lukeleeai/calibforge-rollouts

CalibForge agent rollouts

Every scored CalibForge agent episode we have on TACC Horizon, one config per harness x who played, one split per variant. A row is one episode with its verifier outcome and the full conversation; in relay episodes each assistant turn carries the router's label of who wrote it (owner = student, teacher or router).

• Harnesses: Terminus-2 (harbor, JSON command batches) and mini-swe-agent 2.4.6 in tool mode (one bash tool).
• Players: relay = a student model plays and the Qwen3.8-27B teacher repairs / assists / takes over through the relay router; teacher = Qwen3.8-27B alone; student = a student model alone.
• Scored follows the relay readout (data/relay/pilot/readout.py): a verifier result, no harness error, no verifier timeout, not cut by the run's deadline. Pass = reward >= 1. Skipped trials are counted per variant below.
• Jupiter variants (source = jupiter_rendered) are only the rendered SFT training rows that exist on Horizon (final_v2: at most 2 rows per task, 1:1 passes and failures, QA-filtered), not every episode; their conversation is the 09-21 chat-template text in rendered_text (special tokens kept) with the per-assistant-turn owners in rendered_turn_owners, and messages is empty. Their pass rate is 50 % by construction.

Columns

column	meaning
harness, player, variant, pool	harness, who played (relay / teacher / student), variant (= split), task pool (train, heldout300, pool100)
run, arm, source	run name, relay arm (relay_repair / control / student_only), horizon_trial or jupiter_rendered
student_model, teacher_model	models (null when not in the episode)
task_id, trial_name, session_id	CalibForge task, harbor trial dir, router session id
reward, passed	verifier reward, reward >= 1
failure_cause	for failures: context_overflow, format_loop, timeout, false_done or tests_failed (readout.failure_cause)
exception_type	harbor trial exception (agent ends such as ContextLengthExceededError / AgentTimeoutError)
router_ending	the router's own episode ending (student_budget, teacher_budget, context_hard_end, ...)
takeover_trigger, takeover_turn	first sticky takeover (done_claim, loop, no_progress_wait, context_budget, time_budget) and its turn
n_turns	assistant turns in the conversation
n_student_turns, n_teacher_turns	executed main turns by owner from the router log (one per turn, retries collapsed; Jupiter: the manifest's counts)
n_repairs, n_assists	teacher one-turn repairs of unparseable student replies; guided-run one-turn assists
messages	list of {role, content, reasoning_content, tool_calls (JSON string), tool_call_id, owner}; owner is set on assistant turns only, null if the router join failed
rendered_text, rendered_turn_owners, rendered_n_tokens	Jupiter rendered rows only
started_at, finished_at	trial times from harbor's result.json

terminus2_relay

Harness terminus-2, player relay.

variant	runs	student	teacher	episodes	tasks	pass rate	owner match	skipped trials
horizon_0929_studentArmA	relay_hz_a_s1_20260929, relay_hz_a_s1_20260929_topup, relay_hz_a_s2_20260929, relay_hz_a_s2_20260929_topup, relay_hz_a_s3_20260929, relay_hz_a_s3_20260929_topup, relay_hz_a_s4_20260929, relay_hz_a_s4_20260929_topup, relay_hz_a_s5_20260929, relay_hz_a_s5_20260929_topup, relay_hz_a_s6_20260929, relay_hz_a_s6_20260929_topup	laion/snowball-67b-a2b-relay-sft-a-step246	Qwen/Qwen3.8-27B	9,969	2,557	61.6 %	100.0 %	no_verifier_result 1095
jupiter_0928_student0921	relay_full_relaym6_20260928	grug-datakit-sft-20260921 (open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21)	Qwen/Qwen3.8-27B	1,616	1,225	50.0 %	100.0 %	-

• horizon_0929_studentArmA: Terminus-2 relay (arm relay_repair) on Horizon 09-29, CalibForge remainder pool x4: the student plays, Qwen repairs every unparseable student reply for one turn and takes over for good on a done-claim, a loop, a no-progress wait, or when the student view reaches 32k tokens (context_budget).
• jupiter_0928_student0921: Jupiter 09-26..28 Terminus-2 relay (student 09-21, Qwen teacher, same relay_repair rules, 32k context budget). Rendered SFT training rows only (final_v2: at most 2 per task, 1:1 pass/fail, filtered), not every episode; the conversation is the 09-21 chat-template text in rendered_text.

terminus2_teacher

Harness terminus-2, player teacher.

variant	runs	student	teacher	episodes	tasks	pass rate	owner match	skipped trials
horizon_1001	qctl_s1_20261001, qctl_s2_20261001, qctl_s3_20261001, qctl_s4_20261001, qctl_s5_20261001, qctl_s6_20261001	-	Qwen/Qwen3.8-27B	9,649	2,540	72.9 %	100.0 %	no_verifier_result 603
jupiter_0926	relay_full_baseline6m_20260926	-	Qwen/Qwen3.8-27B	1,616	1,194	50.0 %	100.0 %	-

• horizon_1001: Qwen3.8-27B alone under Terminus-2 (arm control), Horizon 10-01, same task pool as the relay runs.
• jupiter_0926: Jupiter 09-26 Qwen-alone baseline. Rendered SFT training rows only (final_v2: at most 2 per task, 1:1 pass/fail, filtered); conversation in rendered_text.

