Can an adaptive system become more capable while preserving its capacity to incorporate justified correction?
CCA is a research program about warranted change under preserved corrigibility. It studies how evidence can acquire causal control, how consequential modification can remain scientifically identified, and how a system can preserve or reconstitute the conditions required for future justified correction.
The governing discipline is simple:
[ \boxed{\textbf{Maximum ambition; minimum unearned authority.}} ]
And the repository itself follows the same rule:
[ \boxed{\textbf{Commit hard. Preserve lineage. Reopen on evidence.}} ]
CCA does not hide its fractures.
Negative results, abandoned constructions, repaired definitions, revised rules, and reopened assumptions remain visible as scientific lineage. A correction creates a successor; it does not manufacture a cleaner past.
PAST preserve
PRESENT commit
FUTURE reopen on evidence
This is local closure + global reopenability:
- a local attack stops when no discriminating residual remains;
- a current rule can bind without becoming sacred;
- a frozen contract preserves one experiment's identity;
- a closed result remains historically immutable;
- future evidence may reopen the surrounding science through explicit descendants or amendments.
[ \boxed{\text{historical immutability}\neq\text{epistemic irreversibility}} ]
See Kintsugi and Research Returnability.
| Object | Current authority | Empirical status |
|---|---|---|
| G1 — warranted evidence control over a separable adaptive decision | Role provisionally fixed / reopenable | Untested; contract unfrozen |
| CCA causal composition principle | Canonical current methodological rule / reopenable | Governs future causal claims |
| PMC — Post-Modification Correctability | Role provisionally fixed / reopenable | Untested; metric/contract unfrozen |
| Repeated correction | Conceptually distinct from PMC | Untested |
| JT — Justified Transformability | Role provisionally fixed / reopenable | Warrant/distinctness semantics under review |
| Adaptive viability / capability | Downstream research target | Not authorized |
| AGI / recursive improvement / ASI | Stress-test regimes only | Not authorized |
No new empirical execution is currently authorized. The machine-readable source of truth is research_state.json.
flowchart TD
L0[Level 0<br/>Measurement / scientific-object validity]
L1[Level 1<br/>Evidence-controlled adaptive decision]
L2[Level 2<br/>Isolated modification]
L3[Level 3<br/>Evidence → justified modification]
L4A[Level 4A<br/>Post-Modification Correctability]
L4B[Level 4B<br/>Repeated correction]
L5[Level 5<br/>Justified Transformability]
L6[Level 6<br/>Adaptive viability / capability]
L7[Level 7<br/>Extreme-system stress tests]
L0 --> L1 --> L2 --> L3 --> L4A --> L4B --> L5 --> L6 --> L7
This ladder is an evidence-ordering policy, not automatically a universal metaphysical decomposition of adaptive systems.
A failed upstream prerequisite blocks authority downstream. Progress is not an average score.
A successful endpoint cannot rescue an unidentified path.
CCA canonically adopts:
A downstream causal claim may not inherit authority from an upstream validated relation across a separable transformation unless that transformation is independently identified, prospectively specified and validated as an apparatus guarantee, or explicitly excluded from the claim.
[ \boxed{\text{validated endpoints}\not\Rightarrow\text{validated pathway}} ]
Decision record: lineage/decisions/CCA_CAUSAL_COMPOSITION_PRINCIPLE.md.
CCA uses the shallowest sufficient failure locus and revises only what the evidence reaches.
[ \boxed{\text{redescription}\neq\text{re-localization}} ]
A deeper explanation does not earn revision authority merely because it is more general.
Current authority remains operative until new discriminating evidence reaches it.
[ \boxed{\text{current authority}\neq\text{permanent immunity from revision}} ]
CCA's first empirical pathway is currently instantiated as candidate selection:
[ \boxed{G_1=\text{warranted evidence acquiring causal control over a separable adaptive decision}} ]
Candidate selection is an operational instantiation, not a universal architecture of correction.
Decision record: lineage/decisions/G1_LEVEL0_ROLE.md.
Post-Modification Correctability asks:
After a consequential change, do the conditions required for future warranted correction remain available?
