Explode the Code Recoalescence Value

01

Generation is not progress

A fragment is an option. It becomes capability only after it survives a boundary.

For most of software history, producing a plausible implementation was expensive enough to stand in for progress. Agentic systems break that proxy. A model can cheaply produce five patches, four explanations, three test strategies, and a confident recommendation. The pile is larger. The system may be no better.

The distinction is old even if the production rate is new. David Parnas argued in 1972 that modularity depends on the criteria used to decompose a system, not merely on dividing it into pieces. Carliss Baldwin and Kim Clark later described modular designs as collections of options governed by visible design rules. An option has value because it can be exercised under a contract. A disconnected fragment is not automatically an option; without an interface, evidence, or provenance, it may be only debris.

Explosion creates possibility. Recoalescence decides which possibilities become real.

This is why output counts, lines changed, agents spawned, and even passing tests can mislead. AgentLens examined 2,614 software-agent trajectories and found that 10.7 percent of passing trajectories in its evaluation subset exhibited a "Lucky Pass": regression cycles, blind retries, missing verification, or disordered work. The patch passed; the process still carried risk. At a larger organizational scale, a study of 107,000-plus simulated merges of agentic pull requests reported a 27.67 percent conflict rate. Parallelism produced more possible work and more collision surface at the same time.

02

The destruction ledger

Deleting code can create value. Deleting its history can destroy it.

Destruction is usually discussed as failure: lost work, reverted code, a model answer thrown away. In a high-output system, removal is also maintenance. Duplicate implementations increase search cost. Stale modules impersonate live authority. Bloated context competes with the current instruction. Unsupported clauses contaminate otherwise useful prose. The ability to remove precisely is therefore part of the production system, not its opposite.

But "deleted" hides several economically different acts. Lambspring's operating record contains all four:

These states need a ledger. A minimum destruction record contains the object hash, former live location, removal time, author or agent, prompt and model provenance where relevant, rejection reason, last passing tests, dependencies, replacement, recoverability class, and restoration procedure. The live tree can become smaller without forcing institutional amnesia.

One Lambspring retirement removed a paper-only agent persona from live code and a seed record while keeping its historical logs. A separate archive preserved 76 files totaling 5,461,532 bytes with byte-for-byte verification and restoration instructions. In another cleanup, 6,196 lines of archived v1 code were finally deleted after a replacement service and its contracts were established. The operation was not "keep everything" or "delete everything." It was migration from ambiguous live material to a smaller active surface plus an explicit historical boundary.

This resembles the difference between demolition and deconstruction. The U.S. Environmental Protection Agency describes deconstruction as selective disassembly that can preserve entire assemblies or individual materials for reuse. Whether recovery is rational depends on condition, hazards, labor, and future value. Code has the same distinction: a module can be unfit for the current structure while its test fixture, parser, interface, or hard-won failure case remains valuable elsewhere.

03

Rejection must explain itself

A rejection count is not a quality measure. It is a routing outcome with a missing noun.

Rejected because it is false? Rejected because it duplicates accepted work? Rejected because the repository changed underneath it? Rejected because a human never reviewed it? Those outcomes imply very different things about the fragment, the agent, and its future value.

A 2026 study of 11,048 closed agentic pull requests makes the problem concrete. Among rejected PRs, only 35.7 percent had a clear agent failure; 31.2 percent reflected workflow constraints, and 33.1 percent had no observable rationale. Of merged PRs, 15.4 percent still required reviewer involvement. A binary merged/rejected label therefore confounds technical fitness, process fit, timing, and missing evidence.

We use a two-stage rejection rule. First, hard invariants fail closed. A fragment that violates a safety constraint, names the wrong executable target, lacks required provenance, or cannot produce a required proof is excluded from the live path regardless of how persuasive it sounds. Second, fragments that clear the hard gate receive an auditable vector:

V Verification

How strongly tests, checks, or cited evidence support the fragment.

C Contract fit

Whether it satisfies the current interface and exact requested target.

