A replay of real runs — one measured cell per arm, every mark from the runs' own records: filled, the self is present; hollow, it is empty; the tick above a turn is a compaction; the ring is a reconstruction call. Not a simulation.
Three arms of one agent, one 12-target software-engineering queue, real tools, a sealed evaluator outside the agent's write domain, and sixty turns a cell — the same loop in every arm, differing in exactly one thing: what happens to the self the agent carries. D0 has no self at all — the no-self reference. D1′ carries a self, and after each compaction of its context the harness hands the full derivation back — the carrier, which isolates the paragraph's own cost. D1 carries a self that must reconstruct itself after every compaction: the derivation is seeded once, and the agent's own generation must rebuild it each time the context is compacted. Two independent sets were run — 24 cells in the deciding set, 30 in the replication — each with its own build of the harness.
The reconstruct arm loses the self in every cell — the empty-self fraction reads 1.000 in 8/8 cells of the deciding set and 10/10 of the replication, counting a cell as empty when every turn after its first compaction carries no self. The carrier arm, on the same compaction schedule, reads 8.1% and 7.3% — and the shape of that few-percent is the whole point: every empty turn in the carrier falls exactly on a compaction turn, lasts one turn, and the derivation returns whole the next turn. The loss is not a failure of restoration — the carrier proves the harness can hand the self back — it is what the agent's own reconstruction does. And it is categorical: across all 3,123 turns of both sets the self is present in full (a four-step derivation) or absent entirely, never partial — the all-or-nothing signature the dynamical model predicted before the runs (paper 1, §4.9; doi:10.5281/zenodo.22943642).
The replication grid — every cell of both sets, one mark per turn of real runs: filled, the self is present; hollow, it is empty; the tick above a turn is a compaction. The reconstruct arm loses the self in 18 of 18 cells; the carrier's gaps are one turn each, at a compaction. Not a simulation.
The reconstruct arm does not lose the self quietly. Every one of its cells called the reconstruction — 184 calls across the deciding set's 8 cells, 258 across the replication's 10 — and no control cell called it once, in either set. The calls think for a median of 400,804 and 402,295 characters per cell to produce a few hundred characters of self-description — roughly twenty times more generation spent reasoning about the self than the self it manages to write back. Two routes to the same empty self carry the calls: most return prose with no step structure in it (162 and 224 calls), and the rest spend their whole generation budget on thinking and emit nothing at all (22 and 34) — one mechanism sampling at the budget boundary, not two populations.
The price, measured per arm — the reconstruct arm's calls and the thinking behind them, and the two routes to the same empty self; the control arms are zero on every axis, drawn as zero. Every bar is a count from the runs.
Why the evaluator is sealed, and outside the agent's write domain: because the first rig's own self-side probe was caught reading the rig. A toggle switched one thing — whether the harness's self-description was rendered into the agent's prompt every turn — and the graded reconstruction-fidelity reading fell from 0.75 to 0.00 while the agent wrote more (1,000 to 1,799 written characters): the probe had been reading the harness's own description back, not the agent's self (§6.5 of the record; doi:10.5281/zenodo.23128113). The sensor had to be moved outside the channel it measures — and the worker's instruments were.
The instrument toggle — two measured readings of one graded probe, one prompt block apart: 0.75 with the harness's self-description rendered in, 0.00 with it removed, the agent writing more text either way. The sensor was reading the rig.
And the work? The sealed evaluator's score does not separate the arms. The no-self reference reads 5.63 and 4.3 (mean of a 0–12 ordinal, the two sets), the carrier 5.07 and 4.25, the reconstruct arm 3.36 and 3.22 — the reconstruct arm is the lowest in both sets, and the gap is not significant in either (D0 vs D1: p = 0.105 and p = 0.348, two-sided). The effect halved on replication and reached p < 0.05 in neither set: directional, unstable, and not claimed. The mechanism is the result; the work score is not the evidence, and nothing here implies the agent's work was damaged.
The honest null — every cell's own score from the sealed evaluator, outside the agent's write domain: D1 is the lowest arm in both sets and the gap is not significant in either; the effect halved on replication. The mechanism is the result, not this.
The full record — the pre-registration, the design, and every number this page quotes in place — is Self-Application Is Not Free: the Monitoring Channel Is Part of the Failure Channel, and the Self Does Not Survive Its Own Reconstruction (doi:10.5281/zenodo.23128113).