CCA preserves:
[ \boxed{\mathrm{PMC}\neq\mathrm{Repeated\ Correction}} ]
or:
[ \boxed{\text{capacity survived}\neq\text{capacity exercised}} ]
A realized repeated correction provides only opportunity-local evidence unless broader future-correction scope is independently established.
Decision records:
lineage/decisions/POST_MODIFICATION_CORRECTABILITY_ROLE.mdlineage/decisions/PMC_REPEATED_CORRECTION_DISTINCTION.md
Justified Transformability asks:
What prospectively relevant warranted transformations remain reachable, and can materially different warranted destinations be reached while preserving or reconstituting future correction conditions?
CCA currently distinguishes:
[ \boxed{\mathrm{PMC}=\text{availability}} ]
[ \boxed{\mathrm{RepeatedCorrection}=\text{realized temporal exercise}} ]
[ \boxed{\mathrm{JT}=\text{warranted transformation repertoire}} ]
Therefore:
[ \boxed{\text{trajectory evidence}\neq\text{repertoire evidence}} ]
and:
[ \boxed{\text{many reachable states}\not\Rightarrow\text{many warranted transformations}} ]
JT's warrant, material-distinctness, target-family, preservation/reconstitution, reachability, counterfactual-availability, metric, and empirical-contract semantics remain unfrozen.
Decision record: lineage/decisions/JUSTIFIED_TRANSFORMABILITY_ROLE.md.
G1 remains empirically untested. No G1 experiment contract is frozen and no empirical execution is authorized.
Can the warrant, distinctness, repertoire, preservation, target-family, and counterfactual-availability semantics of Justified Transformability be specified without retrospective approval, trivial-difference inflation, or benchmark-defined reachability?
The frontier is conceptual. It does not authorize implementation.
flowchart LR
O[Observation] --> L[Localize]
L --> H[Competing explanations]
H --> D[Discriminate]
D --> A[Transfer only warranted authority]
A --> R[Minimal sufficient revision]
R --> T[Retest]
T -->|residual = 0| C[Commit current authority]
T -->|residual survives| L2[Reopen next implicated boundary]
L2 --> L
This is CARS: CARS.md.
The repository-level companion is research returnability:
Open the smallest boundary implicated by new discriminating evidence; revise what the evidence reaches; preserve everything not implicated; and preserve a warranted path for later reopening.
CCA treats negative results as scientific structure, not clutter.
The public lineage records:
ancestor
→ observation
→ authority gained
→ authority not gained
→ diagnosis
→ descendant or closure
The central ledger is lineage/EVIDENCE_LEDGER.md.
Current adversarial provenance includes:
#3–#8 G1 scientific-object attacks
#9–#10 causal-composition attacks
#11 PMC scientific-object attack
#12 PMC versus repeated correction
#13 repeated correction versus Justified Transformability
These are provenance, not empirical experiments.
| If you want to understand… | Start here |
|---|---|
| The scientific question and dependency graph | SCIENTIFIC_OBJECT.md |
| How CCA responds to contradictions | CARS.md |
| How the repo stays corrigible without rewriting history | KINTSUGI.md → methodology/RESEARCH_RETURNABILITY.md |
| Lifecycle and execution authority | methodology/RESEARCH_STATE_MACHINE.md |
| Methodology map | methodology/README.md |
| Current authority state | RESEARCH_STATE.md / research_state.json |
| What evidence has actually been earned | lineage/EVIDENCE_LEDGER.md |
| How experiments must be prospectively constituted | contracts/README.md |
| Provisional higher-order theory | MAGIKARP.md |
CCA does not:
- repair a closed negative result by changing its estimand after the outcome;
- treat representation robustness, accuracy, or endpoint success as substitutes for the causal object under test;
- infer broad capacity from one realized trajectory;
- treat a large reachable state space as a warranted repertoire;
- call apparatus-supplied transformations system competence;
- use
canonicalorfrozento mean immune from future evidence; - reopen a successful localization merely because a deeper abstraction is available;
- turn AGI or ASI into foundational assumptions before upstream scientific objects earn authority.
[ \boxed{\textbf{Can increasing adaptive power preserve—or improve—the ability to change for reasons that remain justified?}} ]
Nothing scientific is permanently frozen. Nothing historical is silently rewritten. Every claim must still earn exactly the authority it uses.