idkmesh

R4 — Verified Stigmergic Routing

Issue: #97
Status: Synthetic mechanism and frozen reference experiment complete

Research question

Can a local ant-colony/stigmergic memory learn task→worker affinities from verified outcomes, forget stale advantages, adapt after capabilities change, and still let newcomers demonstrate value?

The experiment is intentionally limited to routing. A pheromone trace is not contributor authority, governance power, or evidence that a result is correct.

Reference evidence

The frozen default and lock-in traces, readable comparison, exact generation commands, and artifact digests are published in the R4 reference report. The committed artifact hashes and cross-runtime replay invariants are checked in tests/test_r4_reference.py; byte identity is bound to the recorded Python 3.12 runtime family.

The evidence is deliberately mixed: evaporation plus exploration avoids the permanent-pheromone lock-in failure, but Thompson sampling wins the adversarial lock-in scenario and slightly leads realized success in the default scenario. This is mechanism evidence, not a claim that stigmergy is the preferred router. The issue selection, failed predecessor, and clean current-main continuation are preserved in the issue #97 completion record.

Governing invariant

Activity is not pheromone. Verification is the update gate.

The simulator sends two separate events for every selected worker:

1. activity happened
2. verified outcome arrived

The activity event is deliberately ignored by the stigmergic policy.

Only the verified outcome can change pheromone:

verified success -> positive deposit
verified failure -> optional negative penalty
unverified activity -> exactly zero positive deposit

The result records:

unverified_activity_events
unverified_activity_pheromone_increase
verified_success_deposit_events
verified_success_deposit_total
verified_failure_penalty_events

The required invariant is:

unverified_activity_pheromone_increase == 0

This distinction is essential before any future attempt to map the idea onto GitHub/community activity.

---

1. Pheromone rule

For task class a and worker i, R4 maintains a trace tau[a,i].

At each step, existing pheromone evaporates:

tau <- max(tau_min, (1-rho) * tau)

After verified feedback:

success:
  tau <- tau + success_deposit

failure:
  tau <- max(tau_min, tau - failure_penalty)

The routing weight is approximately:

weight[a,i]
  = tau[a,i]^alpha
  * newcomer_bonus(if untried)

and the policy samples proportionally to these weights unless the explicit exploration floor fires.

With exploration floor epsilon:

with probability epsilon:
  choose uniformly among eligible workers
otherwise:
  sample from pheromone weights

This gives three independently meaningful controls:

• rho — forgetting/evaporation;
• alpha — reinforcement strength;
• epsilon — guaranteed exploration.

---

2. Policies

R4 compares six policies on the exact same replayable task/environment trace.

random

Uniform random eligible worker.

greedy

Choose the highest empirical verified-success posterior mean using a simple Beta(1,1) prior.

This baseline can lock onto incumbents because it does not deliberately explore once an observed worker looks better than an unseen one.

thompson

Per-task-class Thompson sampling with Beta success/failure posteriors.

This is an important conventional stochastic baseline: bio-inspired stigmergy is not assumed to outperform a bandit merely because it resembles nature.

stigmergy-no-evap

• pheromone reinforcement;
• no forgetting;
• no explicit uniform exploration floor.

This is intentionally vulnerable to historical lock-in.

stigmergy-evap

Adds evaporation so old verified evidence gradually loses influence.

stigmergy-evap-explore

Adds both:

• evaporation;
• explicit uniform exploration floor;
• extra routing weight for untried/new worker-task pairs.

This is the main candidate adaptive policy, not a presumed winner.

---

3. Common-random-number outcome design

Policy comparisons should not win because they happened to receive an easier random stream.

R4 therefore generates a task trace once and uses a deterministic hash-derived outcome draw for:

outcome_seed
+ step
+ worker id
+ task class

If two policies choose the same worker on the same task/step, they observe the same synthetic verified outcome.

The task/environment trace is hashed with SHA-256 and every policy result records the same trace digest.

