Polinko's binary eval gates are determined by human judgement.
They generate signals that trace the pattern of its own behaviour as it's evaluated.
Human judgement enters the method as a pass/fail gate.
Polinko records pass or fail first. That
binary gate gives human judgement a computational shape while keeping
the decision explicit. The question is whether the system held the
stated behaviour under the prompt, source material, and constraints
being tested.
Pass
The system held the exact behavioural boundary being tested in that row or case.
Fail
The system crossed, weakened, or contradicted the stated boundary. The failure then receives a source-backed decision.
If a case fails, retain or evict is
recorded after the gate. Retain keeps an in-scope failure as
evidence. Evict removes malformed, duplicate, or noisy material from
the evidence set.
Every gate stays tied to the material that produced it.
Source artefact
A question may begin in a transcript, chat, notebook, database, runtime trace, eval report, screenshot, or direct interaction. The relevant source stays identifiable throughout the work.
Human judgement
I judge a row or case against its stated boundary, record exclusions, and decide what belongs in the evidence set. The source record comes before the summary.
A row or case remains local until repeated evidence can support a lane summary.
The method moves from a bounded question to a bounded claim.
The process can loop when evidence changes the question, the case, or the implementation. The order still matters because interpretation comes after the gate.
-
01
Define one behavioural question, its source material, and the exact boundary the system is expected to hold.
-
02
Work with AI collaborators to reframe, stress-test, and turn the question into an inspectable test.
-
03
Implement the test while preserving the prompt, source interaction, constraints, context, and corrections that shaped it.
-
04
Judge each row or case against its stated boundary and record
passorfailbefore adding a broader interpretation. -
05
Retain in-scope failures as evidence, remove invalid cases with an explicit reason, and rerun only when the evidence boundary allows it.
-
06
Compare repeated evidence at lane level, then decide whether it supports a bounded conclusion, a revised question, or no claim yet.
A public claim is the last step, not the starting point.
Every promoted claim remains traceable through the source, the human judgement, and the repeated evidence that supports it.
- Source
- The interaction, artefact, report, or record from which the research question and evidence are drawn.
- Binary gate
- A human decision about whether a row or case passed or failed the behavioural boundary.
- Lane signal
- A repeated behavioural shape supported by multiple inspectable rows or cases within one research lane.
- Claim
- A bounded interpretation made when the lane signal remains traceable to its judgements and source material.
The current public method is source-first.
Beta 2.3 provides the current evidence baseline. The next method contract keeps active chat and manual-eval artefacts, row or case judgements, exclusions, and lane summaries connected before any method claim is promoted.
OCR still uses case-level pass / fail. Non-OCR lanes
use binary gates where row or case judgement is earned. Across the
research programme, source-first traceability is the active method
shape.
The earlier fail-pressure pulse proposal remains historical method evidence. The forward method now stays closer to source artefacts, row or case judgement, and repeated lane evidence.
Use the method to read the evidence.
The evidence page shows current counts, source-note links, limitations, exclusions, and retained failures. The diagrams page shows how evidence moves through the system. This page names the judgement rule that keeps those routes bounded.