# RESULT — 072 deletion-priority (CAPSTONE, closes block H 066–072)

**Verdict.** ★ **When a sentence is pushed past the ~460-bit knee, SONAR sacrifices the DIRECT
OBJECT (patient / "what-was-acted-on") FIRST and keeps WHEN, HOW-MANY, and WHO longest. The
deletion order is PATIENT → AGENT ≈ PLACE → ACTION → QUANTITY → TIME (dies-first → survives-last).
This REFUTES the pre-registered "gist-over-detail" hypothesis: the syntactic predicate-argument
skeleton {agent, action, patient} is NOT preserved as a unit — its object member is the designated
casualty, collapsing into a single repeated generic noun, while sentence adjuncts (time, quantity)
survive better.** The one robust structural law is SUBJECT-OVER-OBJECT: at matched rarity + length
the agent survives ~4.5× the patient — front-loading / subject-protection (069/071) generalized to
argument structure. Bit-allocation runs OPPOSITE to survival (Spearman ρ = −0.37): SONAR spends the
MOST teacher-forced bits on the rare entities (AGENT 56, PLACE 51) it still loses, and the FEWEST
on the patient it annihilates. All predictions FALSE — **Brier 0.395** (I confidently predicted the
wrong direction; the findings cohere). T3-exploratory.

Run: tmux c100_072, DONE 08:24:52Z (~6.5 min after smoke). All 3 GPUs verified free at claim
(35/15/15 MiB) AND box tmux empty at start; NOT left RUNNING. 6 roles {TIME AGENT ACTION QTY
PATIENT PLACE}, all ~1-token, globally-unique fillers; demand ladder K∈{1..6} concatenated 6-role
facts (K1≈20 tok below knee → K6≈106 tok well past), n=48/cell. Span-align **1.00** every cell.

## What ran
Templated facts "On {time}, {agent} {action} {qty} {patient} at {place}." (3 variants rotate the
time/place adjuncts across front/mid/back positions; SVO core order grammatically fixed). Greedy
round-trip → per-role survival {exact / altered(paraphrase) / dropped}; **PRIMARY metric = PRESENT
(not-dropped = exact|altered)**, EXACT reported as secondary fidelity. Teacher-forced per-role-span
I_spec (069/070/071 machinery; partner null = same-structure different-fillers). Rarity (gpt2 bits)
and token-length measured per role and partialled out via numpy OLS survival~role+rarity+logtok.

## Headline numbers

**Deletion-order curve — per-role PRESENT survival vs demand K:**
| role | K1 | K2 | K3 | K4 | K5 | K6 | mean K≥4 |
|------|----|----|----|----|----|----|----------|
| TIME    | .92 | .82 | .54 | .40 | .39 | .34 | **.376** (survives last) |
| QTY     | .98 | .92 | .76 | .45 | .31 | .24 | **.331** |
| ACTION  | .69 | .54 | .50 | .39 | .30 | .22 | **.302** |
| PLACE   |1.00 | .97 | .74 | .41 | .20 | .14 | **.250** |
| AGENT   |1.00 | .96 | .71 | .42 | .21 | .12 | **.249** |
| PATIENT | .79 | .55 | .21 | .09 | .04 | .04 | **.055** (dies first) |

- **PATIENT collapses ~5–7× below every other role** and starts failing *below* the knee (K2 .55).
  Mechanism (decodes): past the knee the object nouns MERGE into one repeated generic token —
  "…4768 bobbins…9701 flagons…163 flywheels…5330 pennants" → "…cuttings…cuttings…cuttings…cuttings";
  "reliquaries/bollards/gantries/trellises" → "reliquaries/…/relics". The *slot* survives; its
  *content* is overwritten with a placeholder.
- **Subject ≫ object:** AGENT present .249 vs PATIENT .055 at high K (K1: 1.00 vs .79) despite matched
  rarity (13.9 vs 13.6 bits) and length — a pure grammatical-role/position protection.
- **WHEN and HOW-MANY are cheapest to keep:** TIME (.376) and QTY (.331) survive best; numbers ride
  070 (short = cheap), temporal words paraphrase gracefully ("midnight"→"the evening").

**Gist-over-detail is FALSE.** Skeleton {agent,action,patient} mean survival **0.202** < adjuncts
{time,qty,place} **0.319** (gap **−0.117**). SONAR does not privilege the predicate skeleton; the
buried core argument (object) is dropped before the peripheral modifiers.

**Bit-priority ≠ survival (ρ = −0.37).** Per-role teacher-forced I_spec (bits SONAR spends):
AGENT 55.6 · PLACE 51.1 · QTY 42.0 · ACTION 23.0 · PATIENT 20.2 · TIME 18.2. The two most
bit-expensive roles (rare person, rare town — 071's expensive entities) get the biggest allocation
yet only ~.25 survival; the cheap TIME gets the least and survives most. **Allocation tracks
intrinsic entity COST (071), not preservation priority.**

**Role effect survives the rarity/length control (P6).** OLS survival~role+rarity+logtok: rarity
coef +0.008/bit (negligible), logtok −0.15 (longer→slightly worse). Adjusted role survival keeps
PATIENT at **0.089** (role dummy −0.216, dominant) vs AGENT .258 / PLACE .27 / QTY .329 / TIME .305.
Adjusted skeleton−adjunct gap **−0.104** — the object-first deletion is NOT an artifact of rarity or
filler length; it is a genuine role/position effect.

