Research-Stack/6-Documentation/famm/AUTONOMOUS_SPEEDRUN_HARNESS_GATE.md

Autonomous Speedrun Harness Gate

Purpose

Integrate the Prime Intellect auto-nanoGPT lesson into the FAMM stack as an autonomous research harness.

This gate models a long-running research loop as a receipt-bearing search machine:

AGENTS.md / rules
→ goal.md / target invariant
→ plan.md / mutable frontier
→ THREAD.md / durable mission memory
→ variant logs / delta stream
→ run logs / experiment receipts
→ pruning rounds / Warden cleanup
→ source refresh / drift gate
→ statistical receipt

Source boundary

The Prime Intellect run is an external reference pattern, not a dependency. The article describes agents iterating on the nanoGPT speedrun optimizer track, using a markdown harness with durable working memory and run logs. It also documents failure modes: local-loop grinding, insufficient upstream refresh, premature stopping, and weak novelty without external/public-record scaffolding.

This gate imports the harness architecture, not any benchmark-specific optimizer claim.

Core stack

Builder-Judge-Warden
→ Autonomous Speedrun Harness
→ Experiment Delta-DAG
→ FAMM Scar Ledger
→ Source Refresh Warden
→ Statistical Receipt

FAMM object

\mathfrak C_{\mathrm{AutoHarness}}
=
A_{16}(u_{\mathrm{run}})
\otimes
[
\Sigma_{\mathrm{goal}}
+
\Sigma_{\mathrm{plan}}
+
\Sigma_{\mathrm{thread}}
+
\Sigma_{\mathrm{variant}}
+
\Sigma_{\mathrm{run}}
+
\Sigma_{\mathrm{source}}
+
\Sigma_{\mathrm{scar}}
+
\Sigma_{\mathrm{stats}}
+
\Sigma_{\mathrm{receipt}}
]

16D autonomous harness state

X_{\mathrm{harness}}
=
[
\nu,
g,
p,
t,
v,
r,
s,
\mu,
z,
\Delta,
\Omega,
R,
I,
C,
\rho,
\pi
]

Axis	Meaning
0	NUVMAP run/frontier address
1	goal / target invariant
2	plan/frontier state
3	durable thread memory
4	variant / hypothesis stack
5	run-result metric
6	source-refresh freshness
7	semantic mass / promise pressure
8	recurrence / local-loop detector
9	delta-memory / experiment lineage
10	FAMM scar pressure
11	residual / validation gap
12	invariant / benchmark-rule overlap
13	route cost / compute cost
14	receipt strength / statistical validity
15	projection / report gauge

Builder-Judge-Warden mapping

Role	Harness behavior
Builder	proposes optimizer, schedule, initialization, or search-policy mutation
Judge	checks benchmark rules, target metric, seeds/statistical noise floor, and reproducibility
Warden	blocks stale loops, rule violations, overfit/seed hacking, premature pruning, stale upstream source, and unsupported novelty claims

Experiment delta edge

\nu_t
\xrightarrow{\Delta e_t}
\nu_{t+1}

Δe_t = {
  idea_hash,
  patch_hash,
  config_hash,
  run_id,
  seed_set,
  metric_delta,
  compute_cost,
  receipt_hash
}

Source Refresh Warden Gate

Long autonomous runs must periodically refresh source-of-truth state:

upstream PRs
leaderboard / records
paper notes
previous best local stack
external benchmark rule changes

If source freshness exceeds a threshold:

S_{\mathrm{stale}}>\Theta_{\mathrm{source}}
\Rightarrow
\mathrm{WARDEN\_BLOCK}

Coarsening failed ideas

failed idea
→ residual measured
→ scar written
→ local basin coarsened
→ future fine search downweighted
→ reopened only if new evidence/source refresh changes the boundary

Statistical receipt

A run is not promoted by a single attractive metric. It must carry a statistical receipt:

multi-seed result
noise-floor margin
rule compliance
reproducibility status
run-log hash
patch/config hash

Stack placement

MARKOVJUNIOR_16D_PIST_REWRITE_SHIM
→ AUTO_NANOGPT_AUTONOMOUS_RESEARCH_HARNESS_GATE
→ BJW_GEODESIC_CLEANUP
→ NUVMAP_DELTA_DAG
→ FAMM_SCAR_LEDGER
→ SOURCE_REFRESH_WARDEN_GATE
→ STATISTICAL_RECEIPT

No-drift boundary

This gate does not claim autonomous agents can discover truth alone. It models the surrounding harness needed for useful autonomous search: durable memory, receipts, source refresh, pruning, coarsening, and statistical validation.

Project sentence

The Autonomous Speedrun Harness Gate turns long-running agent research into a FAMM-compatible search machine: Builder proposes mutations, Judge validates statistical receipts, Warden blocks stale or invalid loops, failed ideas become coarsened scar basins, and every experiment becomes a Delta-DAG edge with replayable provenance.