feat: kOREInference — Algorithm 4 iDRegEx with MDL scoring + ensemble integration #1

Merged

tobi merged 17 commits from feature/kore-inference into main 2026-07-01 14:08:19 +00:00

Conversation 2 Commits 17 Files changed 10 +1058 -496

tobi commented 2026-07-01 12:50:23 +00:00

Owner

Copy link

What

Implements kOREInference following arXiv 1004.2372 Algorithm 4 (iDRegEx) exactly:

1. For k = 1..kmax, N random trials:
   • iKoa (Algorithm 1) → deterministic k-OA
   • rwr² (Algorithm 3) → k-ORE expression
   • Validate determinism + k-occurrence
2. Score all candidates by MDL (model cost + data cost)
3. Return the best (KOA, expression, k)

Changes

File	Change
bex/kore.py	Rewrote. Removed broken PTA→Shrink→Repair→StateElimination approach. Replaced with proper Algorithm 4 pipeline wrapping ikoa + rwr_sq + MDL
bex/ensemble.py	Added kOREInference as 3rd algorithm in infer_ensemble alongside CRX and iDRegEx. Refactored prefer logic into a clean dispatch table
bex/__init__.py	Export kOREInference and validate_k_ore
tests/test_kore.py	32 tests — core inference, edge cases (empty, single, 20×identical), validate_k_ore, MDL scoring, paper-faithful assertions
tests/test_ensemble.py	19 tests — ensemble runs all three, prefer=crx/idregex/koreinference, edge cases, stochastic stability

Test count: 79 total (28 existing + 32 kORE + 19 ensemble), all passing.

Key finding

Applied to 3,926 real opencode step-boundary tool sequences, kOREInference returns None — rwr0 cannot handle the interconnectivity. The tool graph (read→bash, read→grep, read→glob, bash→read, etc.) is not SORE (Single Occurrence Regular Expression). This is a genuine empirical result: the agent's behavior within steps is probabilistic, not grammatically structured.

Next steps

• Merge when ready
• The examples/ directory (untracked) has exploratory scripts for the session-data analysis

## What Implements `kOREInference` following **arXiv 1004.2372 Algorithm 4** (iDRegEx) exactly: 1. For k = 1..kmax, N random trials: - iKoa (Algorithm 1) → deterministic k-OA - rwr² (Algorithm 3) → k-ORE expression - Validate determinism + k-occurrence 2. Score all candidates by **MDL** (model cost + data cost) 3. Return the best (KOA, expression, k) ## Changes | File | Change | |------|--------| | `bex/kore.py` | **Rewrote.** Removed broken PTA→Shrink→Repair→StateElimination approach. Replaced with proper Algorithm 4 pipeline wrapping `ikoa` + `rwr_sq` + MDL | | `bex/ensemble.py` | Added kOREInference as 3rd algorithm in `infer_ensemble` alongside CRX and iDRegEx. Refactored prefer logic into a clean dispatch table | | `bex/__init__.py` | Export `kOREInference` and `validate_k_ore` | | `tests/test_kore.py` | **32 tests** — core inference, edge cases (empty, single, 20×identical), validate_k_ore, MDL scoring, paper-faithful assertions | | `tests/test_ensemble.py` | **19 tests** — ensemble runs all three, prefer=crx/idregex/koreinference, edge cases, stochastic stability | Test count: **79 total** (28 existing + 32 kORE + 19 ensemble), all passing. ## Key finding Applied to 3,926 real opencode step-boundary tool sequences, `kOREInference` returns `None` — rwr0 cannot handle the interconnectivity. The tool graph (read→bash, read→grep, read→glob, bash→read, etc.) is **not SORE** (Single Occurrence Regular Expression). This is a genuine empirical result: the agent's behavior within steps is probabilistic, not grammatically structured. ## Next steps - Merge when ready - The `examples/` directory (untracked) has exploratory scripts for the session-data analysis

tobi added 1 commit 2026-07-01 12:50:24 +00:00

feat: implement kOREInference (Algorithm 4) with MDL scoring, add to ensemble, 79 tests edd6d9d4dd

tobi added 1 commit 2026-07-01 13:09:17 +00:00

feat: core+outlier analysis via min_coverage parameter, 6 new tests 9045769d57

tobi commented 2026-07-01 13:09:23 +00:00

Author

Owner

Copy link

Neu: min_coverage Core + Outlier Analyse

infer_ensemble(sequences, min_coverage=0.8) findet jetzt zusätzlich den engsten Kern per iterativem CRX + Outlier-Removal:

• Outlier werden nach Symbol-Seltenheit identifiziert (Sequenzen mit seltenen Symbolen zuerst entfernt)
• Bis genau min_coverage (default 1.0 = kein Filter) der Sequenzen übrig ist
• Rückgabe enthält result['core'] mit {grammar, coverage, outlier_count, outliers}

Beispiel: 15 Ansible-Rollen

min_coverage=1.0: fail.include_vars.set_fact.package.file.template.service?.npm?.pip?.lineinfile?  (alle 15)
min_coverage=0.9: fail.include_vars.set_fact.package.file.template.service?.(lineinfile+npm)?       (13/15, outliers: mysql, docker)
min_coverage=0.8: fail.include_vars.set_fact.package.file.template.service?.npm?                    (12/15, + nginx)
min_coverage=0.7: fail.include_vars.set_fact.package.file.template.service?                         (10/15, + phpmyadmin, apache)

Die Outlier sind Rollen mit den seltensten Symbolen (npm, pip, lineinfile). Ein LLM sieht: "10/15 Rollen folgen dem Kern. Nur wer spezifische Tools braucht, fügt extras hinzu."

