HarleyCooper commited on
Commit
588b6b4
·
verified ·
1 Parent(s): 37a4fcf

verifier-loop experiment spec v1.0

Browse files
Files changed (1) hide show
  1. SPEC.md +222 -0
SPEC.md ADDED
@@ -0,0 +1,222 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # The Verifier Loop: executable grammar rules as the RL environment
2
+
3
+ **Spec for the next Dakota1890 experiment.** Version 1.0, 2026-11-12.
4
+ Codebase it extends: [`HarleyCoops/Dakota1890`](https://github.com/HarleyCoops/Dakota1890) (`lexgym/`).
5
+ Data it reads: [`HarleyCooper/riggs1890-lexicon-graph`](https://huggingface.co/datasets/HarleyCooper/riggs1890-lexicon-graph).
6
+
7
+ ---
8
+
9
+ ## 0. What changed since the first framing
10
+
11
+ The first framing of this experiment assumed most of the pipeline had to be
12
+ built. Reading the repos shows otherwise. Three parts of the design **already
13
+ exist** and should not be rebuilt:
14
+
15
+ | Design element | Already implemented as | Where |
16
+ |---|---|---|
17
+ | Reverse clues, answers anchored | `en_to_dak` task: English word → Dakota form, scored against the accepted set | `lexgym/tasks.py`, reward in `lexgym/scoring.py` |
18
+ | Accepted set = every headword sharing the gloss | **Two-witness gold**: only forms attested in both halves of the book are accepted answers | graph builder rule 3 |
19
+ | "May vary the question surface, must not synthesize a new answer" (the Cree contract) | The Cree anchored-expansion run: 490,280 balanced rows, answers dictionary-bound | Cree1865, run `cree1865-synthetic-expansion-v1` |
20
+ | Frozen measures by split | `heldout_rule` (5,284 tasks, unseen families) and `recall` (600 tasks, training vocabulary) | `rl_tasks` config |
21
+ | Lookup / abstain behavior | `LOOKUP: <word>` tool, reward only on final `ANSWER:` | `lexgym/tinker_env.py` |
22
+
23
+ What does **not** exist yet, and is the whole delta of this experiment:
24
+
25
+ 1. An **executable morphology checker** — code that can say *"this form is
26
+ well-built by rule X"* and *"this form breaks rule Y"*. The current Dakota
27
+ grammar scoring is soft overlap (character F1, affix presence); it cannot
28
+ reject a form.
29
+ 2. **Three-label triage** of every synthetic row: attested / rule-licensed /
30
+ rejected, with a promotion rule that keeps rule-licensed rows out of gold.
31
+ 3. The **generate → verify → retrain loop** that grows the dataset to the
32
+ ceiling of the book, with policy-gradient RL reserved for behavior
33
+ (lookup/abstain) only — the 2026-11-10 runs showed RL adds no facts
34
+ (English→Dakota exact match 0.00→0.03 after ~200 GRPO steps).
35
+ 4. The **unverified control** that prices the verifier: the Stoney-RL setup
36
+ (model writes the answers, reward placeholder returns 0.0) run as an arm,
37
+ not as an accident.
38
+
39
+ Also new since the first framing: the Inkling-2-Small SFT run lifted
40
+ memory recall from 3% → 83.5% in three passes. The binding constraint on
41
+ "the maximum of what is in the book" is now (a) the quality of the extracted
42
+ gold and (b) how much surface area the training rows cover — exactly the two
43
+ things this loop works on.
44
+
45
+ ---
46
+
47
+ ## 1. The claim under test
48
+
49
+ > Grammar rules compiled into executable checks can (a) clean the extracted
50
+ > gold, (b) label synthetic rows as licensed or rejected, and (c) raise
51
+ > held-out recall when rule-licensed rows are added to supervised training —
52
+ > without ever promoting an unattested form to gold.
53
+
54
+ Falsified if: checker-approved rows do not change held-out recall, or if the
55
+ rule-violation rate in free output does not fall relative to the baseline arm.
