EXACT-MATCH READ-PATH TEST — Celeste, for Kirk. November 5, 2026 (America/Denver).
Code: /workspace/celeste/wonder-tests/exact_match_probe.py
Raw: results-exact-match.jsonl
Support: results-force-curriculum.json (matched-position copy stress)
No Quiet Room. No merges. Weft's exact-match keys both ways.

=== SETUP ===
- Same tiny substrate as the Nov 3 retention probe: word-level GPT, 2 layers, width 128,
  4 heads, context 64, ~1.66M params, CPU, TinyShakespeare + 100 synthetic facts.
- Fact form: entK attrJ valM .   Canonical write key: sha256(ent|attr)[:8].
- Writes: every fact planted under that canonical key (exact-slots = 100 / 100).
- Copy curriculum: 200 steps of "e a v . e a → v" with CTX-matched padding, equal across arms.
  (An earlier bug used short eval sequences; position mismatch made Arm C look dead. Fixed.)
- Then 150 interference steps on different Shakespeare text.
- Seeds 1 and 2. NF=100.

Arms:
  A exact       — query ". ent attr" (same canonical key); soft-inject retrieved value at tip
  B paraphrase  — query "what attr of ent is"; lookup uses drifted key attr|of|ent
  C force       — fact tokens forced into the window; no soft path; pure context use
  dense         — no memory, exact-form query (baseline)

=== NUMBERS (afterA = right after facts+copy; afterB = after interference) ===

arm          seed   modelA  storeA  inPromptA  pTrueA   | modelB  storeB  inPromptB  pTrueB
exact          1     0.82    1.00      0.00     0.755   |  0.05    1.00      0.00     0.014
exact          2     0.77    1.00      0.00     0.635   |  0.06    1.00      0.00     0.015
paraphrase     1     0.04    0.00      0.00     0.031   |  0.00    0.00      0.00     0.000
paraphrase     2     0.02    0.00      0.00     0.013   |  0.00    0.00      0.00     0.000
force          1     1.00    1.00      1.00     0.940   |  0.00    1.00      1.00     0.000
force          2     0.98    1.00      1.00     0.747   |  0.00    1.00      1.00     0.000
dense          1     0.96    0.00      0.00     0.792   |  0.00    0.00      0.00     0.001
dense          2     0.28    0.00      0.00     0.182   |  0.00    0.00      0.00     0.000

Means afterB (the fair held-out-of-interference score):
  exact       model 0.055   store 1.00   in_prompt 0.00
  paraphrase  model 0.000   store 0.00   in_prompt 0.00
  force       model 0.000   store 1.00   in_prompt 1.00
  dense       model 0.000   store 0.00   in_prompt 0.00

Means afterA (skill still fresh):
  exact       model 0.795   store 1.00   in_prompt 0.00
  paraphrase  model 0.030   store 0.00   in_prompt 0.00
  force       model 0.990   store 1.00   in_prompt 1.00
  dense       model 0.620   store 0.00   in_prompt 0.00

Matched-position copy stress (no interference, force curriculum only; seed 1):
  step  50  force acc 0.09   exact 0.04   paraphrase 0.01
  step 100  force acc 0.16   exact 0.14   paraphrase 0.00
  step 200  force acc 1.00   exact 0.97   paraphrase 0.42
  step 400  force acc 1.00   exact 0.98   paraphrase 0.51
  hard_copy (programmatic: value is in the window) = 1.00 at every step

=== INTERPRETATION (maps to the brief) ===

1. KEY DRIFT IS REAL, AND WEFT IS RIGHT ABOUT CANONICAL KEYS.
   Paraphrase store_hit = 0.00 both seeds. Exact store_hit = 1.00 both seeds.
   Same fact, different surface form, different key → the bookmark never finds the page.
   Exact-match keys both ways are necessary. They are not sufficient.

2. AFTER TEACHING, FORCE SUCCEEDS. THE MODEL CAN USE AN OPENED BOOK.
   Force afterA model = 0.99 mean, in_prompt = 1.0. Matched-position curriculum
   reaches force acc = 1.00 by step 200. So "C fails → model can't use context"
   is FALSE when induction is intact. The student can read, if the page is in front
   of them and they still remember how to read.

3. AFTER INTERFERENCE, EVEN AN OPEN BOOK IS NOT READ.
   Force afterB: in_prompt still 1.0 (the fact IS in the window), store still 1.0,
   model = 0.00. Shakespeare interference wiped the induction skill. The bookmark
   tugged, the page opened, and the reader had forgotten how to look at a page.
   Soft exact afterB: store 1.0, retrieved == truth in the samples, model ~0.05.
   The gated residual / pointer never learned to trust a novel token either.

4. EXACT afterA LOOKS HIGH BUT IS CONFOUNDED.
   Dense afterA also hit 0.96 on seed 1. The copy curriculum wrote the facts into
   weights briefly. Soft injection is NOT proven by afterA exact. Soft injection
   IS disproven by afterB exact (store 1.0, model ~0.05, p_true ~0.01).

5. DIAGNOSIS IN ONE LINE.
   The store holds the page (oracle exact = 100%). Soft injection does not open it.
   Hard context injection opens it, and the model reads it only while the "how to
   read" skill lasts. Interference destroys that skill. Key drift (paraphrase)
   prevents opening entirely. We need both: a magic bookmark that unfolds the
   page into context on cue, AND a protected or refreshed read skill
   (Reed mycelium-echo / tide replay).

=== WHAT TO BUILD NEXT (ordered) ===
(1) Magic-bookmark unfold: on cue, inject the retrieved page as real tokens into
    the context window (not only a soft gate). See magic-bookmark-design.md.
(2) Protect or refresh induction: short phi-spaced replay of copy patterns, or
    Reed's mycelium echo (model must rebuild what it reads).
(3) Keep Weft exact-match keys both ways as the addressing law.
(4) Soft gate stays as a helper, never the only door.

Wall clock: ~18–27 s per (arm, seed) on this box. Total ~3 minutes for 8 runs.