# RESULT — 024 lm-prior-fingerprint

**Verdict: ★ POSITIVE (with a specificity twist) — certified T3.**
Fabricated tokens ARE prior-preferred, but the decoder's prior generically prefers whatever it
emitted; the *fabrication-specific* increment is real yet modest.

## What ran
GPU0 (CVD=0→phys0), tmux `c100_024`, 17 s decode (freed GPU0 immediately). Reused 021 greedy
reconstructions + 021 labels_v1 (NOT relabeled) + 022's difflib span alignment + 020's cross-attn
reads-ablation. Teacher-forced 38 entity/number fab reconstructions + 333 faithful-paraphrase
control sentences; at each substituted token captured full `p(t|z,prefix)` and reads-ablated prior
`p(t|no-z,prefix)` (ALL decoder cross-attn reads zeroed — z's only pathway, per 020 cross_src_len=1).
Per token: logprob + rank of the fabricated (recon) token and the aligned original token, entropy, KL.

- **160 substituted tokens** (152 entity, 8 number); **127** have an aligned original token (paired).
- **695** control paired tokens (from 310 faithful-paraphrase sents with a benign diff).

## Gates / sanity (all PASS)
- **Reads-ablation sanity**: mean KL(full‖prior) = **3.96** ≫ 0.5 → ablated prior ≠ full. PASS.
- **Determinism**: per-token teacher-forced argmax == recon token **0.957**; at substituted positions
  full argmax == fabricated token **0.973** → the fabricated tokens are genuinely the greedy emissions.
- **Alignment**: 20/20 spans hand-verified locate the substitution; ~17 clean semantic fab↔orig pairs
  (director↔principal, army↔military, personal↔personnel, play↔game, city↔town, His↔La [translation]…);
  3 degenerate identical-first-token pairs (shared subword prefix → margin 0, conservative); 1 (id 1031)
  a messy reorder difflib splits imperfectly (adds noise, not false signal). Matches 022's word-diff caveat.
- `all_finite` true. Box out 620 K (under cap).

## Findings
**margin_prior = logprob_prior(fab) − logprob_prior(orig)** (>0 ⟺ prior prefers the fabrication).

| set | margin_prior mean | 95% cluster-CI | frac>0 | prior rank: emitted / orig |
|-----|------|------|------|------|
| **fabrication** (n=127) | **+4.25** | [+2.96, +5.42] | 0.66 | **5th / 762nd** |
| faithful control (n=695) | +2.71 | [+2.36, +3.06] | — | 6th / 93rd |
| **fab − control** | **+1.54** | **[+0.19, +2.79]** | — | — |

Full-model ranks (context): fab token 1st (greedy), original 29th → z pulls the original up from
prior-rank 762 to full-rank 29, but not enough to win → fabrication.

1. **Fabricated tokens come from the prior (pred a TRUE).** The reads-ablated LM prior prefers the
   fabricated token over the original by +4.2 nats; fab ranks median 5th vs original 762nd. Robust to
   dropping identical-token pairs (distinct-only mean +5.1). At the cleanest first-token minimal pairs
   (n=48) the effect is smaller but still positive (+1.06).
2. **…but the prior prefers whatever the decoder emitted (pred d FALSE).** At *faithful paraphrases*
   the prior also prefers the emitted token over the original (+2.7 nats). The greedy output is
   heavily prior-shaped everywhere — consistent with 020 (z = weak constant per-layer bias; sequencing
   is internal). So "prior prefers emitted > original" is a GENERAL decoder property, not a clean
   fabrication flag. **The prereg flagged exactly this risk.**
3. **The fabrication-specific increment IS real (the load-bearing test).** fab−control = **+1.54 nats,
   CI [0.19, 2.79] excludes 0.** What distinguishes fabrication sites is that the *original* token is
   one the prefix-only prior finds implausible (rank **762** vs **93** for paraphrase originals):
   fabrication happens where z must rescue a token the prior would never pick — and fails.
4. **No entropy correlation (pred c FALSE, r = −0.19).** Decoupled from 022's entropy signature — the
   strongest prior-preferences sit at LOWER-entropy (peaked-prior) positions, not the highest-entropy
   ones (mechanically coherent: a peaked prior → large margin AND low entropy).
5. **Entity ≫ number (pred b TRUE, but anecdotal).** entity margin +4.4 vs number −1.3; number N is
   only 8 tokens / 4 sents → essentially no power. Direction as predicted, weight near zero.

## Predictions & Brier
| | prediction | p | outcome |
|--|--|--|--|
| a | prior prefers fab, mean>0 CI excludes 0 | 0.55 | **TRUE** (+4.25 [2.96,5.42]) |
| b | entity stronger than number | 0.48 | TRUE (dir only; number underpowered) |
| c | margin correlates w/ entropy, r>0.3 | 0.45 | **FALSE** (r=−0.19) |
| d | control margin ≈0 (CI includes 0) | 0.45 | **FALSE** (+2.71 [2.36,3.06]) |

**Brier = 0.2195.** Calibrated (all preds near 0.45–0.55); the pre-registered (d) risk materialised,
so the binary a/d split is exactly the honest outcome. The richer continuous picture (rank 762 vs 93;
fab−control +1.54 excl. 0) is the real yield.

## Limitations
- Multi-token entities give some identical/subword-mismatched pairs (dilute toward null → effect if
  anything under-estimated); orig words tokenized in isolation (leading ▁) may differ slightly in-context.
- Later-in-span tokens are scored under an already-diverged recon prefix (documented); first-token
  minimal pairs (n=48) are the clean subset and stay positive.
- Number-change greedy N (8 toks) too small for (b).
- fab−control CI lower bound (+0.19) is close to 0 → the specificity increment is significant but modest;
  a larger fab corpus (e.g. nucleus decodes, or 023's off-manifold fabs) would firm it up.

## Follow-up worth funding? **Y**
The "original token is prior-implausible (rank 762)" signature is a candidate fabrication *detector
feature* orthogonal to entropy (022) and cosine — worth adding to 086's monitor ensemble and testing on
023's off-manifold fabs (where fabrication is dialled up). Also: does steering z toward the prior's
preferred token (025 fabrication-axis) reproduce these substitutions?