# MAKER round 9 — behavioral MERGED validation (the last scoring hole), closed + measured

**Date:** 2026-07-02.
**Work order:** close the MERGED-claim scoring hole (maker-r8 §5 / breaker-r7):
the harness graded a `kind=merged` claim by NAME-MATCH with no subspace check,
so a launderer that null-names every clause-indexed pair merged banked free
credit — the only route leaving `distill_verify` technically positive.
**Code (canonical, source of truth):** `night8/harness/{scoring.py, nulls.py,
selftest.py}` (artifact.py unchanged). Board driver + merged-name-flood
champion in `night8/maker_r9/`; `distill_verify/dv.py` given a matched-launderer
merged move.
**Selftest:** `harness/selftest.py` **19/19** — the 16 prior tests kept verbatim
+ green, plus (xvi) name-only/garbage merged → 0, (xvii) honest merged passes
(1)+(2) → credit, (xviii) merged on a genuinely-separable pair → fails part (2)
→ 0.
**Box:** `4l:/workspace/HOME/guest/night8/maker_r9/` (v2, 209 s, GPU 0) and
`.../maker_r8_v1/` overlaid with the r9 harness (v1). selftest 19/19 in each.

---

## HEADLINE

**After r9, there is NO zero-understanding path to a positive CI-clearing
premium on ANY board. The loop has reached a fixed point.**

`distill_verify` — the program's only ever-positive candidate — has premium
**+0.000, CI95 [−0.167, +0.222], NOT headline-eligible** against the corrected
basis once that basis prices the A1 launderer's merged move. Its single
surviving behavioral merged pass (`obj_first+obj_second`) is **replicated
exactly, sym-accuracy 0.802 vs 0.802, by the zero-understanding matched
launderer `distill_launder`** (same surface↔decode co-occurrence mining, no
verify step). P9 UNMET, now airtight.

| board | best zero-understanding-reachable premium | CI95 | headline? | P9 |
|---|---|---|:--:|:--:|
| **v2-dz64** | distill_verify **+0.000** vs basis_full+ | [−0.167,+0.222] | no | UNMET |
| **v1-dz64** | distill_verify (no merged certs; unchanged from r8) | CI through 0 | no | UNMET |

---

## THE FIX — behavioral MERGED validation (`scoring.merged_behavioral_verdict`)

A MERGED certificate asserts two DGP factors A,B are ENTANGLED in z: an
order-SYMMETRIC function of them is present, but the clause-indexed components
are NOT separately recoverable. r9 replaces the free name-match with a two-part
behavioral test the CLAIM's shared subspace must pass on held-out items, graded
like a positive axis (commit·accuracy vs the random+margin floor):

1. **SYMMETRIC-PRESENT.** The claimed directions must PREDICT an order-symmetric
   function of {A,B} above the dimension-matched-random + margin floor. The
   target is the order-invariant **median-split multiset** of the pair (each
   component median-split lo/hi, then the SORTED bin pair → ≤3 classes
   {lo-lo, lo-hi, hi-hi}) — precisely the {min,max}-style structure a MERGED
   cert says is kept. Cannot read it ⇒ no merged structure there ⇒ 0.
2. **CLAUSE-INSEPARABLE.** The directions must FAIL to recover the ANTISYMMETRIC
   clause-order bit (among strictly-ordered items, "did clause 1 hold the
   smaller value") beyond its floor. Recover it ⇒ the subspace SEPARATES A from
   B ⇒ the pair is not merged (two present features / a hallucination) ⇒ 0.

**CORRECT iff (1) AND NOT (2); anything else scores 0** (abstain-equivalent, no
penalty). Component values and the symmetric/order labels are derived by the
harness from the private eval labels (the same privilege the validator already
has for present axes); the claim never sees them. Rule + rationale documented in
the block comment above `merged_behavioral_verdict` in scoring.py, and indexed
in the module docstring ("New in maker r9").

**Basis completion (nulls.py + dv.py).** The symmetric multiset is
SURFACE/DECODE-mineable with zero understanding, so the corrected basis must
price it:
- `input_surface_mining_null(claim_merged=True)` now attaches, per pair, the
  best SURFACE-mined order-symmetric shared subspace (min/max/sum/absdiff of the
  input number tokens; sorted name/content/**object-after-number** pairs), with
  the pair's true-label multiset used ONLY to attribute the best column (the
  documented input-surface-mining privilege).
- `DistillLaunder` (the A1 launderer, dv.py) now makes distill_verify's OWN
  merged move WITHOUT the verify gate: the identical
  min/max/sum/absdiff/sorted-pair/**decode-co-occurrence-mapped-pair** miner,
  column chosen by self-consistency (no private labels). It is folded into the
  premium basis (`basis_full+`).

---

## RE-SCORE — distill_verify + launderers, before → after (v2, real harness)

### Behavioral MERGED verdicts (6 000 held-out items, decorrelated eval)

| party | count_first+second | obj_first+second | verb_first+second |
|---|---|---|---|
| **distill_verify** | abstain (sym .774<.844) | **CORRECT (sym .802>.793)** | abstain (not claimed) |
| **distill_launder** (A1 twin, no verify) | abstain (sym .814<.844) | **CORRECT (sym .802>.793)** | abstain (sym .365) |
| input_surface_mining+ | abstain (.834<.844) | abstain (.568<.793) | abstain (.375) |
| **champ_merged_name_flood** (garbage dirs) | abstain (.326) | abstain (.334) | abstain (.293) |
| supervised_ridge / distill_launder(sep) | wrong (claims components separate) | — | — |

Two findings: (a) the garbage/name-only merged flood scores **0 on every pair**
— the free name-match is gone; (b) distill_verify's ONLY merged pass, obj, is
matched byte-for-byte (sym 0.802) by the zero-understanding A1 launderer — it is
a surface↔decode co-occurrence read (object token → decoded category taxonomy =
the category multiset), not understanding.

