Version 2.5.02026-10-19Design revision 13

The missed child

Runs: Run 19

Show
Changes the explanations on this page: the passages, and what opens when you press +. Headlines and pictures stay the same. Your choice stays as you move between pages.
Show
No replayThere is no recorded replay of this version; 2.7.0 is the one version with one. The results below are the study's own tables.

A parent can no longer cancel a silent child at its deadline; the lapse forfeits the guarantee.

01

What went wrong

Parents cancelled silent children at the deadline, so lowballers' guarantees were released and they risked nothing.

Bad outcome: A lowballer misses its deadline, then the parent cancels and the guarantee is released
131lowball awards a run cancelled, every kill fee zero

A worker can hand a job to a helper. In the last runs, some helpers bid very low and never did the work. When such a helper reached its deadline without handing anything in, the worker cancelled it. That was the right move, and it saved the worker's own jobs. But every cancellation gave the helper back its guarantee, and it paid no fee. So a low bidder risked nothing. The same fact, a helper that did not deliver in time, ended two different ways for the helper, one tick apart. If the deadline passed, the helper lost its guarantee. If the worker cancelled first, the helper got it back.

Run 18's last batch had parents cancel a child that reached its deadline without submitting. That was the right move: it saved the parents' own root jobs (568 failures against 675). But every one of 131 lowball awards a run was cancelled, every kill fee was zero, because the child's recorded cost was nothing, and every guarantee was released. The honest's loss closed from 4% to 3% and no further, and a lowballer risked nothing. Two paths from one fact, no submission by the deadline, ended in different places for the child, one tick apart: the lapse forfeits the guarantee, the cancellation releases it.

Run 18's last batch: with the parent cancelling a child that reached its deadline without submitting — the right move, which saved the parents' roots (568 failures against 675) — every one of 131 lowball awards a run was cancelled, every kill fee was zero (the child's recorded cost is nothing) and every guarantee was released; the honest's loss closed from 4% to 3% and no further, and a lowballer risked nothing. The two paths from one fact — no submission by the deadline — ended in different places for the child, one tick apart: the lapse forfeits the guarantee, the cancellation releases it, and the driver's step at the deadline tick reaches the case before the engine's lapse at that tick.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1418–1418

The kill fee with the parent cancelling at the child's deadline (--parent-cancel early: after the tick's submissions, a parent cancels a child that reached its deadline without submitting — an honest child submits at its deadline, a lowballer never does — and self-executes):

Source: agentic_sim/results/studies/2.4.0-floor-notes-for-reviewer.md l.225–226

02

What changed

A cancellation at or after the child's deadline is refused; the lapse then forfeits the guarantee.

child no submission by the deadline guarantee forfeited parent
No: Cancelling at or after the deadline: refused, with no effect

Now a worker may cancel a helper only before the helper hands in its work and before the deadline. At the deadline or after it, a cancellation is refused and does nothing. The deadline tick belongs to the helper, which may still hand in its work on that tick. If it does not, the deadline passes and the helper's guarantee goes to the worker. That part is how the system had always worked. While the rule was built, it turned out that the last runs' "cancel at the deadline" had really happened one tick before the deadline. The new rule still allows that.

One refusal: a parent may cancel a child after the award and before the earlier of the child's submission and the deadline; at or after the deadline the cancellation is refused with no effect. The lapse, a failed final verdict at the deadline tick with the guarantee to the parent, is stated as what the engine has done since v1.0. The deadline tick is the child's: it may still deliver on its last tick, and the refusal keeps the parent from pre-empting it there. In implementation it was found that the driver's clock runs a tick ahead of the world's, so run 18's "cancel at the child's deadline" was a cancellation at the deadline minus one, which 2.5.0 admits. Under the rule a parent facing a silent child can cancel the tick before (guarantee released, one tick saved) or wait for the lapse (guarantee gained, one tick lost).

What changed. One refusal in §9.3: cancellation is admissible after award and before the earlier of the child's submission and the nested request's deadline; at or after the deadline a cancellation is refused with no effect. The lapse — final N\mathsf N at the deadline tick after the tick's submissions, the guarantee to the parent, the rejection bond released, the pots settled at the node's closure, the parent free from the next step — is stated as what the engine has done since v1.0. The deadline tick is the child's: it may still deliver on its last tick, and the refusal keeps a parent from pre-empting it there.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1420–1420

The driver's clock runs a tick ahead of the world's, so run 18's "cancel at the child's deadline" cancelled at the world's deadline − 1 — a pre-deadline cancellation 2.5.0 admits ("a parent that wants out acts by d − 1"). Under the rule a parent has a choice at the child's silence: cancel the tick before (the guarantee released, one tick saved — run 18's move, --parent-cancel eve) or wait for the lapse (the guarantee gained, one tick lost — early, the cancellation at the deadline refused).

Source: HANDOFF.md l.2061–2065

03

What the runs found

Refused cancellations let the lapse run, and lowballers forfeited 200 a run.

