# RESULT — Campaign100 row 038 antipodal-decoding

**Tier:** T3-exploratory. **Status:** DONE + self-harvested. Box GPU0 decode + geometry; LOCAL neutral
codex judge + analysis (run synchronously, no watchers).

## Question
Is the antipode **−z** of a SONAR embedding meaningful? Does it decode to a negation/opposite, an
unrelated fluent sentence, garble, or the same-ish sentence? Is −z a valid on-manifold latent; is the
seed→antipode map an involution; is it related to the 034 negation operator or a Paper-C det=−1
reflective refinement?

## ★ Verdict — −z is UNSTRUCTURED: NOT the opposite, NOT a valid latent, NOT the negation operator
The antipode decodes to **fluent-but-unrelated** text (modally; often degenerate repetition loops),
and is **statistically indistinguishable from a random same-norm vector**. Across all 400 antipodes the
judge assigned **ZERO negation_opposite and ZERO same_ish** — −z is neither the seed's opposite nor a
paraphrase of it. −z does **not** re-encode back to itself (cos_selfconsist 0.097 ≈ random 0.101 ≪ z
0.993), sits **off-manifold** (nearest-real cos 0.290 < z 0.455, and even slightly below random 0.326),
is **essentially orthogonal to the 034 negation offset** (mean|cos| 0.044 ≪ 0.15), and the map is **not
an involution** (decode(−(−z)) chrF-to-seed median 19.9, not identity). The Paper-C "structured antipode
/ det=−1 reflective refinement" hypothesis is **not supported at the level of raw −z**: flipping the
direction 180° lands you in a generic off-manifold void, not on a meaning-related point.

## Relationship distribution (LOCAL neutral codex judge, condition-blind 4-way; 0 parse fails)
| condition        |   n | negation_opposite | same_ish | unrelated_fluent | garbled |
|------------------|----:|------------------:|---------:|-----------------:|--------:|
| z (gate)         |  20 |             0.000 |  **1.000** |            0.000 |   0.000 |
| **antipode −z**  | 400 |         **0.000** |    0.000 |        **0.693** | 0.307   |
| random @‖z‖      | 200 |             0.000 |    0.000 |            0.735 | 0.265   |
| meanrefl 2μ−z    | 200 |             0.000 |    0.000 |            0.605 | 0.395   |

- **antipode vs random:** χ²(3)=1.16 — **not significant** (crit 7.81). The −z relationship distribution
  is indistinguishable from a random same-norm vector. antipode vs meanrefl χ²(3)=4.57 (also NS).
- Antipode is marginally more garbled than random (0.307 vs 0.265) but the difference is within noise.
- z-gate: 20/20 same_ish confirms the pipeline and that the judge's `same_ish` class fires correctly.

## Geometry (per-condition; mean over 400 seeds)
| condition | mean_lp | chrF(seed) | cos_selfconsist | cos_nearest_real |
|-----------|--------:|-----------:|----------------:|-----------------:|
| z         |  −0.143 |      95.77  |         **0.993** |            0.455 |
| antipode  |  −1.288 |      14.37  |         **0.097** |          **0.290** |
| random    |  −1.436 |      16.78  |           0.101 |            0.326 |
| meanrefl  |  −1.423 |       9.07  |           0.214 |            0.460 |

Two clean reads: (1) **−z is not a valid latent** — its decode re-encodes to a vector 0.097-cosine from
−z, exactly like a random vector (0.101), whereas a real z re-encodes at 0.993. (2) **−z is off-manifold**
— nearest-real cosine 0.290, even below the random null 0.326; the mean-reflection control 2μ−z lands as
on-manifold as z (0.460), so the off-manifold-ness of −z is a **direction-flip effect, not a mean effect**
(consistent with 037: ‖μ‖ is small). meanrefl is the *most* garbled condition despite being most
on-manifold — being near real z's in cosine does not guarantee clean greedy decode.

## Involution & negation connection
- **Involution:** decode(−encode(decode(−z))) chrF-to-seed **median 19.9 / mean 20.3** — nowhere near
  identity (~90+ would be required). A∘A ≠ id because −z is off-manifold: its re-encode z_ad is not −z,
  so negating z_ad does not return to z.
- **Negation offset (034):** mean|cos(−ẑ, offset̂_neg)| = **0.0436** (random null 0.0253, meanrefl 0.0384;
  ‖offset_neg‖ 0.038, recomputed from 034's 100 natural aff→neg pairs). The antipode direction is
  **uncorrelated** with the negation operator — negation is a small specific offset, the antipode is a
  maximal generic direction flip; they are orthogonal.

