Skip to content

tryskian/probaboracle

Repository files navigation

Probaboracle

Research Stage Polinko Model Polinko Toy Factory Model Refactor

probably a mini oracle. definitely a mini chatbot

Note

Current status: The Polinko research model is being staged for the next beta.

This is an active refactor window for the model contract, evidence snapshots, docs, and supporting tools. Current builds are kept stable while the repo surfaces are simplified, tested, and aligned for the next release.

Probaboracle is a small, local, agent-backed CLI mini chatbot using the Polinko research model.

It only accepts four question types:

  • what
  • when
  • why
  • where

That narrow surface is the point. Probaboracle is not trying to be a general chat tool. It is a small instrument for studying whether a model can stay coherent, vague, answer-shaped, and product-specific inside tight interaction guardrails.

Current research stage:

  • Polinko research model refactor
  • source state: under audit against the Polinko-family governance contract
  • live eval gate: paused
  • next step: reconstruct the instruction/config blueprint before any runtime or beta claim

Most recent diagnostic snapshot:

  • Research Beta 6.0
  • method: fail-pressure pulse
  • first fixed-prompt pulse: FAIL
  • result: 1 anchor / 13 counted seams / 0 excluded
  • status: diagnostic snapshot, not a clean baseline

Most recently closed beta:

  • Research Beta 5.1
  • retain + evict

In this repo, a new beta gets pinned when the method change alters what the evidence means, not just when wording or procedure gets tidier. The previous clean-baseline reset frame is now under quarantine until the family governance contract is reconstructed and checked against the established Polinko method.

The next baseline is not a live eval run. It is a method-contract repair pass: separate structural config from runtime instruction language, identify patched-in instruction drift, and require human approval before any instruction change becomes a research or beta claim.

What This Repo Demonstrates

  • constrained one-node generation through a fixed prompt surface
  • structural config plus positive agent-owned routing/output targets
  • coherence-first evaluation instead of one overloaded product verdict
  • beta-specific evidence gates:
    • row-level baselines use PASS / FAIL
    • closed Beta 5.1 failures can then earn RETAIN / EVICT
    • the Beta 6.0 snapshot judged one fixed-prompt pulse at a time
    • pulse rows are only evidence inside that pulse:
      • anchor
      • counted_seam
      • excluded_noise
    • the pulse receives one PASS / FAIL verdict
    • the clean-baseline candidate keeps the same pulse method before the next beta-boundary decision

Data Viz Direction

Probsie charts follow the eval shape first. Shared chart families are useful only when the data shape naturally matches.

The initial visualisation set is:

  • pulse charts:
    • stacked horizontal bars for anchor, counted_seam, and excluded_noise
    • grouped or faceted pulse comparison for snapshot versus clean baseline
  • detail table:
    • row id, prompt, output, pulse label, reason, and seam note below the chart
  • row and lens charts:
    • row-level pass / fail / pending stack by prompt type
    • prompt-by-lens table heatmap
    • fail-family horizontal bars
    • correction slope only when true before/after pairs exist

Run It

make install
probaboracle

The app opens a compact terminal loop. Choose a question type with the arrow keys, press enter, or hit esc to exit.

The operator commands, eval workflow, and setup checks live in the runtime runbook.

Core operator commands:

make start
make end-preflight
make end-git-check
make caffeinate-status
make decaffeinate
make check

Read Next

Licence

Apache-2.0. See licence.


Probaboracle is not a resource for yes, no, or maybe.

Contributors