Datasetpublic

tts-sft/round5-rt-cells

Round-5 RT-cut (2026-08-19): repair-side self-verifier mechanisms Motivated by the 08-19 returns: evidence-side mechanisms are saturated (labels +2.5pp -> 1 crack; disputes -> 1; property tests -> 1) while the one repair-side mechanism (tr7) converted a DIFFERENT problem set at equal rate. This cut targets conversion directly, with the seed-variance lesson applied: every arm runs 3 replicates (GENSEED 52004/62011/72019, the r5sv2 replicate values) and ships its own concurrent… See the full description on the dataset page: https://huggingface.co/datasets/tts-sft/round5-rt-cells.

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes148downloads

Dataset Card

Round-5 RT-cut (2026-08-19): repair-side self-verifier mechanisms

Motivated by the 08-19 returns: evidence-side mechanisms are saturated (labels +2.5pp -> 1 crack; disputes -> 1; property tests -> 1) while the one repair-side mechanism (tr7) converted a DIFFERENT problem set at equal rate. This cut targets conversion directly, with the seed-variance lesson applied: every arm runs 3 replicates (GENSEED 52004/62011/72019, the r5sv2 replicate values) and ships its own concurrent sv7d baseline.

cell	mechanism
`trx7`	instrumented-trace feedback: on a candidate's first failing test the (patched) probe harness re-executes it under `sys.settrace` and the repair prompt shows the last executed lines WITH variable changes (DebugRepair/TraceCoder-style runtime-state channel; env `TRACE_EXEC=1`)
`rst7`	blocked-restart diversity injection: problems blocked >=50% for >=2 loops get a renderer-side RESTART round — only 4 failed parents shown + a structural-divergence directive (`--blocked-restart`)
`rt7`	routed combo: pwsel base (`--wb-pair --emit-order failcount --emit-top-k 20`) + property validators on any-valid problems + blocked-restart + trace — each mechanism self-gates on problem state
`sv7d`	concurrent baseline, r01/r02 only (r00 = the completed priority run)

22 units = (trx7/rst7/rt7 x r00-r02 + sv7d x r01-r02) x u00/u01 (the 295 screen slice; same problems/budgets/loop-0 as all sv cells). Runner: run_shard_r5_rt.sh (same UNIT+RUN_DIR contract; it additionally exports LCB_TRACE_EXEC and LCB_FB_RESTART from the unit env / restart file). Minipatch = the D-cut minipatch + RT patches; everything opt-in, bit-identical off. CPU validation: validate_rt.py (10 tests) + validate_abcd.py (17-test regression) — run both after cloning, no GPU needed.

Analysis contract: compare cells as 3-seed means, paired per (replicate, problem); never rank single seeds (measured: sv7d moved 34->26/589 across seed contracts). Return whole RUNDIRs incl. `*.summary.jsonl`, `*.restart.jsonl`, `fbaudit.jsonl (its blocks now may carry trace fields), selffb/, metrics_loops.json`.