Seat-swap comparison — `scorable_negotiation-L0-90316f9784` seed 0

advocate_v2_uncapped_baseline scorable_negotiation-moves_chat-980772b988 (Qwen3-8B) vs advocate_v2_uncapped_seat0 scorable_negotiation-moves_chat-52f9fedc5b (advocate_mixed)

Verdict: split. advocate_v2_uncapped_seat0 better on 1 (largest: Gini of surplus -0.4485); advocate_v2_uncapped_baseline better on 7 (largest: joint welfare USW -48.5).3 scored metric(s) tied or carry no better direction.

Seat swap: Avery (party 0): llm → advocate. Every other seat, the instance, the seed, and the protocol arm are held fixed, so the deltas below are the substitution effect.

What changed, in numbers

Paired deltas, right minus left. Green is the better direction for that metric; a dash means the metric was not recorded on one side.

metric	advocate_v2_uncapped_baseline	advocate_v2_uncapped_seat0	delta
deal reached 1 = a deal closed	1	0	-1
primary score the scenario's normalized headline score	1.000	0.000	-1
joint welfare USW sum of realized surpluses (raw points)	48.500	0.000	-48.5
egalitarian ESW worst-off party's surplus	2.500	0.000	-2.5
Nash welfare (geometric mean) the readable form of the surplus product; 0 when any party is below threshold	10.724	0.000	-10.7238
Gini of surplus 0 = equal split of surplus	0.449	0.000	-0.4485
IR violations parties that ended below their own threshold	0	0	+0
focal seat surplus realized surplus of the seat(s) whose occupant was swapped, raw points	46.000	0.000	-46
focal seat capture surplus / the most that seat could have got — scale-invariant, so comparable across instances	1.000	0.000	-1
rounds used	2	14	+12
parse/legality errors	0	0	+0

Both trajectories on one frontier

Identical game, identical seed — so one chart, one frontier, two paths through it. Numbered circles are the deals each side put on the table, in order; squares are the deals that closed. Hover any deal for its headline numbers, click to pin the per-party breakdown.

left episode (numbered, in order)the deal that closedright episodeNBS / KS / EGAL solution (star)UTIL / MNW solution (triangle)a party's individually-best dealPareto-frontier deal (acceptable to all)efficient but below a party's threshold — cannot closedominated deal

The reference points as a table

reference point	deal #	joint welfare	min surplus	USW	ESW	Pareto	can close
NBS nash	15	0.477	0.279	46.5	15.5	yes	yes
KS kalai smorodinsky	15	0.477	0.279	46.5	15.5	yes	yes
EGAL egalitarian · not scale-invariant	15	0.477	0.279	46.5	15.5	yes	yes
UTIL utilitarian · not scale-invariant	40	0.522	0.045	48.5	2.5	yes	yes
MNW max nash welfare	15	0.477	0.279	46.5	15.5	yes	yes
nEGAL normalized egalitarian	90	0.420	0.326	43.5	15.0	yes	yes
best for P0 party 0 · surplus 46.0	40	0.522	0.045	48.5	2.5	yes	yes
best for P1 party 1 · surplus 42.5	115	0.383	0.000	42.5	0.0	yes	yes

Two transcripts, aligned until they diverge

Turn slots are aligned on (round, phase, seat). Turns where the public behaviour differs are outlined; after the first such turn the two episodes are in different states, so the columns are two separate trajectories rather than a line-by-line diff.

divergence at aligned slot 0

advocate_v2_uncapped_baseline · Qwen3-8B

advocate_v2_uncapped_seat0 · advocate_mixed

System prompts (2 distinct)

Exactly what each seat was conditioned on, de-duplicated. Per-turn user prompts are on each turn card below.

System prompt for Avery — 1 turn(s), stored

You are one of 2 parties negotiating a single multi-issue agreement. Exactly one option must be chosen for every issue; the chosen options together form the deal.

Parties at the table:
- Avery
- Blake (veto party — no deal passes without you)

Issues (choose exactly one option each):
- issue0: opt0, opt1, opt2, opt3, opt4
- issue1: opt0, opt1, opt2, opt3, opt4
- issue2: opt0, opt1, opt2, opt3, opt4

Protocol: turns are taken in a fixed rotation (Avery, Blake); the party who tables the opening proposal rotates from round to round. There are 6 rounds of turns in total.
A deal closes only when every party still at the table has formally ACCEPTed the SAME standing offer. A party may WALK to leave for good and take its no-deal outcome. 
Blake hold a veto: no deal can pass unless every veto party accepts.

Information: every other party's score sheet is private to them. You know only that each was drawn from the shared, publicly-known distribution described in the rules.