This makes the experiment reproducible while allowing policies to route differently.

---

4. Default environment: specialization, shift, newcomers, churn

The default environment has three task classes:

code
test
security

and several workers with different specialties.

Before the change point:

• code-specialist is strongest on code;
• test-specialist is strongest on tests;
• security-specialist is strongest on security;
• generalist is moderate across classes.

At shift_step, affinities change:

• the old code/test specialists effectively swap strengths;
• the security specialist degrades;
• the generalist remains stable.

Later, two newcomers appear:

• newcomer-strong — genuinely useful, especially for security;
• newcomer-weak — low quality.

The generalist also has a temporary unavailable window to introduce simple churn.

This makes the routing algorithm solve several problems at once:

learn specialization
+ forget stale specialization
+ survive temporary absence
+ give newcomers trials
+ distinguish strong vs weak newcomer

The default environment is a descriptive stress scenario, not a causal one-factor experiment.

---

5. Deliberate lock-in trap

lockin is a separate synthetic scenario designed specifically to produce a harmful reinforcement regime.

Workers:

early-incumbent
steady-backup
late-expert

Before the shift:

early-incumbent success ~= 1.00
steady-backup    success ~= 0.55
late-expert      unavailable

After the shift:

early-incumbent success ~= 0.05
steady-backup    success ~= 0.55
late-expert      success ~= 0.95 and becomes available

A no-evaporation pheromone policy can carry a large historical advantage for the now-bad incumbent, while evaporation + explicit exploration should have a mechanism to discover the late expert and forget the old route.

The test suite requires this fixture to expose harm from no-evaporation stigmergy rather than allowing the bio-inspired mechanism to be reported only in favorable regimes.

---

6. Metrics

Every policy reports raw machine-readable routing metrics.

Verified outcome

• verified successes;
• overall verified-success rate;
• pre-shift verified-success rate;
• post-shift verified-success rate.

Regret / specialization

For every step the simulator knows the synthetic success probability of every eligible worker.

It records:

oracle success probability
selected worker success probability
expected regret = oracle - selected

Aggregate metrics include:

• cumulative expected regret;
• mean expected regret;
• optimal-assignment rate.

The oracle is for measurement only; routing policies do not get to use it.

Adaptation

adaptation_recovery_steps searches the post-shift trace for the first 25-step window whose mean selected/oracle expected-quality ratio reaches at least 90%.

If no such window occurs, recovery is null.

This is a descriptive diagnostic rather than a formal optimal stopping criterion.

Stale-route persistence

For task classes whose best worker changes across the shift, R4 measures the fraction of early post-shift assignments that still go to the old best route.

This helps expose historical lock-in directly.

Concentration

• assignment entropy;
• assignment Herfindahl-Hirschman concentration (assignment_hhi);
• assignment counts;
• longest same-worker streak;
• longest failed same-worker lock-in streak.

High concentration can represent useful specialization or dangerous lock-in; it must be interpreted with success/regret together.

Newcomers

• first assignment step per newcomer;
• newcomer assignment share;
• newcomer assignment count;
• newcomer availability opportunities.

A good policy should not maximize newcomer share blindly. It should provide enough opportunity to discover a strong newcomer without permanently subsidizing a weak newcomer.

Pheromone integrity

For stigmergic policies:

• verified success deposit events/total;
• verified failure penalty events/total;
• unverified activity events;
• unverified activity pheromone increase;
• periodic full pheromone snapshots.

---

7. Pheromone snapshots

R4 stores full machine-readable pheromone state periodically, including at important transition steps:

• start;
• capability shift;
• newcomer arrival;
• end;
• configured snapshot interval.

This allows a contributor to audit whether a route strengthened because of verified success or remained strong only because evidence decayed too slowly.

The raw event trace separately stores per-step routing diagnostics before and after verification when event retention is enabled.

---

8. Why evaporation matters beyond routing

R4 is partly testing a broader IDKMesh principle:

Useful memory should usually have a half-life unless the environment is truly stationary.

