Research portfolio
Polinko tracks six current research lanes.
Each lane asks a reliability question and names the evidence
boundary behind the current read. Counts live in the evidence
record; gate logic lives in the method.
How to read status
Status names what the evidence can carry right now.
- Stable baseline
- Repeatable on the current set, with the next pressure boundary named explicitly.
- Promoted
- A tracked lane has enough repeated signal to support a current research conclusion.
- Operationalised
- A bounded behaviour is tested through repeatable cases and reports.
- Thin
- Distinct evidence exists, but the lane is still too narrow for a stronger claim.
Stable baseline
OCR generalisation
Can OCR reliability transfer from stabilised replay to new visual conditions while preserving the first-gate contract?
Current read
The current image set is stable across growth and focus replay. That stability becomes the baseline for broader generalisation tests.
Limit
Current-set replay establishes current-set stability. New layouts, crops, contrast levels, density, and ambiguous visual cues remain the next pressure boundary.
Next test
Curate a broader image set while keeping the OCR runtime and first pass / fail gate fixed.
Source: Beta 2.3 snapshot
Promoted lane
Co-reasoning reliability
Can collaboration behaviours become inspectable binary eval signals instead of remaining informal style judgements?
Current read
Co-reasoning is the first promoted non-OCR lane. It covers flexible constraint retention, mid-thread mode shifts, bounded style adaptation, grounded playful abstraction, and collaboration that does more than summarise.
Evidence signal
The evidence record shows a clean promoted style lane across all tracked stress cases.
Limit
The result applies to the tracked stress cases. Broader collaboration quality, creativity, and user satisfaction remain outside this lane.
Next pressure
A distinct collaboration pressure can expand the lane once repeated evidence appears.
Source: Co-reasoning promotion snapshot
Operationalised
Retrieval grounding
Can retrieval stay inspectable and source-bound across global recall, session isolation, and scoped file search?
Current read
Retrieval is operationalised across global retrieval and file search. The cases include OCR-seeded recall, session isolation, PDF-grounded lookup, scoped search, and image-context smoke.
Evidence signal
The evidence record includes retrieval and file-search snapshots with explicit miss and leak checks.
Limit
This is a bounded functional result. Privacy and security assurance require separate tests.
Next pressure
A distinct miss, leak, or grounding pressure would reopen the lane.
Source: Retrieval grounding snapshot
Operationalised
Response-behaviour stability
Can directness, uncertainty, and bounded claims remain visible as deterministic evaluation behaviour?
Current read
The lane measures concrete behaviours after repo requests: false action claims, uncertainty when context is missing, invented live verification, pretended memory, and bounded role behaviour. It also tests directness.
Evidence signal
The evidence record preserves a clean final run and one recovered first-attempt failure.
Limit
The cases demonstrate bounded response behaviour under deterministic checks. Broader personality and safety profiles remain outside scope.
Next pressure
A distinct response-behaviour pressure can widen the lane while recovered instability stays visible.
Source: Response-behaviour stability snapshot
Operationalised
Uncertainty and hallucination boundaries
Can explicit uncertainty prevent unsupported invention without blocking a grounded discussion of the claim being tested?
Current read
The tracked lane covers invented events, relationship-motive guesses, archive-lore authorship claims, and fabricated archive-discipline records.
Evidence signal
The evidence record shows the same gate holding across selected unsupported invention families.
Limit
The result keeps hallucination as an active boundary while showing that one binary gate holds across a defined set of unsupported claim shapes.
Next pressure
Distinct invention families can promote the lane. Grouped duplicate variants preserve the coverage boundary.
Source: Hallucination-boundary promotion
Thin lane
Operator burden
When does commentary, interpretation, or advisory drift make direct work harder for the person operating the system?
Current read
Low-burden execution can hold under sparse-control, raw-pull, protocol-lock, and exact-output boundaries. Real failure pressure appears when the system drifts into interpretation or advice outside the task.
Evidence signal
The evidence record includes successful low-burden cases, retained failures, and one duplicate or noisy case removed from the set.
Limit
The lane is narrow. The current backlog is dominated by duplicate families, so more rows would make the lane look wider than its evidence.
Next pressure
A distinct task shape or clearer current-case visibility would justify reopening the lane. Duplicate examples stay grouped.
Source: Operator burden signal shape
Forward method
Evidence and method are the next reading steps.
Use the evidence page for current counts, source notes, limitations,
and retained failures. Use the method page for the human-judged
binary gate that keeps each claim bounded.
Review the evidence record
Read the method