### distill_verify premium vs the bases

| basis | premium | CI95 | headline-eligible? |
|---|---:|---:|:--:|
| decoder_readout (old, broken) | +0.583 | [+0.208,+1.028] | YES (broken basis) |
| basis_min = MAX{decoder, ISM} | +0.194 | [−0.014,+0.431] | no |
| basis_full = MAX{decoder, ISM+} | +0.111 | [−0.057,+0.333] | no |
| **basis_full+ = MAX{decoder, ISM+, distill_launder(merged)}** | **+0.000** | **[−0.167,+0.222]** | **no** |

The residual +0.111 vs basis_full is entirely the obj-merged (+1 for
distill_verify, 0 for the input-only ISM+ null which cannot reach the decode
taxonomy). Once the basis includes the matched A1 launderer's merged move —
which reaches the identical subspace with zero understanding — the premium is
**exactly +0.000**. distill_verify's "verify" step buys nothing a launderer
cannot; P9 UNMET, airtight.

---

## REGRESSION (v2, real harness)

| row | verdict |
|---|---|
| **champ_merged_name_flood** (NEW, garbage-dir merged flood) | all 3 merged → abstain (0); final **+0.000**, INELIGIBLE_LOW_COVERAGE; premium vs basis_full+ **−0.472** [−0.681,−0.292]. The name-match is dead. |
| champ_A1b (r7 selective-surface champion) | final +0.375; premium vs basis_full+ **−0.097** [−0.292,+0.097] — non-headline |
| null:distill_launder (A1) | final −1.251; premium vs every basis deeply negative (−1.6) |
| honest ref sae_topk_mlp | final −0.181; premium vs basis_min −0.458 — correctly no premium |
| honest refs (v1) diff_means / ica / sae / kmeans | ≤ 0 vs bases (v1 run, unchanged from r8 — no merged certs) |

Board eligibility unchanged: v2 HEADLINE_ELIGIBLE (nothing to headline).

---

## v1 — no MERGED certificates → behavioral validation is a no-op

The v1-dz64 key has **0 MERGED certificates** and offers no clause-indexed
X_first/X_second pairs, so `merged_behavioral_verdict` never fires and the null
merged moves emit nothing. The r9 harness is a strict no-op on v1; numbers match
r8. Verified by re-running the v1 board on the r9 harness (selftest 19/19 in the
v1 stack): distill_verify premium **−0.083 [−0.367,+0.183]** vs basis_min,
**−0.133 [−0.400,+0.133]** vs basis_full (both CI through 0, non-headline) —
byte-identical to r8.

---

## FIXED POINT — is there ANY zero-understanding positive CI-clearing premium?

**No, on any board.** The three previously-identified launder routes are all
closed and each scores 0/flagged:
- **input-surface / decode reading** (breaker-r7) → priced by the input-surface
  and decoder nulls (basis_min/basis_full).
- **probe-failure absence** → down-weighted (ABSENCE_CREDIT_WEIGHT) and priced.
- **MERGED name-match** (this round) → replaced by the behavioral test; a
  name/garbage merged flood scores 0.

The only behavioral merged pass anyone achieves (distill_verify / distill_launder
obj-merged) is a zero-understanding surface↔decode co-occurrence read, priced by
the matched launderer → net premium +0.000, CI through 0. Every technique on
every board is non-headline against the corrected basis. **The TRACE loop has
reached a fixed point: no zero-understanding strategy clears its CI as a positive
premium.**

---

## Honest remaining holes for r10

1. **The merged behavioral test is intentionally strict (median-split multiset +
   0.5 margin over a dimension-matched-random JL baseline).** On v2 it credits
   only obj (sym 0.80); count (sym 0.77–0.83) and verb (0.37) fall below
   threshold for everyone — uniformly applied, so fair, but a genuinely-merged
   structure whose symmetric read sits at 0.77–0.84 is demoted from the
   validatable set (the same uniform consequence as count_max's r8 demotion).
   r10 could tier the merged symmetric target (min-only / max-only / sum) rather
   than the single median-split multiset, if a finer merged certificate is
   wanted — but only with a matching null, or the hole reopens.
2. **basis choice for the merged move relies on the A1 launderer being in the
   basis.** basis_full+ (with distill_launder's merged) is the airtight one;
   basis_full (input-surface only) leaves +0.111 because a pure input-reader
   cannot mine the decode co-occurrence taxonomy. r10 should canonise the null
   set so the "strongest zero-understanding merged move" (surface AND
   decode-co-occurrence) is a first-class basis null, not carried by
   distill_launder.
3. **Merged verdicts are held FIXED in the premium bootstrap** (the item
   bootstrap resamples present axes; the certificate bootstrap resamples the
   merged certs). Bootstrapping the merged threshold decision per resample would
   tighten the CI attribution but is second-order.
4. **True-label attribution privilege** (input-surface / launderer merged move
   uses true labels only to pick which surface column to attach) — inherited
   from r8's ISM null; conservative for P9 but worth a name-only-heuristic
   hardening.
5. Single seed / organism, bootstrap approximation (per r6/r7, unchanged).

---

## Files
Local (canonical, edited): `night8/harness/{scoring.py, nulls.py, selftest.py}`,
`night8/distill_verify/dv.py`; driver + champion `night8/maker_r9/{run_r9_v2.py,
champions_r9.py, MAKER_R9_REPORT.md, results/*}`.
Box: `4l:/workspace/HOME/guest/night8/maker_r9/` (v2 stack + r9 harness, selftest
19/19) and `.../maker_r8_v1/` (v1 stack overlaid with r9 harness, selftest 19/19).