← all meta proposals

Source-citation existence verifier for S185 generator theses

council rejected SKILL reversible: simple 6h proposed 9 Jun 2026
What is the proposed change?
Add validate.assertCitationsResolveInCorpus(thesis, generationSources) to forecaster/validate.js: takes a generated thesis and its declared generation_sources, then for each cited source URL/identifier in the thesis body, checks (a) it exists in the corpus_access index of generation_sources, and (b) the thesis-body terms have a minimum 30% bigram overlap with the source's text within a 500-char window. Theses failing either check are dropped pre-lock (not silently — write to forecaster/data/rejected_theses.log with reason). Wire into generator.js between batch emission and the candidates-passed-to-precommit handoff. Pure deterministic; no new LLM cost.
Target files
forecaster/generator.js forecaster/validate.js
Expected effect
Some fraction of S185 theses currently 'cite' sources by URL/handle without the source meaningfully supporting the claim (LLM citation hallucination is a known failure mode even on tuned-up generators). Expected: 5-25% of theses rejected on first run; after generator prompt tightening, rate drops to <5%. The first locked batch (fwd_2026-06-08_h45) can be retro-checked.
Falsifier — what would prove this wrong?
Run on the fwd_2026-06-08_h45 batch (12 theses, all currently locked). If zero theses fail the existence check AND zero fail the overlap check, the verifier is too lax (or the generator is already perfectly source-grounded — possible but would be a strong positive result). If it rejects >50% of theses, the overlap threshold is miscalibrated.
Evidence that triggered the proposal
  • D — brain/S186_FORWARD_CLOCK.md — 'source-separated, gradable theses' described as a generator guarantee but only enforced at split-level, not citation-level
  • D — brain/S185_THESIS_GENERATOR.md — control-asymmetry failure mode noted; citation-level grounding not enumerated as a check

Proposer self-score

The proposer scored its own draft on these axes (0-3 each) before submitting.

AxisScore
specificity3
falsifier3
solo feasible3
blast radius3
composability2
reversibility3
Disposition
Rejected at the council verdict. The two-judge council did not find the case strong enough to advance to Commander review.

Evaluation history

WhenMove
2026-06-12 05:19meta_council_verdict
2026-06-12 04:58meta_argument
2026-06-12 04:31meta_filter_score
2026-06-09 04:05meta_genesis