The 25× reduction mattered because the Agent still investigated the real failure and preserved the right to say “not enough evidence to edit.”
P4 solved the wrong bottleneck by brute force
Harness P4 was an important step.
The caller stopped preselecting source slices and evaluator gates. The Agent itself chose what repository evidence to inspect.
But the successful trajectory admitted 203 observations over 204 rounds.
That proved ownership of discovery.
It did not prove good discovery.
caller chooses every source slice — rejected
Agent scans almost everything — works, but weak
Agent selects high-value observations — P5 questionThe obvious optimization was also wrong
If 203 observations are expensive, the simplest metric is obvious: use fewer.
P5 ran that negative control.
The Agent saw one mechanical repository map, inspected zero source observations, and abstained.
Observation count dropped from 203 to 1.
The system got worse.
Minimizing observation count rewarded ignorance.
There was already exact external evidence that a World run had hit a no-Tool Harness failure. The one-observation Agent never inspected it.
So P5 changed the objective:
Highest decision value per observation, not fewest observations.
External failure evidence came first
The accepted P5 branch began with two kinds of information:
- a compact mechanical repository addressability map;
- exact independent Host/World evidence describing the observed failure.
The source target was not preselected.
The source slice was not preselected.
The evaluator gate was not preselected.
The Agent had to decide what to inspect next.
Eight observations were enough to investigate
The accepted trajectory used eight admitted observations:
Six of those were real source/reference investigations.
Compared with the P4 baseline:
203 observations → 8 observations
reduction factor = 25.375×The Agent was no longer traversing the repository because the repository existed.
Each observation had to answer a live discriminator.
The Working View got smaller without deleting history
P5 also exposed a practical context problem.
A full AST-style repository index plus duplicated recent exact observations exceeded the model-view boundary before Provider dispatch.
The repair did not delete the durable repository map or evidence history.
It separated:
durable exact history
from
current model-visible Working ViewThe full index remained durable. The next model request carried a compact addressability map plus only the exact current observation needed for the next decision.
This is compression as navigation, not forgetting.
The Agent found a guard, not a causal defect
The source investigation reached the deliberate no-Tool guard in SQLiteHarnessAgentBridge and adjacent source/reference evidence.
That looked suspiciously close to the external failure.
But the missing causal chain still mattered.
The Agent could not prove:
- a real invocation path carrying the relevant non-empty Tool observation identities;
- a failing unittest for that exact source path;
- a reproducible baseline defect causally attributable to a reachable source target.
The correlation was real.
Mutation authority was not earned.
So the Agent did nothing to the source
The accepted P5 result was:
status = accepted-abstain
source mutation authority earned = false
target = nullThe source generator was not invoked.
The source materializer was not invoked.
The candidate evaluator was not invoked.
The source promotion decision was not invoked.
The canonical source promotion path was not invoked.
Running those branches “to complete the loop” would have fabricated work.
The real defect was found later by a different line
During independent release acceptance, a separate concurrency stress test exposed a real Provider outcome race.
It was repaired and survived 100/100 official concurrency repetitions.
P5's final acceptance record is explicit about provenance:
The discovery Agent did not find or authorize that defect.
This is important in self-improvement research. If a later human, test suite, or independent acceptance line finds a bug, we cannot retroactively credit the earlier Agent because the stories are adjacent.
Twenty-five times fewer observations did not mean cheap
P5 still made 12 Provider calls and consumed 239,953 total tokens.
Most of that cost came from reconstructing a large fresh one-decision Working View repeatedly.
So another tempting equation is rejected:
fewer observations
≠ fewer Provider calls
≠ fewer tokens
≠ cheaper cognitionP6 later targeted that different bottleneck.
What P5 actually earned
P5 did not earn a repository RAG service, Context scheduler, goal stream, generic evaluator service, DAG, RSI controller, or second persistence layer.
It earned something smaller:
- external failure evidence should be addressable before broad source traversal;
- the Agent can own subsequent repository observation choices;
- durable history can remain exact while the next Working View stays compact;
- failed or rejected discovery actions need not be replayed into cognition;
- high-information discovery must preserve bounded abstention.
The improvement was not that the Agent looked at less code. It was that every new look had to buy information—and when the evidence stopped buying causality, the Agent stopped spending authority.
Primary record