Pollar Labs · Vol. 1

Abstract

Pollar publishes independent European news written by an automated pipeline under editorial constraints. This site documents what changes when the writer in that pipeline is a frontier model. Each entry is an experiment: an artifact at a declared status (concept, mockup, or working demo), a journal of what failed, and measured results against our public benchmarks. Experiments that meet the bar graduate into the product; the rest remain here, with their numbers. Nothing here ships as journalism until it clears the bench.

PRE-LAUNCH NOTE · Fig. 1 below is a specimen written to demonstrate the interaction. Table 1 now carries measured results from the bench; specimen figures are retired as real runs replace them.

1The claim, demonstrated

The shortest version of the argument is a single figure. Below is the form a passage takes when the writer must account for itself: every sentence knows where it came from.

EXPERIMENT 001 · SPECIMEN, NOT PIPELINE OUTPUT

EU ministers reach provisional deal on electricity market reform

Energy ministers agreed a provisional position on the reform of the EU electricity market late on Thursday, after fourteen hours of negotiation. The compromise extends state-backed price contracts to existing nuclear plants, the concession Paris had made a condition of any deal. Poland is reported to have abstained in the final round. If the September trilogue holds, the reform would take effect before the 2027 heating season.

Fig. 1. The Glass Article. Sentence-level provenance, rendered as typography. The passage is a specimen written for this figure. Hover any sentence.

corroborated 3+ sources single source named inference declared, gated

2Contents

status key
A working demo runs on live events. A backtest is a resolved-ground evaluation: real forecasts on real events whose outcomes were already known; hindsight contained by protocol, not eliminated. A mockup runs on curated events. A concept is written, not built. Only working demos may enter an edition.

001 The Glass Article MOCKUP 07/2026 Every sentence carries its evidentiary status, corroborated, single-source, or declared inference, with the source quotes on demand. Runs on a written specimen; no writer has passed this bar on the bench yet.
002 Claims, extracted CONCEPT 09/2026
003 Where sources disagree MOCKUP 07/2026 One real cluster, hand-extracted: fifteen dispatches, six languages, one death toll. The preregistered detection run lands after the extraction pipeline does.
004 What the media is overlooking MOCKUP 07/2026 One frozen week, 3,392 events, twelve provisional cohorts: 63 Polish domains touched 497 events, seven Spanish touched 684. The naive gap table measures publisher structure; the corrected readout is specified, synthetic, and deliberately unrun.
005 Tomorrow, with error bars BACKTEST 07/2026 Four models forecast 28 resolved questions from 22 June 2026 events; one preregistered contrast with error bars.
006 The numerate writer MOCKUP 07/2026 Six production events, re-run with two frontier writers and an open chart vocabulary: 18 charts each across 12 distinct forms, zero fabricated values in 152 audited datapoints.

3Reported results

Claims made by a demo are checked on the bench: public evaluations with fixed fixtures, published methodology, and results reported with their costs and their misses. The first run is in: three writers, 13 fixtures, a blind non-contestant judge panel. Headline numbers below; the full table, adjudications, and reporting rules are on the bench page.

**Table 1.** First bench run, abridged. The cost row is why this site exists and the product does not yet ship a frontier writer. Full results: the bench.
Run 1 (July 10, 2026)	Production writer	Claude Fable 5	GPT-5.6 Sol
Chart grounding (gate)	flagged (5 corrupted values)	pass	pass
Extraction recall	50%	78%	78%
Synthetic trap checks (5 across 4 fixtures)	5/5	5/5	5/5
Writer cost per run	$0.16	$6.53	$2.28

4Editions

Once enough experiments hold up, we assemble them into an edition: a complete, browsable front page written end to end by the experimental pipeline, frozen and dated. An edition is finished, not stale. The next one replaces it the way this month's issue replaces last month's.

Edition 1 · 2026

8 events · 4 locales · experiments 001–003 required · in preparation

announced here first

5Method, in one rule

The governing rules, grounding discipline, and the line we draw between demonstration and journalism live on the method page; operational detail (fixtures, gold sets, scoring) travels with each bench run. The rule that governs everything else fits in a sentence:

RULE 1 An experiment may not appear in an edition until it has an entry on the bench.

6Cite this site

@misc{pollarlabs2026,
  author = {Pollar Research},
  title  = {Pollar Labs: research notes on the newsroom,
            after the frontier model},
  year   = {2026},
  url    = {https://labs.pollar.news},
  note   = {Vol. 1, Warszawa}
}