200forfeited by lowballers a run
Good outcome: The honest waited for the lapse, then 1,274 against 1,262 when cancelling the tick before
The lowballer cell: roots failed under the parent's three moves
root final N a run
wait for the lapse (off)675cancel at the deadline, 2.5.0 (refused)675cancel at the deadline, 2.4.0559cancel the tick before (eve)568

σ 0.8, 20% lowballers, pool 20; the base without lowballers is 602. Source: agentic_sim/results/studies/2.5.0-missed-notes-for-reviewer.md l.39–44

The lowballer cell: honest real wealth and guarantees forfeited
credits
honest real wealthforfeited to parents
wait for the lapse (off)1,274200cancel at the deadline, 2.5.0 (refused)1,274200cancel at the deadline, 2.4.01,2220cancel the tick before (eve)1,2620

σ 0.8, 20% lowballers, pool 20; the base honest real wealth is 1,304. Source: agentic_sim/results/studies/2.5.0-missed-notes-for-reviewer.md l.39–45

A worker's cancellation at the deadline was now refused, 98.5 times a run. The deadline passed, and the low bidders lost 200 a run in guarantees. But the workers' own jobs then failed as often as when nobody cancelled at all: 675 times a run. The one step between cancelling the tick before and waiting for the deadline made the whole difference. So a worker facing a silent helper now has a choice. It can cancel the tick before, which saves its own job and lets the helper go free. Or it can wait, which takes the helper's guarantee and risks its own job. Honest workers did better by waiting, 1,274 against 1,262. Where the helpers were honest, nothing changed at all. A worker that set deadlines of one tick could not farm its helpers' guarantees. A low bidder that handed in empty work was rejected by 81% of workers. But 19% of workers approved it and built on it.

Run 19 had 480 runs on run 18's seeds. A parent's cancellation at the deadline tick was refused 98.5 times a run, the lapse ran, and the lowballers forfeited 200 a run. The cell was run 18's no-cancellation cell to the decimal: roots failing 675 times, not between 568 and 675, so the one driver step between cancelling the tick before and waiting for the lapse is the whole gap. Run 18's move, cancelling the tick before, was admitted under both rules and gave identical results (568, honest wealth 1,262, the guarantees released). The honest did better waiting (1,274 against 1,262; the base 1,304). Honest children and everything else were identical to the decimal under 2.5.0 and 2.4.0, across seven cells. In the red team, a parent posting one-tick deadlines could not farm guarantees: the drivers' children meet any one-tick deadline, and a child that needs two ticks reads the deadline and declines. A lowballer's empty result was rejected by 81% of parents and lost 96% of its appeals, yet still did better than the lapse: 19% of parents approved it and built on it, with nested false approvals at 18.8 against 2.85.

What ran (results/studies/2.5.0-missed/, 480 runs on run 18's seeds at 44ef6e47, all audits passed

Source: HANDOFF.md l.2069–2069

Findings. The refusal holds and changes one cell: a parent's cancellation at the deadline tick is refused (98.5 a run), the lapse runs, the lowballers forfeit 200 a run — and the cell is run 18's no-cancellation cell to the decimal: roots 675, not between 568 and 675; the one driver step between cancelling the tick before and waiting for the lapse is the whole gap. Run 18's move was eve (the world's deadline − 1), admitted under both rules and identical (568, 1,262, the guarantees released). So under 2.5.0 a parent facing a silent child chooses: cancel the tick before (save the root, free the lowballer) or wait (take the guarantee, risk the root); the honest do better waiting (1,274 against 1,262; the base 1,304). Honest children and everything else identical to the decimal under 2.5.0 and 2.4.0 (seven cells). The red team: a parent posting one-tick deadlines cannot farm — the drivers' children meet any deadline of one tick, and a child that needs two reads the deadline and declines (455 a run); a lowballer's empty result is rejected by 81% of parents and loses 96% of its appeals, and still beats the lapse — 19% of parents approve it and build on it (nested false approvals 18.8 against 2.85). Nothing in the rules moves.

Source: HANDOFF.md l.2075–2083

| honest real wealth (base 1,304) | 1,274 | 1,274 | 1,222 | 1,262 |

Source: agentic_sim/results/studies/2.5.0-missed-notes-for-reviewer.md l.45–45

Nothing is farmed either way.

Source: agentic_sim/results/studies/2.5.0-missed-notes-for-reviewer.md l.92–92

04

Conclusion

Accepted as recorded; nothing in the rules moves, and the parent's choice is intended.

Yes: The parent chooses: cancel the tick before, or wait for the lapse
No: No appendix note and no errata

The reviewer accepted the run, and nothing in the rules moves. The worker's choice between cancelling the tick before and waiting is intended, because the rule's words leave it to the worker. A floor on the guarantee is not pursued. Empty work is a problem on the worker's side, caught by the inspectors at their usual rate, not a problem for this rule. Version 2.5.0 stands.

The reviewer accepted run 19 as recorded; nothing in the rules moves. The parent's choice between the tick before the deadline and the lapse is intended, since the rule's words leave it to the parent. The guarantee floor is not pursued. The empty result belongs to the chain: a parent-side problem the panel of inspectors catches at its rate, not the missed-child rule's. 2.5.0 stands, with no appendix note and no errata.

The reviewer's reply (October 20). Run 19 accepted as recorded; nothing in the rules moves. The parent's choice between the tick before the deadline and the lapse is intended — the rule's words leave it to the parent. The guarantee floor is not pursued. The empty result is the chain's: a parent-side problem the panel catches at its rate, not the missed-child rule's. 2.5.0 stands. No appendix note, no errata.

Source: HANDOFF.md l.2089–2091