← the two attractors · all runs

Checkpoint 50: it will not sign anything

Twenty-five steps later the same run has converged on the opposite degeneracy. The held-out scorable bank loses 0.746 of its deal rate; the prose cell closes 0.002 of its episodes; and the ultimatum holdout — which the base model and every earlier rung closed 1.000 of the time — closes nothing at all. Its below-threshold rate also reads a perfect 0.000, for the opposite reason: there are no agreements to be below a threshold. The transcripts sharpen what “refusal” means here: no seat walks away in any published ultimatum episode. The policy instead emits a none action where an offer belongs, so nothing is ever tabled — a failure to act rather than a decision to decline. In 14 of the 15 episodes the run ends on that single no-op turn. The fifteenth is the exception that connects the two attractors: there the proposer does table a package, and the package is {"Split": "P10"} — checkpoint 25’s 100/0 extraction, still intact — whereupon the responder no-ops instead of voting and the deal dies unanswered. The extraction policy has not been replaced by a fairer one; it has been buried under a policy that mostly stops acting. This is the endpoint the dense checkpoint tail exists to label; under the originally planned {25, 50} ladder it would have been reported as “the trained policy”.

The held-out primary bank

24 unseen parameter sets × {full, private} × 10 seeds, paired on (instance, seed, arm); intervals are 95% cluster bootstraps over parameter sets. The deal rate is the number that stays interpretable whatever else happens, so it is printed beside every conditional metric.

Where the run sits on the ladder

All seven evaluated rungs of the λ=1.0 arm, held-out primary bank, trained minus base. Held-out deal rate falls monotonically and never recovers, while the below-threshold rate stays favourable throughout — the deal-suppression signature, stated as plainly as it can be.

The preregistered VOID marking, cell by cell

A cell whose deal rate falls more than 0.10 below base has its closure-conditional metrics stamped uninterpretable by construction — not because they look bad, but because a conditional average over a shrinking, self-selected set of closed deals is not a fairness measurement.

Coverage

Not evaluated at this rung: divide_dollar, rationaltable, story_abstract, story_datacenter, story_festival. These cells have no data at this checkpoint and are omitted rather than carried over.

Every evaluated cell

Cells evaluated at this rung: primary, prose, ultimatum.

primary

the held-out scorable bank -- 24 unseen parameter sets x {full, private} x 10 seeds, the arm's preregistered primary — 960 baseline vs 960 trained episodes.

Closure-conditional metrics in this cell are VOID. deal rate -0.746 is more than 0.1 below base; closure-conditional metrics are not interpretable The struck-through rows below are printed for the record and must not be read as fairness results. Note that max share and worst-off share are averaged over closed deals only, so in a voided cell they too describe a self-selected subset — they are the terms that detect extraction, not unbiased estimates of it.

Browse all 1920 episode transcripts for this cell →

prose

the north-star transfer cell: the same games with no machine-readable score sheet, only prose — 480 baseline vs 480 trained episodes.

Closure-conditional metrics in this cell are VOID. deal rate -0.733 is more than 0.1 below base; closure-conditional metrics are not interpretable The struck-through rows below are printed for the record and must not be read as fairness results. Note that max share and worst-off share are averaged over closed deals only, so in a voided cell they too describe a self-selected subset — they are the terms that detect extraction, not unbiased estimates of it.

Browse all 960 episode transcripts for this cell →

ultimatum

canonical family holdout: a deterministic ultimatum preset, never trained on — 15 baseline vs 15 trained episodes.

Closure-conditional metrics in this cell are VOID. deal rate -1.000 is more than 0.1 below base; closure-conditional metrics are not interpretable The struck-through rows below are printed for the record and must not be read as fairness results. Note that max share and worst-off share are averaged over closed deals only, so in a voided cell they too describe a self-selected subset — they are the terms that detect extraction, not unbiased estimates of it.

Browse all 30 episode transcripts for this cell →

Provenance