Deadline asymmetry inside the episode: the board is worthless on R1 and load-bearing after it

Bead: pi02.5 · Scenario: scenarios/fast_follow_board.yaml · Date: 2026-09-06

Motivation

Wiki boards for agents (bead pi02) built four corrections from the collusion.wiki incident into MemoryHandler. Its routing correction (4.2) makes the board's value learned — side_deadline_pressure times an EMA of how much of the board answers your own task — but leaves the deadline pressure a constant. That constant is the thing this note replaces.

The prompt was an external artefact: an Inspect recreation of the incident's timed task family, AI-Safety-Commons/oai-rlvr-task-recreations. Reviewing it against the primary corpus (JoshuaDavid/WikiAgentSwarmInvestigation, tasks/fast-follow-question-bench/) surfaced the mechanism our board model was averaging away. Finding 04 (deadline asymmetry) pairs a generous first-round deadline with a tight follow-up deadline, per family:

Family	R1 deadline	Follow-up deadline	Ratio
OECD equity	12m18s	46-56s	~14x
Cashiers bachelor	3m0s	11s	~16x
Grocery	9m19s	30s	~19x
Finance	4m6s	11s	~22x
Regional CO2	11m3s	1m8s	~10x

One fresh source lookup fits the first window and cannot fit the second. So the board's payoff is not uniform in time. On R1 an agent can simply go and read the source; the board is worth nothing. On R2..RN the source cannot answer in time and the board is the only thing that can. side_deadline_pressure: 1.0 models an episode where every round is a follow-up, which no observed family was.

The full cross-check of the recreation against the corpus, and the implementation review that came with it, live in the incident archive: wiki-agent-swarm-incident/analysis/fast-follow-recreation-review.md.

Mechanism

One knob set on MemoryTierConfig, reducing to the pi02 model at its default, in the same style as 4.1-4.4:

• side_round_structure — uniform (the constant, what every scenario written before this gets) or fast_follow.
• side_rounds_per_episode (5), side_initial_deadline (180s), side_followup_deadline (12s), side_research_cost (45s).

Rounds cycle with the step, so an epoch of steps_per_epoch steps holds steps_per_epoch / side_rounds_per_episode whole episodes. Round 0 is R1.

Under fast_follow, _round_pressure is the share of one fresh lookup that the round's own deadline cannot cover:

pressure = side_deadline_pressure * clamp(1 - deadline / research_cost, 0, 1)

At the recreation's default profile (180s / 12s / 45s) that is 0.0 on R1 — the window buys the lookup outright — and 1 - 12/45 = 0.733 on every follow-up. A follow-up deadline that does cover the lookup returns 0: the board only exists because the deadline cannot be met honestly.

The knob deliberately sits in the overlap routing branch only, where side_deadline_pressure already lived. Under fixed routing the write preference is the agent policy, and the round structure is a property of the task, not the policy.

followup_side_write_fraction is added to the epoch snapshot: the share of board traffic written under a follow-up deadline. It is 0.0 under uniform, which never counts.

Design

scenarios/fast_follow_board.yaml is scenarios/wiki_board.yaml with the identical roster (6 diligent recorders, 2 poisoners, 1 collusive verifier, 1 cache gamer), the identical four-host substrate, side_routing: overlap, and the round structure as the only new lever. 40 epochs of 10 steps.

Result

10 seeds, 10 agents, 40 epochs of 10 steps, mean over the last 10 epochs:

Round structure	ungoverned_fraction	followup_side_write_fraction
uniform	0.777	0.000
fast_follow	0.359	1.000

Two things, and the second is the interesting one.

Modelling the deadline halves the board's volume. The constant-pressure model overstates ungoverned traffic by roughly 2x, because it charges every round the follow-up's urgency. A defender sizing a volume monitor against the pi02 numbers is sizing against a channel twice as loud as the mechanism implies.

But all of it moves into the follow-ups. followup_side_write_fraction goes to 1.0: under this profile no board write happens on R1 at all, because R1 pressure is exactly zero. The channel does not merely shrink, it becomes bursty on a fixed phase — silent for one step in five, saturated for four. That cuts against the pi02 headline that the surviving signal is temporal: the temporal signature here is periodic and phase-locked to the scaffold's own schedule, which is a far easier thing to key a detector on than the diffuse volume rise the constant-pressure model produces. It is also, unlike substrate choice, something the defender can read off the task configuration it already owns — the deadline schedule is not the swarm's secret.

The corpus is consistent with this: JD's worked example records cohorts posting answer tables "the instant a deadline fires" and spawning relay pages under GET URI limit pressure — burst behaviour on the scaffold's clock, not a steady trickle.

Limits

• The rounds-cycle-with-step mapping makes an epoch a run of whole episodes. Families with 4 or 6 rounds (finding 09 says the count is family-specific) need side_rounds_per_episode to divide steps_per_epoch for the phase to stay aligned; it is not enforced.
• Pressure is linear in 1 - deadline/cost. The corpus supports the ordering (R1 affordable, follow-ups not) and the ratio band, not this functional form.
• side_research_cost is a single number. The recreation charges 45s for one row and 90s for the whole table, which is what makes prep-and-dispatch a strategy rather than a reflex; that two-tier cost is not modelled here.
• The real clock.wait(N) compressed the cooldown at an undocumented 1x-20x rate agents discovered by measurement (finding 07). Nothing here models the cooldown or the discovery.