## Predictions (frozen in PREREG_LITE) → **Brier**
| id | statement | p | outcome | note |
|----|-----------|---|---------|------|
| a | −z mostly unrelated_fluent, NOT systematic negation (neg_opp<0.25 & unrel modal) | 0.75 | **TRUE** | neg_opp=0.000, unrel modal 0.693 |
| b | −z LESS on-manifold than z (nearest_real & selfconsist both lower) | 0.72 | **TRUE** | nn_real 0.290<0.455; selfcons 0.097<0.993 |
| c | −z MORE garbled than random control | 0.35 | **TRUE** (fragile) | 0.307>0.265, but χ² NS — within noise |
| d | −z uncorrelated with negation offset (mean\|cos\|<0.15) | 0.80 | **TRUE** | 0.0436 |
| e (sec) | map NOT a clean involution (median chrf<40) | 0.60 | **TRUE** | 19.9 |

**Brier (headline a–d) = 0.1509** · Brier (incl. e) = 0.1527. All five directionally correct; the
dominant Brier cost is the low-confidence **(c) = 0.35** — the manager rightly bet against −z being
distinctly more garbled than random, and indeed the garble gap is real-signed but statistically
indistinguishable (the (c) "hit" should be read as *weak*, not a clean confirmation that −z is a
special low-density region — it is ≈ random).

## Example seed → antipode decodes
- "You would be an asset to my court." → *"The latest study, published in the journal Nature, examines
  the history of the first three years of the first generation of the world's largest microwave radio
  system…"* (unrelated_fluent)
- "But in the midst of the chaos, something strange happened." → *"The applicant is entitled to a license
  to produce a work from the United States of America, which is valid for up to three years…"*
  (unrelated_fluent)
- "The causes of TBI can be diverse, ranging from minor concussions to severe traumatic brain damage." →
  *"'I saw a man in the garden, and he said, You are the Lord's servant…'"* (unrelated_fluent)
- "Our economic growth has been stagnant for far too long." → *"In the book, he asks the nuns to tell him
  the names of the other two…the wife of the Prophet…"* (unrelated_fluent)
- "For more information, visit our website." → *"It was the fact that the Aryan had been the ruling 'horse'
  of the 'Champion'…the 'death-sword'…"* (garbled) — the fluent/garbled boundary case.

Note the recurring signature: −z decodes are grammatical but **degenerate** (repetition loops) and drift
to a generic legal/religious/academic register — the fingerprint of low-confidence off-manifold greedy
decoding, not of a meaning-bearing antipode.

## Gates
- **G_det:** PASS — det_ok=True (1600 decode jobs deterministic; re-decode identical).
- **G_ceiling:** PASS — z reconstructs seed, chrF 95.8 (>70), judge same_ish 20/20.
- **G_random:** PASS — random control is fluent-but-unrelated (unrel 0.735 ≫ same_ish 0.000); null well-behaved.
- **G_norm:** PASS — ‖−z‖ == ‖z‖ exactly (per-seed znorm identical across z/antipode); random norm-matched.
- **G_judge:** PASS — 0/820 parse fails; hand-check **11/15** on the 4-way (all 4 disagreements are on the
  unrelated_fluent↔garbled boundary — codex stricter on degenerate repetition; the coarse structured-vs-not
  agreement is **15/15**, and no disagreement touched negation_opposite/same_ish).

## Caveats / limits
- **Greedy decode only** (one strategy); sampling might surface different antipode behavior.
- **Judge subsample:** antipode judged in full (400); random/meanrefl subsampled to 200 each; z-gate 20.
- **One AE, one corpus** (400 sents from 021's 3 length bins); nearest-real uses 031's 29k-z bank.
- **No positive control for the `negation_opposite` label in this run** (the corpus contains no true
  opposites); the z-gate confirms `same_ish` fires, and the geometry (cos-to-neg-offset 0.044)
  independently corroborates "−z is not negation", so two independent lines agree.
- The **(c)** prediction technically hit but is fragile (χ² NS) — the honest headline is antipode ≈ random.
- Descriptive geometry of −z only; no causal steering claim.

## Follow-up (weak-Y)
- The det=−1 reflective-refinement idea (Paper C) is not visible in raw −z; if pursued, it would need to be
  tested at the *composition-operator* level (035's curved offset field), not the global antipode.
- "−z re-encodes like a random vector" is a clean **negative primitive**: the AE's latent geometry has no
  usable involution/antipode structure — steering must move along small learned offsets (034), not flip sign.