P Provenance

Whether model, prompt, source, dependencies, and transformations are known.

N Non-redundancy

Whether it adds distinct capability rather than coordination surface.

Each component lies between 0 and 1; higher is better. J therefore rises as the reason to keep a fragment outside the live system grows. Weights must be declared by domain and calibrated against downstream outcomes. They are not natural constants. A high-stakes executable path may assign most weight to verification and contract fit. A research notebook may assign more to provenance and recoverability.

These bands are illustrative. The vector and reason code remain canonical because a scalar destroys information. Every rejected unit should also carry one disposition: invalid, incompatible, superseded, deferred, salvageable, or archived. Reintroduction is allowed only when evidence changes: a new test passes, the contract changes, provenance is repaired, or the duplicate becomes necessary. The old score is not overwritten; a new score is appended under the new boundary.

04

The mathematics of coming back together

Value retained and value created by integration are different terms.

Let an objective be exploded into candidate fragments F = {f₁ ... f_n}. Give each fragment a pre-integration estimate u_i^* in one consistent unit: expected dollars, hours saved, defect reduction, or a declared utility scale. The gross option pool is:

G is not realized value. It is the optimistic pre-integration pool and is expected to be wrong. It creates a denominator against which the cost and selectivity of the process can be inspected.

Rejected material may still contain a future option. For fragments outside the live set S, let ρ_i be recoverability: the probability that provenance, compatibility, and physical availability will permit future reuse. Let u_i,future be expected future utility and c_i,restore restoration cost:

Now evaluate the integrated system itself. U(S) is end-to-end utility after integration; U(B) is the baseline without the run. C_E is the cost of exploding the work, and C_R is the cost of reconciling it:

The multiple R is not bounded between 0 and 1. It can be negative when the swarm creates more coordination cost than utility. It can exceed 1 when interactions among accepted fragments create capability greater than their separate estimates. That interaction is worth measuring directly:

A positive Δ_I is genuine synthesis: two parts become more useful together. A negative value is coordination debt: regressions, duplicated abstractions, interface mismatch, review time, or operational complexity consume the local gains.

05

What the field record says

The observations do not prove the metric. They show why its terms are necessary.

Destructive compression

In June 2026, a local review stage appeared to be a model failure. It produced malformed or non-discriminating outputs and passed too much work onward. A controlled replay tested two models against three payload forms. With the full approximately 11,975-character payload, each model produced 0 valid outputs in six prompts. Removing a duplicated brief improved each to 3/6. A lean approximately 1,899-character contract produced 6/6 for each.

The first diagnosis - replace the model - was rejected. The stronger explanation was payload bloat colliding with a 4,096-token context limit. Deleting roughly 84 percent of the prompt restored the required structure. This is a small internal replay, not a model benchmark, but it demonstrates the mechanism: context can have negative marginal value when it obscures the boundary.

Selective routing

In the instrumented review window, Compass recorded 1,477 local assessments. The local stage classified 616 as WRONG and dropped all 616 before the frontier judge. It classified 861 as GOOD, OUTLIER, or UNCLEAR and escalated every one. Fifteen ultimately surfaced.

Across the longer review history, 1,925 of 2,379 records were dropped and 23 surfaced. These figures are selectivity telemetry, not quality telemetry. "Surfaced" is not truth, and "dropped" is not model failure. The important design choice is directional: confident local wrongness terminates; uncertainty escalates. Unknown labels do not silently become rejection. That asymmetry prevents a cheap gate from becoming an invisible censor.

Salvage below the document level

A later report-take failure exposed a different boundary. One generated take contained a supported sentence followed by an unsupported causal claim. Rejecting the whole take would lose valid material; accepting it would launder the unsupported dependency. The sanitizer now operates on independent units: supported text survives, unsupported numeric or causal clauses are removed, unknown provenance markers reject the whole take, and an empty result fails closed. The change passed 52 focused tests and the then-current 320-test suite.