Tests

85 Tests insgesamt (+6 neue für min_coverage), alle grün.

## Neu: `min_coverage` Core + Outlier Analyse `infer_ensemble(sequences, min_coverage=0.8)` findet jetzt zusätzlich den **engsten Kern** per iterativem CRX + Outlier-Removal: - Outlier werden nach **Symbol-Seltenheit** identifiziert (Sequenzen mit seltenen Symbolen zuerst entfernt) - Bis genau `min_coverage` (default 1.0 = kein Filter) der Sequenzen übrig ist - Rückgabe enthält `result['core']` mit `{grammar, coverage, outlier_count, outliers}` ### Beispiel: 15 Ansible-Rollen ``` min_coverage=1.0: fail.include_vars.set_fact.package.file.template.service?.npm?.pip?.lineinfile? (alle 15) min_coverage=0.9: fail.include_vars.set_fact.package.file.template.service?.(lineinfile+npm)? (13/15, outliers: mysql, docker) min_coverage=0.8: fail.include_vars.set_fact.package.file.template.service?.npm? (12/15, + nginx) min_coverage=0.7: fail.include_vars.set_fact.package.file.template.service? (10/15, + phpmyadmin, apache) ``` Die Outlier sind Rollen mit den seltensten Symbolen (`npm`, `pip`, `lineinfile`). Ein LLM sieht: "10/15 Rollen folgen dem Kern. Nur wer spezifische Tools braucht, fügt extras hinzu." ### Tests 85 Tests insgesamt (+6 neue für min_coverage), alle grün.

tobi added 1 commit 2026-07-01 13:16:31 +00:00

docs: add min_coverage to MCP tool + README, include core in output 036a84cc76

tobi commented 2026-07-01 13:16:34 +00:00

Author

Owner

Copy link

Letzter Commit: MCP-Tool + README aktualisiert.

• infer_best_grammar hat jetzt min_coverage Parameter (default 1.0 = deaktiviert)
• Nur wenn Agent explizit setzt (min_coverage=0.8), kommt Core+Outlier-Analyse in der Antwort
• README dokumentiert den Parameter und die Algorithm Selection Guide

Letzter Commit: MCP-Tool + README aktualisiert. - `infer_best_grammar` hat jetzt `min_coverage` Parameter (default 1.0 = deaktiviert) - Nur wenn Agent explizit setzt (`min_coverage=0.8`), kommt Core+Outlier-Analyse in der Antwort - README dokumentiert den Parameter und die Algorithm Selection Guide

tobi added 1 commit 2026-07-01 13:18:34 +00:00

docs: update README and SHOWCASE for kOREInference + core/outlier analysis 0886e5f3bc

tobi added 1 commit 2026-07-01 13:21:57 +00:00

docs: fix Go lint description (both optional), format outliers in SHOWCASE usage 5e0674bf77

tobi added 1 commit 2026-07-01 13:24:03 +00:00

docs: badges row, nav links, language tags on code blocks 0be1a7fd79

tobi added 1 commit 2026-07-01 13:37:12 +00:00

docs: add Buy Me a Coffee button 929a50c95d

tobi added 1 commit 2026-07-01 13:40:05 +00:00

chore: ignore examples/ a70024397f

tobi added 1 commit 2026-07-01 13:41:39 +00:00

docs: center logo, move coffee button below nav links 4969819dbb

tobi added 1 commit 2026-07-01 13:41:58 +00:00

docs: bump logo 180→216px bb193d000f

tobi added 1 commit 2026-07-01 13:42:17 +00:00

docs: extra blank line above coffee button bc81bbdce1

tobi added 1 commit 2026-07-01 13:48:38 +00:00

docs: syntax highlighting tags in SHOWCASE.md fbdb39bf7c

tobi added 1 commit 2026-07-01 14:00:08 +00:00

feat: replace single-chart Helm with cross-project convention (15 charts, 6 publishers) b34e39d4b9

tobi added 1 commit 2026-07-01 14:02:17 +00:00

docs: xkcd-style chart d96e99d84f

tobi added 1 commit 2026-07-01 14:03:22 +00:00

fix: chart bars use Dervish purple #853E91 ce4088e705

tobi added 1 commit 2026-07-01 14:04:57 +00:00

fix: grouped bar chart with Without Dervish vs With Dervish 2562519718

tobi added 1 commit 2026-07-01 14:05:53 +00:00

fix: broken y-axis bar chart for readability 90d8c69aa7

tobi merged commit d2d57bc431 into main 2026-07-01 14:08:19 +00:00

tobi referenced this pull request from a commit 2026-07-01 14:08:19 +00:00

Merge pull request 'feat: kOREInference — Algorithm 4 iDRegEx with MDL scoring + ensemble integration' (#1) from feature/kore-inference into main

tobi deleted branch feature/kore-inference 2026-07-01 14:08:19 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: tobi/grammar-inference-engine#1

No description provided.