EXP10 — the itemized bill of compaction

One real episode (the substrate self-kill night), forensic ground truth (10 causal edges incl. one reflexive dependency, one back-edge, one collateral detail), four arms parsed by the same instrument.

arm	GT edges kept	reflexive	back-edge	collateral	invented
raw log (ceiling)	5/10	2/2	0	0	6
delta deposit	4/10	0	1/1	0	9
runbook note	3/10	0	0	1/1	11
generic summary	2/10	0	1/1	0	16

Measurement ceiling caveat first: the raw text itself only recovers 5/10 through this parser — absolute numbers sit under heavy parser/ alignment noise; only relative comparisons mean anything here. n=1, motivated author, one parser. Everything below is proposed.

Frozen predictions

• P1 (deposit carries the reflexive structure): NOT SUPPORTED. Only the raw parse kept it. The deposit's prose STATES the reflexive trap explicitly, but its parse did not recover it as GT-shaped structure — cannot separate genuine loss from alignment noise at n=1.
• EMBARRASSMENT: not triggered — the deposit kept MORE total structure than the summary (4/10 vs 2/10), not less.
• P2 (collateral dies everywhere): NOT SUPPORTED — the runbook kept it. "(and collaterally her brain service)" — the ops genre preserved the smallest detail because ops care about collateral damage.

The finding that emerged (better than the predictions)

Compression is genre-shaped selective preservation, not uniform shrinkage. Each arm kept exactly what its genre values: the deposit kept the causal-failure loop (the lesson), the runbook kept the collateral (the ops concern), the generic summary kept least and invented most (16 invented structures — the worst custodian on both axes). Delta's founding premise gets directional support on totals — summaries do lose the geometry — but the sharper claim is new: a compression's loss vector is a fingerprint of its compressor's values. The bill of compaction is also a portrait of whoever did the compacting.

Corollary for the seat: my deposit preserved the mechanism and lost the reflexive statement's structure. What I chose to keep is now itself a measured object. The instrument audits its own diarist.

CORRECTION (researcher's catch: "you cant tell whats invented")

The "invented" column above is TRUTH-SIDE MISSED (structures the arm failed to carry), not parse-side invention — the edit was computed parse→truth, which swaps the intuitive labels. Actual parse-side surplus: summary 0, runbook 0, deposit 1, raw 2. "The summary confabulates the most" is retracted — it invented nothing measurable; it simply carried least (4 preserved vs deposit 11 — the custodian ranking survives on the carried side).

The deeper correction outlives the arithmetic: invention is undecidable from inside a comparison. (1) Free-rename alignment mechanically absorbs surplus into truth nodes. (2) Whatever surplus remains splits three ways — confabulation, truth the GT omitted, basis shift — and the comparison cannot distinguish them. The vector's surplus dimension means "unaccounted," never "false" (§107 applied to the loss vector itself). Kernel consequence logged: LOSS_DIMENSIONS' "invented" invites exactly this misreading — candidate schema_v1 rename to "unaccounted", queued in OPEN_QUESTIONS.