Permanent historical advantage can create:

• stale worker routing;
• incumbent contributor lock-in;
• outdated verifier trust;
• obsolete topology links;
• architecture choices that survive long after evidence changes.

But too much evaporation wastes durable evidence and causes repeated rediscovery.

The interesting question is not “should history exist?” It is:

what evidence should persist
for how long
under which change signal
with what exploration floor?

R4 begins with the lowest-risk version of that question: synthetic task routing.

---

9. Why this is not yet a reputation system

Do not directly convert R4 pheromone into contributor authority, merge rights, voting power, resource ownership, or governance weight.

Routing and governance have different threat models.

Community/reputation deployment needs separate experiments for:

• Sybil attacks;
• collusion;
• reciprocal reinforcement;
• spam/activity gaming;
• identity resets;
• incumbent capture;
• newcomer fairness;
• appeals/correction;
• strategic withholding;
• malicious verification;
• transparency/privacy.

R4’s strongest community lesson, if supported, may be architectural rather than numerical:

reward verified durable outcomes
+ decay stale evidence
+ preserve newcomer exploration

not “use this exact pheromone score for people.”

---

10. Run

Default specialization/change/newcomer scenario:

python -m randomness_lab.r4 \
  --scenario default \
  --steps 800 \
  --shift-step 400 \
  --task-seed 42 \
  --outcome-seed 4242 \
  --policy-seed 1337 \
  --output results/r4-default.json

Deliberate lock-in trap:

python -m randomness_lab.r4 \
  --scenario lockin \
  --steps 500 \
  --shift-step 100 \
  --task-seed 11 \
  --outcome-seed 1111 \
  --policy-seed 77 \
  --output results/r4-lockin.json

For smaller result files, omit per-step events while retaining aggregate metrics and pheromone snapshots:

python -m randomness_lab.r4 --scenario default --no-events --output results/r4-summary.json

---

11. Acceptance criteria mapping for #97

The initial implementation targets:

• interchangeable routing policies;
• random/greedy/Thompson baselines;
• no-evaporation stigmergy;
• evaporation stigmergy;
• evaporation + explicit exploration/newcomer bonus;
• reproducible non-stationary environment;
• reproducible newcomer arrival;
• verified-outcome-only pheromone deposits;
• raw routing diagnostics and pheromone snapshots;
• harmful lock-in fixture;
• tests proving activity cannot increase pheromone.

---

12. Next experiments after the first reference run

Factor-isolate evaporation

Sweep:

rho = 0, 0.001, 0.005, 0.01, 0.02, 0.05, 0.10

under identical capability-shift traces.

This can estimate the exploration-memory trade-off rather than comparing only two hand-picked settings.

Factor-isolate exploration

Sweep explicit exploration floor separately from evaporation.

Measure:

• newcomer time-to-discovery;
• weak-newcomer wasted assignments;
• regret;
• routing concentration.

Reinforcement strength

Sweep alpha, success-deposit size, and failure penalty.

Aggressive positive feedback may learn specialization quickly but amplify misleading early success.

Task→verifier routing

Apply the same mechanism to choosing independent verifier bundles by failure class, but only after task→worker routing is understood.

Community path simulation

Create an abstract contributor/task graph with Sybil/collusion attackers before any real community scoring is considered.

---

13. Evidence standard

A useful result is not:

“Ant colony algorithms work in nature, therefore IDKMesh should use them.”

A useful result is a map like:

stationary specialization:
  stigmergy learns quickly

capability shift:
  no evaporation locks in
  moderate evaporation recovers

newcomer arrival:
  zero exploration delays discovery
  small exploration discovers strong newcomer
  excessive exploration wastes work on weak newcomer

high noise / misleading early success:
  Thompson may beat stigmergy

If the conventional bandit wins, that is a successful R4 finding.

The biological analogy proposes a local rule. Reproducible evidence decides whether the rule belongs in IDKMesh.

This site is open source. Improve this page.