docs/EVALUATION_PROTOCOL.md

# Joint representation selection versus separate repair: proposed protocol

Status: an executable evaluation design for future work, not a completed benchmark
or evidence of an intelligence advantage. The current exact-family tests tie strong
conventional comparators.

## Primary comparison

Compare two systems with identical proposal language, initial state, task evidence,
candidate operator pool, quotient compiler, coding methods, acquisition channels,
evaluation data and total budgets:

- **Joint selection:** choose an abstraction with its predicted future memory and
  regeneration costs in the admission objective.
- **Separate selection and repair:** use a tuned abstraction learner, then optimize
  memory and repair with the same tools. Permit lookahead and cross-validation;
  the baseline must not be artificially denied relevant cost information.

Include pooled learning, a conventional observable-quotient compiler, and exhaustive
bundle search on small cases as reference comparators. Match either the actual
compute budget or a calibrated hardware-normalized cost; live operator capacity
alone is insufficient. Report both selection and deployment costs.

## Task design

Use at least three independently generated families, such as supplied linear
operators, compositional symbolic tasks, and a trained-model continual-learning
benchmark. For each, predefine available sensors, allowed proposal language,
damage distribution, protected queries and deployment tasks. Outside the exact
backend, provide an explicit approximation contract instead of claiming exactness.

Freeze evaluation splits before selecting methods. Keep new task families and
later capability additions unavailable during proposal selection. Independently
sample families for uncertainty estimates; many words from one model are correlated
observations, not independent trials. The five old seed labels are integer RNG seeds.

## Resource ledger

Record model inference/training, proposal count, rejected candidates, compiler and
closure work, sensing calls, retries, search coordination, coding/decoding, memory
bytes, metadata, wall time, hardware and deployment cost. Charge offline data and
teacher/oracle access. Preserve failed runs in the ledger.

Before frozen-offspring evaluation, stop exploratory workers and remove histories
not included in persistent state. Persist every surviving artifact and its byte
count. Give each offspring the same deployment budget on unseen tasks. Do not call
an exploration model as an uncharged fallback.

## Outcomes and ablations

| Outcome | Definition | Interpretation |
|---|---|---|
| Durable utility | Preregistered held-out task score after capability additions and damage | Retained usable behavior |
| Consolidation cost | Total charged cost to produce the frozen offspring | Benefit must account for overhead |
| Next-discovery productivity | Improvement in held-out score divided by fully charged discovery cost, averaged over independent families | Candidate evidence of recursive acceleration |
| Abstention calibration | False exact certificates, justified ambiguity, inconsistent input handling | Reconstruction reliability |
| Storage / latency | Full persistent bytes and recovery/deployment time | Practical tradeoff |

Remove joint selection, tether metadata, mixed-operation synthesis, and regenerative
coding separately while sharing all remaining tools. Include a baseline using the
discovered primitive to distinguish invention value from a privileged implementation.
Preregister how cost-zero cases and zero or negative improvements are handled.

Use paired comparisons at the independent-family level, intervals for effect size,
and all preregistered seeds. Tune both methods with equal resources. Report null
and negative results, sensitivity to budget, and the tradeoff between performance,
storage and recovery time. Select a minimum practical effect before observing data.

## Claim gates

1. Correct finite execution establishes the implementation contract.
2. A replicated equal-resource improvement supports an enabling architecture claim.
3. Transferable mixed-operation gains support synthesis if removal ablations explain
   the gain and strong shared-tool baselines fail to match it.
4. Repeated increases in next-discovery productivity across independent families
   support recursive acceleration. Deployment score increases alone do not.

If the strongest controlled baseline continues to tie, report CQW as an architectural
integration with explicit reconstruction obligations. That is an informative outcome.