Skip to content

PR8 Second Adversarial Progression Review v1

Purpose

This review attacks the deterministic PR8 progression projection after the first progression-authority pass. The target is not recommendation quality. The target is whether subtle snapshot, coverage, overlap, temporal, or witness semantics could make an advisory projection look more authoritative or more complete than its inputs justify.

The reviewed baseline remains:

explicit selected governed state dimensions
        +
direct accepted PR1 relations
        +
explicit request-local focus
        +
explicit prerequisite checks
        +
explicit exploration preservation
        ↓
ProgressionFrontier

The projection remains advisory:

COULD BE CONSIDERED != SHOULD DO

Finding 1 — focus/seed collapse

Before this pass, the same exact capability concept could be used both as an explicit frontier seed and as an explicit focus. Because focus creates a FrontierCandidate, a capability already serving as progression basis could reappear as projected frontier output without any semantic change in level, scope, or capability identity.

That collapses two distinct roles:

SEED BASIS != PROJECTED DIRECTION

Repair

During source-backed derivation, exact focus concepts must remain distinct from exact selected seed concepts.

FOCUS == EXACT SEED CONCEPT -> REJECT

This does not infer a latest or preferred state. The caller still explicitly chooses the single state to use as the seed for that concept.

A focus may still overlap a concept that independently appears as a direct adjacent frontier candidate. In that case the candidate truthfully preserves both facts:

explicit_focus = true
adjacency witness exists

but neither fact creates rank or priority.

Finding 2 — repeated-state witness amplification

Before this pass, two different PersonalCapabilityStateId values for the same exact capability concept could both be selected as frontier seeds. The same direct semantic edge could then appear twice as adjacency provenance, differing only by state id.

That creates a presentation-amplification hazard even though PR8 has no numeric score.

Repair

Within one progression request:

ONE EXACT SEED CONCEPT -> AT MOST ONE SEED STATE

This is deliberately not a latest-state rule:

AT MOST ONE != AUTOMATICALLY CHOOSE NEWEST

The caller must explicitly choose which state is the seed.

Distinct seed concepts may still independently produce multiple witnesses for one candidate. Those are genuinely distinct semantic bases and remain inspectable provenance, not ranking.

Partial prerequisite-binding coverage

A candidate can have more than one categorical REQUIRES relation. A caller may supply an explicit PrerequisiteCheckBinding for only a subset.

PR8 must preserve that distinction exactly:

BOUND REQUIRES -> assessed_prerequisites
UNBOUND REQUIRES -> unassessed_prerequisites

A gap is derived only for an assessed relation whose requested dimensions are not represented as supported under the selected state.

Therefore:

NO GAP FOR UNASSESSED PREREQUISITE != SATISFIED
PARTIAL ASSESSMENT != COMPLETE PREREQUISITE COVERAGE

PR8 v1 intentionally has no coverage_complete, all_prerequisites_satisfied, or similar authority-bearing convenience field.

Conflict-bearing supported seeds

PR3 deliberately separates support standing from conflict status. A dimension can be:

standing = SUPPORTED
conflict_status = UNRESOLVED

PR8 permits such a dimension as an explicit seed because the support-content standing is still SUPPORTED. It does not silently resolve, penalize, rank, or suppress the conflict.

The exact state id and selected dimension keys remain in the adjacency witness, and source-backed verification requires the supplied PR3 state snapshot.

SUPPORTED + UNRESOLVED != CONFLICT-FREE SUPPORT
UNRESOLVED CONFLICT != PRIORITY PENALTY
UNRESOLVED CONFLICT != AUTOMATIC SEED REJECTION

Any UI that needs to display the conflict must inspect the exact referenced state basis; PR8 does not duplicate or reinterpret PR3 conflict semantics.

Cross-snapshot semantic substitution

The first adversarial pass added exact source-backed re-derivation verification. That proves that a frontier is consistent with the supplied semantic, epistemic, frame, and state snapshots.