## Gates
G2 knee-crossed **PASS** (K6 median 106 tok >70; overall exact drop .86→.14 = .73). G3 bits valid
**PASS** (pad-gate 3.8e-5 <0.01; content I_spec[K1] 26.9 ≥6; align 1.00). G4 z-norm ratio 1.14
**PASS**. G5 content rarity median 12.9 bits **PASS**; matched-triple {agent,patient,place} rarity
spread only 1.27 bits (well matched). **G1 positive-control FAIL as coded** (min-role present 0.688):
this is a *measurement-sensitivity* miss, not a broken instrument — the four clean roles
{AGENT,PLACE,QTY,TIME} are captured at K1 .92–1.00; the two low values are (a) ACTION 0.688 because
verbs PARAPHRASE ("appraised"→"assessed", "hoisted"→"lifted") and synonym substitution scores as
dropped → **ACTION survival is a LOWER BOUND**; (b) PATIENT 0.79 because a few rare plural nouns are
mangled past the edit threshold even below the knee. Neither affects the object-first ordering.

## Predictions (Brier 0.395 — all FALSE)
- P1 (.85) K1 min-role present ≥.90 → **FALSE** (.688; ACTION verb-paraphrase undercount, see G1).
- P2 (.70) gist skeleton > adjunct by ≥.10 → **FALSE, reversed** (gap −.117; adjuncts survive more).
- P3 (.55) TIME dies first → **FALSE, reversed** (TIME survives LAST; PATIENT dies first).
- P4 (.55) ACTION survives best (top-2) → **FALSE** (top-2 = QTY, TIME; ACTION mid, and undercounted).
- P5 (.55) bit-priority predicts survival (ρ≥.5) → **FALSE, reversed** (ρ −0.37).
- P6 (.50) role-effect beyond rarity (adj gap ≥+.10) → **FALSE, reversed** (−.104: real role effect,
  but the object-sacrifice direction, opposite the skeleton-protection I predicted).

**The six misses cohere into a cleaner, single story than the pre-reg:** deletion priority is
neither "gist over detail" nor "cheap over expensive" — it is **OBJECT-FIRST**, a structural
compression law (drop the direct object, protect the subject, keep the peripheral when/how-many),
independent of rarity and length, and *anti-correlated* with how many bits SONAR nominally spends.

## Limitations
Greedy decode only. Object-role vs medial-POSITION are not fully separable: the SVO core order is
grammatically fixed, so PATIENT is always sentence-medial (after the number) while AGENT is
subject-initial — "object dies first" may be partly "buried-medial slot dies first" (both are the
069/071 front-loading law; the design decorrelates only the time/place adjuncts). ACTION survival is
a lower bound (verb paraphrases uncounted; a formal codex neutral judge on ACTION/PATIENT ambiguous
calls is the top deferred follow-up — manual decode audit already confirms the direction). gpt2
mean-per-token rarity is a noisy proxy for multi-piece temporal words (TIME rarity 19 is inflated);
I_spec is the trustworthy cost signal. English, templated (no clean 6-role parse of arbitrary FLORES
→ no external-validity cell). n=48/cell.

## BLOCK H SYNTHESIS (066–072, capacity)
SONAR's bottleneck is a **fixed ~460-bit information budget**, confirmed five independent ways: raw
bits-accounting (066), an English-text plateau (068), the conjunction ceiling (069 LONG), the number
ladder (070), and the entity ladder (071). The capacity **knee is denominated in BITS, not tokens**
(068: the token-knee moves with per-token surprisal but the bit-knee is constant ~330–350; RAND
knees at half the tokens). Composition is **subadditive** — even unrelated clauses compress ~8%
(069, R=.92) — so packing more facts buys sub-linear survival. Degradation past the knee is
**graceful and fluent, never silent**: magnitude/rounding errors on numbers (070), entity merges and
confabulation (071), and here object-noun collapse into a repeated placeholder — the decoder always
emits a fluent sentence, it just overwrites the least-protected content. **Cost is dominated by rare,
high-entropy content**: a rare entity runs ~50 bits (071), 4–6× a short number (070), and these
expensive slots get the largest bit-allocation yet are the first to degrade under pressure. **072 is
the capstone map**: turning the budget into a semantic priority ranking, the compression hierarchy is
**object-first / subject-last among core arguments, with when & how-many the cheapest to retain** —
SONAR preserves WHO-did-something and WHEN/HOW-MANY, and sacrifices WHAT-was-acted-on first; the
predicate-argument skeleton is not stored as an atomic gist, and nominal bit-allocation is
*anti-correlated* with what actually survives.

## Follow-up worth funding? **Y (narrow)**
(1) Separate object-role from medial-position: put the patient sentence-initial (passive / OSV
templates) and re-measure — is it the *object relation* or the *buried position* that dooms it?
(2) Formal codex neutral judge (~100) on ACTION-paraphrase and PATIENT-merge calls to tighten the
lower bounds. (3) Push the same 6-role ladder with a bit-budget x-axis (not K) to place each role's
half-survival demand on the common 460-bit scale. (4) Object-collapse mechanism: is the repeated
placeholder the most frequent / lowest-cost noun, i.e. a prior-fallback?