terminus2_student

Harness terminus-2, player student.

variant	runs	student	teacher	episodes	tasks	pass rate	owner match	skipped trials
armA_heldout300	heldout_hz_A_20260930	laion/snowball-67b-a2b-relay-sft-a-step246	-	286	286	10.5 %	100.0 %	no_verifier_result 14
0921_pool100	stu100_t2_0921b_20261008	grug-datakit-sft-20260921 (open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21)	-	191	99	5.2 %	100.0 %	no_verifier_result 9

• armA_heldout300: Arm A student alone on the 300 held-out CalibForge tasks (NOT the training pool), Horizon 09-30.
• 0921_pool100: 09-21 student alone on a 100-task slice of the pool, Horizon 10-08.

msa2_relay

Harness mini-swe-agent-2.4.6, player relay.

variant	runs	student	teacher	episodes	tasks	pass rate	owner match	skipped trials
takeover32k	msa_full6_s1_20261003, msa_full6_s1b_20261003, msa_full6_s1c_20261003, msa_full6_s1e_20261003, msa_full6_s2_20261003, msa_full6_s3_20261003, msa_full6_s3e_20261003, msa_full6_s4_20261003, msa_hz_pilot1	grug-datakit-sft-20260921 (open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21)	Qwen/Qwen3.8-27B	5,400	2,610	63.7 %	100.0 %	no_verifier_result 25
hint_takeover32k	msa_hint_s1_20261004, msa_hint_s2_20261004, msa_hint_s2b_20261005, msa_hint_s3_20261004, msa_hint_s4_20261004	grug-datakit-sft-20260921 (open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21)	Qwen/Qwen3.8-27B	5,294	2,612	65.2 %	100.0 %	no_verifier_result 95
hint_takeover16k	msa_early_cf_s0b_20261005, msa_early_cf_s1b_20261005, msa_early_cf_s2b_20261005, msa_early_cf_s3b_20261005	grug-datakit-sft-20260921 (open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21)	Qwen/Qwen3.8-27B	5,147	2,610	69.2 %	100.0 %	no_verifier_result 19
guided_v1	msa_assist_cf_s0b_20261007, msa_assist_cf_s1b_20261007, msa_assist_cf_s2b_20261007, msa_assist_cf_s3b_20261007	grug-datakit-sft-20260921 (open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21)	Qwen/Qwen3.8-27B	5,102	2,606	54.3 %	100.0 %	no_verifier_result 123
guided_v2	msa_assist2_cf_s0_20261007, msa_assist2_cf_s1_20261007, msa_assist2_cf_s2_20261007, msa_assist2_cf_s3_20261007	grug-datakit-sft-20260921 (open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21)	Qwen/Qwen3.8-27B	5,201	2,609	65.2 %	100.0 %	no_verifier_result 25

• takeover32k: mini-swe-agent tool-mode relay: Qwen repairs unparseable student replies and takes over (sticky) on done-claim / loop / no-progress, or when the student view reaches 32k tokens.
• hint_takeover32k: As takeover32k plus the teacher hint: a teacher-only context-economy instruction in the teacher's system message (short outputs, targeted edits, background long jobs, submit once checked).
• hint_takeover16k: Teacher hint, context takeover at 16k tokens of student view instead of 32k.
• guided_v1: Guided: non-sticky one-turn teacher assists on a command timeout, a 4-failure streak, a repeated output and context check-ins at 20k and 35k tokens (at most 6, gap 3); sticky hand-off at 50k; teacher hint on.
• guided_v2: Guided v2: the same assists, sticky hand-off at 40k tokens plus a time hand-off once the student has used 60 % of the task's agent budget; teacher hint on.

msa2_teacher

Harness mini-swe-agent-2.4.6, player teacher.

variant	runs	student	teacher	episodes	tasks	pass rate	owner match	skipped trials
horizon_1003	msa_full6_s1_20261003, msa_full6_s1b_20261003, msa_full6_s1c_20261003, msa_full6_s1e_20261003, msa_full6_s2_20261003, msa_full6_s3_20261003, msa_full6_s3e_20261003, msa_full6_s4_20261003, msa_hz_pilot1	-	Qwen/Qwen3.8-27B	5,852	2,607	73.6 %	100.0 %	no_verifier_result 21

• horizon_1003: Qwen3.8-27B alone under mini-swe-agent (arm control of the msa_full6 runs and msa_hz_pilot1).

msa2_student

Harness mini-swe-agent-2.4.6, player student.

variant	runs	student	teacher	episodes	tasks	pass rate	owner match	skipped trials
msafin_pool1310	msa_selfd_cf_b_20261005	laion/snowball-67b-a2b-msa-relay-sft-step2394	-	2,590	1,306	17.6 %	100.0 %	no_verifier_result 30
0921_pool100	stu100_msa_0921b_20261008	grug-datakit-sft-20260921 (open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21)	-	199	100	5.0 %	100.0 %	no_verifier_result 1

• msafin_pool1310: msafin (the MSA2 relay-SFT student) alone under mini-swe-agent, 1,310-task CalibForge pool x2.
• 0921_pool100: 09-21 student alone under mini-swe-agent on a 100-task slice of the pool, Horizon 10-08.