Olague-Secret/epago-sn36-deepresearch-trajectories

Epago SN36 — trajectories, baselines, diagnostics and mining tooling

Everything produced while investigating model mining on Bittensor subnet 36 (Epago). All data was generated locally by running the subnet's own harness against its own bundled corpus. Nothing here is copied from the subnet's private artifacts.

---

⚠️ Read this first: coronation is currently impossible

On EpagoFoundation/epago @ 7ddfef0 (latest origin/main as of 2026-09-10), no challenger can ever be crowned, regardless of quality.

The mechanism requires clearing the bar twice: a round duel, then a confirmation duel. The round duel mints its exam correctly, dispatching on the release type (service.py:1305-1312):

release = self.cfg.eval.taskgen_release          # "POOL1"
if not is_sealed_release(release):               # POOL1 -> False, takes the sealed-pool path
    return self.deps.generate_tasks(...)

The confirmation duel skips that dispatch (service.py:856-861):

public_tasks = self.deps.generate_tasks(
    seed=round_confirmation_public_seed(start.block_hash, start.round),
    release=self.cfg.eval.taskgen_release,       # "POOL1" handed straight to the generator
    ...

and the generator does not know that release:

generate_tasks(release='POOL1') -> ValueError: unknown taskgen release: 'POOL1'
                                   (known: ['R1','SCI1','SCI2','SCI3','SCI4'])

The except Exception at service.py:877 swallows it, records confirmation_failed, and returns None. The caller then discards the winner:

if confirmation is None or not confirmation.outcome.accepted:
    # A failed confirmation *mint* demotes as well, never crowns
    winner = None

CORONATION_CONFIRMATION_DUELS defaults to 1, so this path always runs. The same bug hits _maybe_calibrate (service.py:1633), so noise_floor never calibrates and stays pinned at the static CROSS_GPU_NOISE_BUDGET = 0.03.

Consequence for miners: the best reachable outcome is NEAR_MISS, which earns zero emissions (the arena pays former kings only), permanently spends the hotkey, and grants one re-duel that hits the identical bug.

Two caveats: the constant is env-overridable, so a validator running EPAGO_CORONATION_CONFIRMATION_DUELS=0 would skip confirmation and coronation would work — but that deviates from the "must clear twice" property the whitepaper relies on, and is not observable from outside. And the fix is one line (_confirm_winner should call self._public_tasks(...)).

---

Provenance

Model	Alibaba-NLP/Tongyi-DeepResearch-30B-A3B @ 4b0ac5767427a55d08a254f0367e2934976598e0
Harness	epago-harness-v4.3 — search + visit, native <tool_call> JSON, <answer> tags
Corpus	data/corpus-science-5792 bundled in the repo — 5,792 papers, all four OpenAlex domains
Generator	epago.taskgen.generator.generate_tasks, release SCI4
Budgets	40 turns · 100,000 chars · 32,768 ctx · 2,048 action tokens · temp 0.0 · seed 42
Hardware	1× RTX PRO 6000 Blackwell (95.59 GiB, sm_120)

The real duel corpus (EpagoFoundation/epago-corpus-science-2, 50,420 papers) returns HTTP 401 and is not fetchable; the whole EpagoFoundation org has zero public repos. The bundled 5,792-paper corpus is the proxy — same harvest_holdout.py → build_corpus.py pipeline, 9× smaller.

---

results/

diag-sci4-200.jsonl — the decisive measurement

200 SCI4 tasks, gold evidence resolved for all 200, each wrong answer attributed to finding vs deciding by checking whether the gold document's url/title ever entered a <tool_response>.

ACCURACY  80/200 = 40.0%
  wrong_DECIDING (had all evidence)   93   46.5%
  correct                             80   40.0%
  abandoned                           25   12.5%
  wrong_FINDING (never saw evidence)   2    1.0%

Retrieval is essentially solved: 93 vs 2. In 98% of clean wrong answers the model found the gold paper and still failed. The failure is synthesis, commitment and answer format — not search.

Fields: task_id, template, correct, error, turns, gold, answer, answer_len, evidence_total, evidence_seen.

baseline-sci4-64.jsonl — greedy baseline

King under validator-identical settings (temp 0.0, seed 42, repetition_penalty 1.05).

ACCURACY 24/64 = 37.5%    TURNS median 13, max 33 (cap 40)
answered 58 (90.6%) · context_budget 3 · malformed_action 3
ZERO turn_cap deaths, ZERO timeouts — the turn budget is not binding

Per template: constrained_study 7/10 (70%) · named_set_superlative 7/21 (33%) · named_set_count 10/33 (30%). Research errors are systematically under-counts (gold 4 → 3, gold 3 → 0).

The answer-length effect — ANSWER_MAX_CHARS = 200 truncates rather than rejects:

answer length	n	correct
≥ 200 (hit the cap)	11	0
60–199	21	12
< 60	26	12

Every answer that hit the cap scored zero — the model writes a markdown essay and truncation discards the conclusion before it states an answer.

traj-sci4-192x4.jsonl — 123 SFT trajectories

Rejection-sampled: 192 tasks × 4 rounds at temperature=0.8, top_p=0.95 through the real harness. Kept only episodes that finished cleanly and answered exactly correctly, then the cheapest survivor per task.