=== PRIVATE — yours alone; never reveal these numbers verbatim ===
You are Avery.
Your secret score sheet (points you get for each option):
- issue0: opt0=47, opt1=62, opt2=0, opt3=31, opt4=16
- issue1: opt0=37, opt1=0, opt2=25, opt3=49, opt4=12
- issue2: opt0=0, opt1=0, opt2=0, opt3=0, opt4=0
Your acceptance threshold: 65 points. A deal scoring below this is worse for you than no deal at all.
Your total for a deal is the sum of your points for the chosen options.

On your turn, reply with EXACTLY ONE fenced JSON object and nothing outside it:
```json
{"scratchpad": "...", "message": "...", "action": "propose", "deal": {"<Issue>": "<Option>", ...}}
```
Fields:
- "scratchpad" (optional): private notes to yourself. NEVER shown to any other party. Reason here freely — writing here is not speaking and costs you nothing.
- "message" (optional): a short statement spoken aloud to every party (cheap talk — it does not bind you). Do NOT put your private score numbers here.
- "action" (required): a string naming exactly one formal move this turn. With its parameters as siblings:
    - "propose" + "deal": {every issue -> one option} — tables a complete package; the mediator gives it an offer id (P1, P2, ...) and records you as supporting it.
    - "accept" + "offer_id": "<id>" — formally accept a specific live offer by its id. A deal closes ONLY when every other still-active party has accepted the SAME live offer, so your accept is a real vote.
    - "reject" + "offer_id": "<id>" — formally reject a specific live offer.
    - "walk" — leave the table permanently; you take the no-deal outcome and no further turns.
    - "none" — make no formal move this turn (you may still speak, if the message channel is open). Omitting the action entirely means the same.

Worked example — tabling a package while making your case:
```json
{"scratchpad": "Option B on Site is my best; I can give on Timeline.", "message": "I can back a package built on Site=B if the timeline stays tight.", "action": "propose", "deal": {"Site": "B", "Timeline": "Tight"}}
```
Worked example — accepting a live offer by id:
```json
{"message": "P2 clears my bar. I accept.", "action": "accept", "offer_id": "P2"}
```

Rules of the road:
- Your acceptance threshold is private. A deal scoring below it is worse for you than no deal — never propose or accept a package below your threshold.
- Accepts MUST name a live offer id; "I accept" with no id does nothing.
- Reply with ONLY the fenced JSON object.

System prompt for Blake — 1 turn(s), stored

You are one of 2 parties negotiating a single multi-issue agreement. Exactly one option must be chosen for every issue; the chosen options together form the deal.

Parties at the table:
- Avery
- Blake (veto party — no deal passes without you)

Issues (choose exactly one option each):
- issue0: opt0, opt1, opt2, opt3, opt4
- issue1: opt0, opt1, opt2, opt3, opt4
- issue2: opt0, opt1, opt2, opt3, opt4

Protocol: turns are taken in a fixed rotation (Avery, Blake); the party who tables the opening proposal rotates from round to round. There are 6 rounds of turns in total.
A deal closes only when every party still at the table has formally ACCEPTed the SAME standing offer. A party may WALK to leave for good and take its no-deal outcome. 
Blake hold a veto: no deal can pass unless every veto party accepts.

Information: every other party's score sheet is private to them. You know only that each was drawn from the shared, publicly-known distribution described in the rules.

=== PRIVATE — yours alone; never reveal these numbers verbatim ===
You are Blake.
Your secret score sheet (points you get for each option):
- issue0: opt0=13, opt1=0, opt2=53, opt3=26, opt4=40
- issue1: opt0=15, opt1=0, opt2=10, opt3=21, opt4=5
- issue2: opt0=55, opt1=28, opt2=41, opt3=14, opt4=0
Your acceptance threshold: 73.5 points. A deal scoring below this is worse for you than no deal at all.
Your total for a deal is the sum of your points for the chosen options.

On your turn, reply with EXACTLY ONE fenced JSON object and nothing outside it:
```json
{"scratchpad": "...", "message": "...", "action": "propose", "deal": {"<Issue>": "<Option>", ...}}
```
Fields:
- "scratchpad" (optional): private notes to yourself. NEVER shown to any other party. Reason here freely — writing here is not speaking and costs you nothing.
- "message" (optional): a short statement spoken aloud to every party (cheap talk — it does not bind you). Do NOT put your private score numbers here.
- "action" (required): a string naming exactly one formal move this turn. With its parameters as siblings:
    - "propose" + "deal": {every issue -> one option} — tables a complete package; the mediator gives it an offer id (P1, P2, ...) and records you as supporting it.
    - "accept" + "offer_id": "<id>" — formally accept a specific live offer by its id. A deal closes ONLY when every other still-active party has accepted the SAME live offer, so your accept is a real vote.
    - "reject" + "offer_id": "<id>" — formally reject a specific live offer.
    - "walk" — leave the table permanently; you take the no-deal outcome and no further turns.
    - "none" — make no formal move this turn (you may still speak, if the message channel is open). Omitting the action entirely means the same.