56
+
57
+ ---
58
+
59
+ ## 2. The checker
60
+
61
+ ### 2.1 Rule schema
62
+
63
+ Every compiled rule is a pure function over a canonical form plus a rule file
64
+ that admits it:
65
+
66
+ ```python
67
+ @dataclass
68
+ class MorphRule:
69
+ name: str # "pronoun-affix-insertion"
70
+ scope: str # "v." | "n." | "adj." | ... (Riggs POS abbreviations)
71
+ inventory: list[str] # surface strings, e.g. ["wa", "mda", "ya", ...]
72
+ apply: Callable[[str, str], list[str]] # (stem, affix) -> candidate surface forms
73
+ notes: str # citation: Riggs page / grammar section
74
+
75
+ @dataclass
76
+ class Admission:
77
+ rule: str
78
+ n_attested: int # attested forms the rule explains
79
+ n_explained_total: int
80
+ coverage: float # n_explained_total / n_attested_in_scope
81
+ false_positives: int # rule-accepted forms NOT attested, sampled and read by hand
82
+ admitted: bool # coverage >= threshold AND FP rate <= threshold
83
+ ```
84
+
85
+ ### 2.2 The admission test
86
+
87
+ A rule is **admitted** only after it passes the book's own attested forms:
88
+
89
+ - **Coverage gate.** On the attested inflected forms of in-scope entries,
90
+ the rule must explain ≥ 60% (the feasibility probe's simplest rule already
91
+ explains 63% of 6,989 forms; stricter per-rule thresholds are set at
92
+ admission time).
93
+ - **Precision gate.** Every form the rule *accepts* that is *not* attested is
94
+ a candidate false positive. A sample is read against the page images; a rule
95
+ whose accepted-but-unattested sample is > 10% wrong builds is not admitted.
96
+ - **Family holdout.** Coverage and precision are computed on train families
97
+ first; the rule is then frozen and re-scored on held-out families. A rule
98
+ that only works on families it was tuned on is not admitted.
99
+
100
+ ### 2.3 What v0 checks without any compiled morphology (already shippable)
101
+
102
+ From `lexgym/orthography.py`, which is grounded and tested:
103
+
104
+ 1. **Alphabet check** — every character in the Riggs alphabet
105
+ (`ALPHABET` set); anything else is `is_suspect` and is never gold.
106
+ 2. **Shape check** — the form parses as headword syllables + one accent mark
107
+ (`parse_headword`); accent mark terminates its syllable; at most one accent.
108
+ 3. **Attested check** — the form is in the entry's `inflected_forms`, or is a
109
+ headword, or is in the family's attested set. This is the gold criterion.
110
+ 4. **Near-match report** — the form differs from an attested form only by
111
+ n/ŋ (1,494 known extraction suspects). Reported, never scored correct.
112
+
113
+ ### 2.4 Milestone 1: run the checker against the gold (first build, no GPU)
114
+
115
+ The checker's first job is cleaning the gold. The feasibility probe's rule
116
+ ("inflected form = headword with a pronoun affix inserted, optionally ± `-pi`")
117
+ explains 63% of the 6,989 inflected forms; the ~2,576 unexplained forms are a
118
+ **triage list**, and the sampled misses are mostly extraction errors (e.g.
119
+ `abawamdeca` for a headword printed with ć) or forms attached to the wrong
120
+ entry. Output of M1:
121
+
122
+ - `triage_unexplained.parquet` — one row per unexplained form: entry_id, page,
123
+ source_image, candidate correction, suspected error class.
124
+ - `rule_admission.json` — per-rule coverage and precision on train and
125
+ held-out families.
126
+ - Every fix flows back into `lexgym/build_graph.py` inputs, which raises the
127
+ ceiling the last SFT run hit.
128
+
129
+ **Affix inventory is M2, not M1.** The 1,497 extracted grammar rules live in
130
+ LFS and the pronoun/paradigm tables are Riggs grammar pp. 40–57 (scan pages,
131
+ already imaged in `docs/collages/tinker_grammar_tables.jpg`). The inventory
132
+ must be read from those sources, not invented. A rule can license a form but
133
+ not choose it: the affix is inserted word-internally ~60% of the time and
134
+ prefixed ~40%, and which one is a fact about each verb — so licensed forms are
135
+ a superset, and only attestation picks the winner.