768 episodes -> answered 667 · usable(correct+clean) 368 · context_budget 66 · malformed 35
coverage 123/192 tasks = 64.1%   kept: median 11 turns, max 32

368/768 = 47.9% sampled vs 37.5% greedy. Sampling beats greedy decoding by ~10 points — the competence is already there; greedy just doesn't reach it.

Token stats (king's tokenizer): full sequence median 12,529 / p90 22,004 / max 28,974 (all fit 32,768). Assistant-only median 3,147 / p90 5,708 / max 8,414. Total 1,699,909 tokens, of which 437,801 (25.8%) are assistant tokens.

⚠️ Loss masking is mandatory

Train on assistant turns only. The user turns containing <tool_response> are environment output — real corpus text, 74% of all tokens. Training on them teaches the model to hallucinate search results instead of calling the tool.

⚠️ Edge cases in this file

95 records are fully clean. 25 contain a _NUDGE — meaning the episode had a malformed assistant turn that then recovered; training on that turn teaches malformed output and risks the format probe. 3 contain a _BUDGET_MSG. Both are user messages that are not <tool_response>, which additionally moves last_query_index in the chat template and strips <think> from every assistant turn before them.

---

Chat-template behaviour (verified empirically)

chat_template.jinja computes ns.last_query_index by scanning backwards for the last user message that does not both start with <tool_response> and end with </tool_response>. Assistant turns after that index keep their <think>; earlier ones have it stripped.

For a normal episode [system, user(Question), assistant, user(tool_response), …] that index is 1, so all assistant turns keep their reasoning, and turn-k prompts are exact token prefixes of turn-(k+1) prompts (verified: prefix checks True, <think> count grows 1 → 2).

One flat training sequence per episode is therefore valid — no per-turn expansion needed — except for the 28 records above.

---

scripts/

script	purpose
diagnostic_eval.py	baseline + the DECIDING/FINDING attribution
baseline_eval.py	accuracy + episode death-mode breakdown
gen_trajectories.py	rejection sampling; captures episodes by subclassing harness.Episode and patching the module attribute, so the shipped batched loop runs unmodified
export_legacy.py	un-fuses transformers-5.x checkpoints back to the king's 18,867 legacy tensor names, plus verify_against_king()
preflight_full.py	shipped preflight plus norm_sanity_probe, which epago miner preflight omits
train_lora.py	LoRA SFT loop (designed and benchmarked, not yet run)
bench/	supporting micro-benchmarks (chunked CE, layer timing, tiny-model fixtures)

The tensor_set trap

norm_sanity_probe requires the challenger's tensor name set to equal the king's exactly:

if chall_norms.keys() != king_norms.keys():
    failures.append(ProbeFailure("tensor_set", ...))

Measured on the real checkpoint:

king tensors on disk:              18867      (128 experts × 48 layers × 3 projections + 435)
transformers 5.17 state_dict:        531      (fused 3-D experts, gate+up packed)
missing 18432 · extra 96  ->  a plain save_pretrained FAILS with the hotkey already spent

Un-fusing mapping (verified empirically with a tiny model where gate=1.0, up=2.0):

fused	slice	legacy
experts.gate_up_proj (E, 2I, H)	[e, :I, :]	experts.{e}.gate_proj.weight (I, H)
	[e, I:, :]	experts.{e}.up_proj.weight (I, H)
experts.down_proj (E, H, I)	[e]	experts.{e}.down_proj.weight (H, I)

Order is [gate ; up], no transpose. Had it been reversed, the export would still produce 18,867 correctly-named tensors and pass tensor_set while silently swapping every expert's gate and up projections.

Gates that are not binding

The norm ratio caps (per-tensor 20.0, global 5.0) are degeneracy detectors, not training constraints. A merged LoRA at ρ = ‖ΔW‖/‖W‖ = 4.9 still passes; r=64/α=128 gives ratio 1.0004. Do not design around them. Also note the upstream king repo ships a README.md, which is not an allowed file type and must be deleted before upload.

---

Caveats

• SCI4 is the family behind the duel's private half. The public half (800 of 1,000 tasks, carrying the 99.9% LCB bar) is a sealed POOL1 intersection pool, not reproduced here — minting it needs an LLM verbalizer API.
• Proxy corpus is 9× smaller than the real one, so retrieval is easier; 37.5–40% is optimistic against the ~22% the chain contract records for the king on the real POOL1 exam.
• n=64 / n=200 → roughly ±12pp / ±7pp at 95%.
• Every trajectory is a task the base model already solved at least once in four samples. This data sharpens greedy decoding toward the model's own best behaviour; it cannot teach new capability.
• Prior evidence is negative: docs/EVALUATION.md records rl5 at 19.5% vs base 22.2%. The maintainers trained this model on this exam and got approximately nothing.