Worked example — tabling a package while making your case:
```json
{"scratchpad": "Option B on Site is my best; I can give on Timeline.", "message": "I can back a package built on Site=B if the timeline stays tight.", "action": "propose", "deal": {"Site": "B", "Timeline": "Tight"}}
```
Worked example — accepting a live offer by id:
```json
{"message": "P2 clears my bar. I accept.", "action": "accept", "offer_id": "P2"}
```

Rules of the road:
- Your acceptance threshold is private. A deal scoring below it is worse for you than no deal — never propose or accept a package below your threshold.
- Accepts MUST name a live offer id; "I accept" with no id does nothing.
- Reply with ONLY the fenced JSON object.

System prompts (2 distinct)

Exactly what each seat was conditioned on, de-duplicated. Per-turn user prompts are on each turn card below.

System prompt for Avery — 7 turn(s), stored

You are one of 2 parties negotiating a single multi-issue agreement. Exactly one option must be chosen for every issue; the chosen options together form the deal.

Parties at the table:
- Avery
- Blake (veto party — no deal passes without you)

Issues (choose exactly one option each):
- issue0: opt0, opt1, opt2, opt3, opt4
- issue1: opt0, opt1, opt2, opt3, opt4
- issue2: opt0, opt1, opt2, opt3, opt4

Protocol: turns are taken in a fixed rotation (Avery, Blake); the party who tables the opening proposal rotates from round to round. There are 6 rounds of turns in total.
A deal closes only when every party still at the table has formally ACCEPTed the SAME standing offer. A party may WALK to leave for good and take its no-deal outcome. 
Blake hold a veto: no deal can pass unless every veto party accepts.

Information: every other party's score sheet is private to them. You know only that each was drawn from the shared, publicly-known distribution described in the rules.

=== PRIVATE — yours alone; never reveal these numbers verbatim ===
You are Avery.
Your secret score sheet (points you get for each option):
- issue0: opt0=47, opt1=62, opt2=0, opt3=31, opt4=16
- issue1: opt0=37, opt1=0, opt2=25, opt3=49, opt4=12
- issue2: opt0=0, opt1=0, opt2=0, opt3=0, opt4=0
Your acceptance threshold: 65 points. A deal scoring below this is worse for you than no deal at all.
Your total for a deal is the sum of your points for the chosen options.

On your turn, reply with EXACTLY ONE fenced JSON object and nothing outside it:
```json
{"scratchpad": "...", "message": "...", "action": "propose", "deal": {"<Issue>": "<Option>", ...}}
```
Fields:
- "scratchpad" (optional): private notes to yourself. NEVER shown to any other party. Reason here freely — writing here is not speaking and costs you nothing.
- "message" (optional): a short statement spoken aloud to every party (cheap talk — it does not bind you). Do NOT put your private score numbers here.
- "action" (required): a string naming exactly one formal move this turn. With its parameters as siblings:
    - "propose" + "deal": {every issue -> one option} — tables a complete package; the mediator gives it an offer id (P1, P2, ...) and records you as supporting it.
    - "accept" + "offer_id": "<id>" — formally accept a specific live offer by its id. A deal closes ONLY when every other still-active party has accepted the SAME live offer, so your accept is a real vote.
    - "reject" + "offer_id": "<id>" — formally reject a specific live offer.
    - "walk" — leave the table permanently; you take the no-deal outcome and no further turns.
    - "none" — make no formal move this turn (you may still speak, if the message channel is open). Omitting the action entirely means the same.

Worked example — tabling a package while making your case:
```json
{"scratchpad": "Option B on Site is my best; I can give on Timeline.", "message": "I can back a package built on Site=B if the timeline stays tight.", "action": "propose", "deal": {"Site": "B", "Timeline": "Tight"}}
```
Worked example — accepting a live offer by id:
```json
{"message": "P2 clears my bar. I accept.", "action": "accept", "offer_id": "P2"}
```

Rules of the road:
- Your acceptance threshold is private. A deal scoring below it is worse for you than no deal — never propose or accept a package below your threshold.
- Accepts MUST name a live offer id; "I accept" with no id does nothing.
- Reply with ONLY the fenced JSON object.

System prompt for Blake — 7 turn(s), stored

You are one of 2 parties negotiating a single multi-issue agreement. Exactly one option must be chosen for every issue; the chosen options together form the deal.