136
+
137
+ ### 2.5 Labels
138
+
139
+ | Label | Meaning | Use |
140
+ |---|---|---|
141
+ | `attested` | in the book, passes alphabet+shape | gold, always |
142
+ | `licensed` | rejected by no admitted rule, not attested | training only in the licensed arm; never gold |
143
+ | `rejected` | breaks an admitted rule or the alphabet | excluded; logged as the model's error signal |
144
+
145
+ **Promotion rule:** a `licensed` row becomes `attested` only with a speaker
146
+ confirmation or a second source. The loop amplifies its own errors the moment
147
+ this is skipped — this is the failure mode the Stoney arm is designed to
148
+ measure.
149
+
150
+ ---
151
+
152
+ ## 3. The loop
153
+
154
+ ```
155
+ lexicon graph (dakota_entries)
156
+ │
157
+ ┌──────────▼──────────┐ ┌─────────────────────┐
158
+ │ generator │ │ executable checker │
159
+ │ (clue templates + │───────▶│ alphabet, shape, │
160
+ │ model rephrasings, │ │ attested, rules │
161
+ │ form proposals) │ └─────────┬───────────┘
162
+ └─────────────────────┘ │ 3-label triage
163
+ ▼
164
+ attested → SFT gold (pass 2)
165
+ licensed → licensed arm only
166
+ rejected → generator feedback
167
+ │
168
+ ┌───────────▼───────────┐
169
+ │ SFT pass 2 (Tinker) │
170
+ │ + RL for lookup/abstain│
171
+ └───────────┬───────────┘
172
+ ▼
173
+ frozen eval: rephrased-clue recall,
174
+ held-out families, violation rate
175
+ ```
176
+
177
+ Rejected rows go back to the generator as *negative* signal (the proposal that
178
+ broke rule X is not retried verbatim), which is what makes this a loop and not
179
+ a filter.
180
+
181
+ ---
182
+
183
+ ## 4. Arms
184
+
185
+ | Arm | Adds to pass-2 SFT | Tests |
186
+ |---|---|---|
187
+ | A. Baseline | current mixed file, no new rows | reference |
188
+ | B. Rephrased clues | model-varied clue wording, answers anchored (Cree contract) | is knowledge tied to prompt wording? |
189
+ | C. Rule-licensed forms | checker-approved unattested forms, labeled as such | does licensed surface area help or pollute? |
190
+ | D. Unverified (Stoney-style) | model-written answers, no checker | what is the verifier worth? |
191
+
192
+ C vs A is the main read; D vs C prices the checker. All arms share the same
193
+ base model, base SFT recipe (LoRA rank 32 on Tinker), and frozen eval.
194
+
195
+ ## 5. Measures (frozen before any arm runs)
196
+
197
+ 1. **Recall under unseen phrasings** — held-out clue templates written after
198
+ training data is fixed; answers scored against the two-witness accepted set.
199
+ 2. **Held-out families** — the existing `heldout_rule` split, unchanged.
200
+ 3. **Rule-violation rate in free output** — sample free generation, run the
201
+ checker, report % rejected. Arm D is predicted to be highest.
202
+ 4. **Neither-attested-nor-licensed rate** — forms the checker can't place.
203
+ This is the error-amplification detector.
204
+
205
+ ## 6. Compute shape
206
+
207
+ - M1 (checker + triage): CPU only. No training.
208
+ - M2 (inventory compile): CPU + page-image reading.
209
+ - M3–M4 (generation + arms B/C/D + baseline): Tinker LoRA, ~4 passes
210
+ comparable to the $11.23/pass Inkling run; generation rollouts via the
211
+ existing Tinker sampling path.
212
+ - Eval: the frozen sets, temperature 0, 1,500 tasks — same harness as the
213
+ 2026-11-12 run.
214
+
215
+ ## 7. Limits
216
+
217
+ - A licensed form is a hypothesis about 1890 orthography rules, not about
218
+ Dakota as spoken. It never enters gold without a speaker or a second source.
219
+ - The book is one witness with missionary bias; the checker verifies
220
+ well-formedness *per Riggs*, nothing more.
221
+ - One seed per arm to start; the D arm exists partly because a single noisy
222
+ run of the Stoney pipeline already misled once.