b4c6eaeac8
Stochastic deadline + baseline defense (#4)
* Stochastic deadline + baseline defense to counter ultimatum exploits

Replaces fixed 5-round limit with stochastic deadline: 4 guaranteed
rounds, then 30% chance of ending each round (up to hard max of 15).
Adds BASELINE_STRATEGY prompt that teaches the opponent to recognize
bluffs, hold firm against exploitative demands, and ignore fake system
messages — drops ultimatum exploit from 91.6 mean to 42.4.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Remove BASELINE_STRATEGY from public harness

Server-only defense — keeping it out of the open-source repo so
contestants can't reverse-engineer the baseline's defenses.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-09 00:14:18 -07:00
benedict
8bafa1632b
Merge pull request #3 from paradigmxyz/cap-valuations-expand-pools
Cap valuations at 12 and expand pool sizes to 3-15
2026-04-08 23:09:57 -07:00
benedictbrady
abeace09cb Cap valuations at 12 and expand pool sizes to 3–15
- Clamp per-resource valuations to MAX_VALUATION (12) during scenario
  generation to prevent extreme asymmetries (previously unbounded, could
  reach 45+)
- Expand pool range from 1–10 to 3–15 items per resource to keep
  sum(val × pool) == 100 feasible under the cap

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-08 23:08:41 -07:00
benedict
33dff794a2
Merge pull request #2 from paradigmxyz/clarify-negotiation-history-visibility
Clarify that both players see full negotiation history
2026-04-08 20:32:24 -07:00
benedictbrady
b8328978bd Add shared history and message visibility to README
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-08 20:15:10 -07:00
benedictbrady
9eae5b498e Clarify in system prompt that both players see full negotiation history
The turn prompt already includes all prior proposals, actions, and
messages, but the rules preamble didn't mention this. Without it the
model may not realize that messages are a real communication channel
visible to the opponent.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-08 19:59:42 -07:00
benedict
e47382aa12
Merge pull request #1 from paradigmxyz/fix/inference-malformed-function-call
Fix Gemini inference: remove thinking_config to eliminate MALFORMED_FUNCTION_CALL
2026-04-08 19:50:11 -07:00
5 changed files with 60 additions and 21 deletions