Process state, not just final state

The operating journals hold 112 imported entries, 309 active knowledge nodes, and 23 explicit tension records in open, held, or resolved states. That count is not value. Its significance is structural: dissent can be preserved without granting it production authority. Agents may speak freely in their own journals; a bridge summarizes current state; contracts govern admission; tests govern execution. Free expression and strict integration are complements.

The distinction also changes prompting. A good production prompt is not the longest account of everything the organization knows. It states the current role, one decision, the exact input contract, allowed evidence, failure behavior, and output schema. Broader memory remains retrievable and attributable outside the hot path. This reduces prompt collision without erasing history.

06

The same shape appears in physical systems

These are structural analogies, not claims that code, buildings, cells, and spacecraft are equivalent.

The common pattern is recoverable modularity under explicit interfaces. A beam without known dimensions, a biological cell without a useful connection, a flight component without pedigree, and a generated function without contract evidence all fail for different reasons. Yet each shows why local existence is weaker than system membership.

07

An operating protocol for agent swarms

Freedom belongs inside the workspace. Authority is earned at the boundary.

1. 01

   Declare the design rules before fan-out

   Name the objective, interfaces, invariants, evidence classes, write scope, and owner. Parallel work without a shared contract is multiplication, not decomposition.

2. 02

   Give every worker an attributable journal

   Store the full prompt, role, model, parent, timestamps, actions, evidence, artifacts, and next step. Preserve uncertainty and dissent in the worker's own voice. Never require private chain-of-thought.

3. 03

   Keep hot prompts narrow

   Load the decision contract and only the evidence needed for that decision. Retrieve broader memory by reference. Measure payload length, truncation, structured-output validity, and retry behavior.

4. 04

   Gate the smallest verifiable unit

   Score clauses, files, migrations, tests, or patches at their actual dependency boundary. Do not accept a whole report because one sentence is cited, or destroy a whole branch because one file is stale.

5. 05

   Reject with a vector and a reason

   Apply hard invariants first, then preserve verification, contract, provenance, and redundancy dimensions. Unknown must remain unknown; it must not be laundered into either acceptance or failure.

6. 06

   Archive before irreversible deletion when the option is plausible

   Hash the artifact, record dependencies and replacement, test restoration, and assign carrying cost. Purge only when recovery value is dominated by risk or maintenance.

7. 07

   Verify the assembled system

   Run unit, integration, end-to-end, rollback, and operational checks. A set of green branches is not a green product.

8. 08

   Measure the right denominator

   Track reconciliation hours, human forensic-triage time per admitted branch, regression rate, dormant recovery, and end-to-end utility. Output volume belongs in the cost column unless it survives.

08

Limits, failure modes, and what would change our mind

Recoalescence value is a proposed accounting frame, not a validated universal metric. Lambspring's evidence is observational, comes from one evolving system, and spans telemetry eras. Its routing labels are operational states rather than ground-truth judgments. The prompt replay had six items. Code deletion and later stability occurred in sequence, but sequence alone does not establish causality.

The framework also invites gaming. A team can inflate G by exaggerating local estimates, improve apparent selectivity by rejecting aggressively, or manufacture positive integration surplus with a weak baseline. The defense is not more algebra. It is preregistered units, external outcome measures, versioned contracts, explicit costs, and calibration against later failures and recoveries.

Evidence about AI productivity already warns against universal claims. METR's early-2025 randomized trial found that 16 experienced open-source developers working in familiar mature repositories took 19 percent longer with AI, even while believing they were faster. METR's 2026 follow-up says newer tools likely provide more speedup, but selection effects make the size unreliable. Both results fit the same principle: capability is conditional on task, workflow, user, contract, and measurement boundary.

The goal is not a high rejection rate. It is a high-fidelity system with low avoidable reconciliation cost and recoverable knowledge. Over-rejection creates blindness. Under-rejection creates contamination. Recoalescence is the disciplined search for the boundary between them.