It does not authenticate where those snapshots came from.

Two independently supplied catalogs can contain the same exact concept refs and the same relations while carrying materially different concept text under the same ref. If the changed text does not alter the deterministic inputs used by the frontier algorithm, the same frontier can re-derive against both snapshots.

This is an explicit limit:

VERIFIED AGAINST SUPPLIED SNAPSHOT != AUTHENTICATED SOURCE SNAPSHOT
EXACT CONCEPT REF != CONTENT HASH
CATALOG JSON != SIGNATURE
CONSISTENCY != PROVENANCE AUTHENTICITY

PR8 does not add a local SHA field and call that authenticity. A future durable archive may content-address exact snapshots, and a future governance layer may authenticate publisher/issuer identity, but those are separate contracts.

Historical as_of reconstruction limit

ProgressionFrontier.as_of bounds the selected personal state basis. It does not prove that the supplied capability catalog or competence-frame catalog existed in exactly that form at as_of.

A historical frontier can therefore be consistently reconstructed using a supplied snapshot, but PR8 cannot claim:

this was the exact shared semantic snapshot available at that historical time.

Without an archived authenticated snapshot contract:

HISTORICAL FRONTIER as_of
!=
HISTORICAL CATALOG AUTHENTICITY

The same distinction applies to frame snapshots.

Source-snapshot authenticity boundary of the verifier

validate_progression_frontier_v1(...) is a deterministic consistency verifier. It:

  1. reconstructs the exact stored effective request;
  2. validates selected source records/states against supplied snapshots;
  3. reruns the frozen PR8 derivation;
  4. requires exact equality with the supplied frontier.

It does not:

authenticate catalog publisher
verify a digital signature
prove an archive timestamp
prove global snapshot identity
prove that same-ref content was never substituted

Accordingly:

SOURCE-BACKED VERIFICATION != SOURCE AUTHENTICATION

Additional invariants frozen by this pass

SEED BASIS != PROJECTED DIRECTION
FOCUS == EXACT SEED CONCEPT -> REJECT
ONE EXACT SEED CONCEPT -> AT MOST ONE SEED STATE
AT MOST ONE SEED STATE != LATEST-STATE POLICY

PARTIAL PREREQUISITE ASSESSMENT != COMPLETE COVERAGE
UNASSESSED PREREQUISITE != SATISFIED PREREQUISITE
NO GAP != ALL PREREQUISITES SATISFIED

SUPPORTED + UNRESOLVED != CONFLICT-FREE SUPPORT
CONFLICT STATUS != FRONTIER PRIORITY

MULTIPLE DISTINCT SEMANTIC WITNESSES != HIGHER PRIORITY
REPEATED STATE IDENTITY != ADDITIONAL SEMANTIC SUPPORT

VERIFIER CONSISTENCY != SNAPSHOT AUTHENTICITY
EXACT REF != CONTENT HASH
HISTORICAL as_of != HISTORICAL CATALOG PROOF

Non-goals preserved

This pass does not introduce:

  • latest-state selection;
  • recency weighting;
  • witness scoring;
  • prerequisite-completeness scoring;
  • automatic conflict resolution;
  • catalog or frame signatures;
  • content-addressed persistence;
  • historical semantic archive infrastructure;
  • recommendation ranking;
  • permission or safety inference;
  • history/Legend feedback into progression.

Adversarial regression surface

The executable review is primarily in:

tests/progression/test_progression_second_adversarial_v1.py

It covers:

  • exact focus/seed overlap rejection;
  • repeated seed-state amplification rejection;
  • partial REQUIRES binding with explicit unassessed remainder;
  • SUPPORTED + UNRESOLVED seed semantics;
  • legitimate focus plus adjacency on the same derived candidate;
  • same-ref cross-snapshot semantic substitution as an explicit authenticity limit;
  • historical as_of without false historical-catalog claims.