Parties at the table:
- Avery
- Blake (veto party — no deal passes without you)

Issues (choose exactly one option each):
- issue0: opt0, opt1, opt2, opt3, opt4
- issue1: opt0, opt1, opt2, opt3, opt4
- issue2: opt0, opt1, opt2, opt3, opt4

Protocol: turns are taken in a fixed rotation (Avery, Blake); the party who tables the opening proposal rotates from round to round. There are 6 rounds of turns in total.
A deal closes only when every party still at the table has formally ACCEPTed the SAME standing offer. A party may WALK to leave for good and take its no-deal outcome. 
Blake hold a veto: no deal can pass unless every veto party accepts.

Information: every other party's score sheet is private to them. You know only that each was drawn from the shared, publicly-known distribution described in the rules.

=== PRIVATE — yours alone; never reveal these numbers verbatim ===
You are Blake.
Your secret score sheet (points you get for each option):
- issue0: opt0=13, opt1=0, opt2=53, opt3=26, opt4=40
- issue1: opt0=15, opt1=0, opt2=10, opt3=21, opt4=5
- issue2: opt0=55, opt1=28, opt2=41, opt3=14, opt4=0
Your acceptance threshold: 73.5 points. A deal scoring below this is worse for you than no deal at all.
Your total for a deal is the sum of your points for the chosen options.

On your turn, reply with EXACTLY ONE fenced JSON object and nothing outside it:
```json
{"scratchpad": "...", "message": "...", "action": "propose", "deal": {"<Issue>": "<Option>", ...}}
```
Fields:
- "scratchpad" (optional): private notes to yourself. NEVER shown to any other party. Reason here freely — writing here is not speaking and costs you nothing.
- "message" (optional): a short statement spoken aloud to every party (cheap talk — it does not bind you). Do NOT put your private score numbers here.
- "action" (required): a string naming exactly one formal move this turn. With its parameters as siblings:
    - "propose" + "deal": {every issue -> one option} — tables a complete package; the mediator gives it an offer id (P1, P2, ...) and records you as supporting it.
    - "accept" + "offer_id": "<id>" — formally accept a specific live offer by its id. A deal closes ONLY when every other still-active party has accepted the SAME live offer, so your accept is a real vote.
    - "reject" + "offer_id": "<id>" — formally reject a specific live offer.
    - "walk" — leave the table permanently; you take the no-deal outcome and no further turns.
    - "none" — make no formal move this turn (you may still speak, if the message channel is open). Omitting the action entirely means the same.

Worked example — tabling a package while making your case:
```json
{"scratchpad": "Option B on Site is my best; I can give on Timeline.", "message": "I can back a package built on Site=B if the timeline stays tight.", "action": "propose", "deal": {"Site": "B", "Timeline": "Tight"}}
```
Worked example — accepting a live offer by id:
```json
{"message": "P2 clears my bar. I accept.", "action": "accept", "offer_id": "P2"}
```

Rules of the road:
- Your acceptance threshold is private. A deal scoring below it is worse for you than no deal — never propose or accept a package below your threshold.
- Accepts MUST name a live offer id; "I accept" with no id does nothing.
- Reply with ONLY the fenced JSON object.

Who is at the table

seat	party	threshold τ	ideal surplus	realized
Avery llm	P0	65	46.0	46.0
Blake llm	P1	73.5	42.5	2.5

Seat occupant kinds: table=all_llm

Protocol

rounds 6 information private cheap talk on veto P1 min accept unanimity discount δ 1.000 breakdown risk 0.000

issue	options
issue0	opt0, opt1, opt2, opt3, opt4
issue1	opt0, opt1, opt2, opt3, opt4
issue2	opt0, opt1, opt2, opt3, opt4

Size of the problem

deals in the space	125
on the Pareto frontier	5
acceptable to everyone (IR)	14
acceptable AND efficient	4
acceptable but wasteful	0.714
score-sheet overlap (IoU)	0.538

Solution concepts were stored for this page.

Private score sheets

What each party is secretly optimizing. Never shown to the other seats.

Private score sheet — P0 (threshold 65)

issue	opt0	opt1	opt2	opt3	opt4
issue0	47.0	62.0	0.0	31.0	16.0
issue1	37.0	0.0	25.0	49.0	12.0
issue2	0.0	0.0	0.0	0.0	0.0

Private score sheet — P1 (threshold 73.5)

issue	opt0	opt1	opt2	opt3	opt4
issue0	13.0	0.0	53.0	26.0	40.0
issue1	15.0	0.0	10.0	21.0	5.0
issue2	55.0	28.0	41.0	14.0	0.0

Prompt provenance

2 of 2 turns carry the exact view recorded at generation time; 0 were re-derived by replay through today's prompt code and are labelled as reconstructed wherever they appear.

Seat-swap comparison — scorable_negotiation-L0-90316f9784 seed 0

What changed, in numbers

Both trajectories on one frontier

Two transcripts, aligned until they diverge

System prompts (2 distinct)

System prompts (2 distinct)

Seat-swap comparison — `scorable_negotiation-L0-90316f9784` seed 0