Agent routing (labels: tier:fable / tier:standard)
Heuristic: Fable designs and judges; Opus/GPT-class agents build and assemble. Delegate work whose failure mode is loud (tests fail, tables don't reproduce); keep on Fable work whose failure mode is plausible-and-silently-wrong (statistical semantics, claim-bearing prose, sensitive framing). Issues carry the label; acceptance tests are written into the standard-tier issues so the harness judges the output, not the agent.
Two standing Fable roles (instead of Fable-everywhere):
- PI/editor — reviews every claim-bearing PR (results sections, abstracts, site explainer copy) and runs an adversarial pass before any submission.
- Semantics guardian — changes to the following files require Fable review regardless of green CI: popdgp
metrics/views/floors; every repo's holdout/split logic; every data loader's weight handling; promotion-gate thresholds and floor bands; populace-fit's model.py/qrf.py weight semantics. (This portfolio's three caught bugs — the SCF int16 overflow, harness tail-blindness, fragility-bolding semantics — all passed lint and tests while being scientifically wrong.)
Fable-tier items: popdgp#2 (paper design), populace#302 (holdout masking — load-bearing wall for every non-self-referential claim), populace#304 (system paper), calibration-paper results prose, composition interpretation, Belgium note review (#303, EUROMOD framing). Everything else labeled standard. Human-only: authorship, funding, #290 merge, Axiom-paper timing.
🤖 Generated with Claude Code
Agent routing (labels:
tier:fable/tier:standard)Heuristic: Fable designs and judges; Opus/GPT-class agents build and assemble. Delegate work whose failure mode is loud (tests fail, tables don't reproduce); keep on Fable work whose failure mode is plausible-and-silently-wrong (statistical semantics, claim-bearing prose, sensitive framing). Issues carry the label; acceptance tests are written into the standard-tier issues so the harness judges the output, not the agent.
Two standing Fable roles (instead of Fable-everywhere):
metrics/views/floors; every repo's holdout/split logic; every data loader's weight handling; promotion-gate thresholds and floor bands; populace-fit'smodel.py/qrf.pyweight semantics. (This portfolio's three caught bugs — the SCF int16 overflow, harness tail-blindness, fragility-bolding semantics — all passed lint and tests while being scientifically wrong.)Fable-tier items: popdgp#2 (paper design), populace#302 (holdout masking — load-bearing wall for every non-self-referential claim), populace#304 (system paper), calibration-paper results prose, composition interpretation, Belgium note review (#303, EUROMOD framing). Everything else labeled standard. Human-only: authorship, funding, #290 merge, Axiom-paper timing.
🤖 Generated with Claude Code