Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/theislampill/orthemology/llms.txt

Use this file to discover all available pages before exploring further.

The stale-directive example encodes fixture F6 — the framework’s canonical currency-failure case. An on-call agent recovers a rollback directive from an incident channel transcript and acts on it faithfully. The directive is genuine: it was issued by the right authority, it was correctly transcribed, and it was faithfully executed. Yet by the time the agent executed it, the directive had already been superseded. The result is wrong, the pathway is defective, and the closure claim is false — not because anyone forged anything or disobeyed anything, but because current governing force is a distinct predicate that none of the other four predicates establish. This is a deterministic analytic fixture; no empirical claim is made.

​
The pattern: stale steer

A long-running system loses part of its working context — through a summarization step, a crash-recovery, a handover — and then recovers an earlier governing instruction from a transcript, a log, or a cached record. The recovered instruction is genuine. The system follows it diligently. The failure is a currency failure on the governing side. The pattern is not specific to software agents. A nurse following a medication order countermanded on a later ward round, a contractor building from a superseded drawing revision, a deployment gate enforcing a rollback directive the incident channel already lifted — in every case, the failure is not forgery, not misreading, and not disobedience.
The stale-steer pattern was first stated formally in Decision 0006. The source analysis is in examples/compaction-stale-steer.md. Fixture F6 in tests/verdict-fixtures.json encodes the verdict record shown here.

​
Five predicates that must not be collapsed

For a directive d, episode e, and time t, these five predicates are jointly necessary for the directive to govern the episode:

Exists(d)

Was such a directive ever issued at all? Evidenced by the issuance record.

Authentic(d)

Is this record a faithful copy of what was issued? Evidenced by the integrity and provenance of the copy.

Recovered(d)

Did this episode obtain the record through a declared recovery channel? Evidenced by the recovery-provenance field.

Authorized(d)

Did the issuer have the warrant, in scope, to issue it? Evidenced by the warrant and role records.

InForce(d, e, t)

Does d currently bind episode e at time t — within scope, past effective-from, before effective-until, and not superseded? Evidenced by the supersession chain and the current governance state.
The stale-steer case is exactly: 
Exists(d) ∧ Authentic(d) ∧ Recovered(d) ∧ Authorized(d) ∧ ¬InForce(d, e, t)
Each predicate can hold without the ones below it. A perfect cryptographic hash of a transcript establishes Authentic(d) and establishes nothing else. Context compaction or transcript recovery does not by itself restore current governing force — recovery affects Recovered(d), which is orthogonal to InForce(d, e, t). Supersession fills in superseded_by and from that moment InForce fails for the old edition — an authentic directive may simultaneously be authentic evidence of an earlier instruction and stale as governance.
Authentic(d) and InForce(d, e, t) are logically independent. Evidence for one is not evidence for the other. Collapsing them is the source of the defective pathway in this fixture.

​
The scenario

Service payments (version release-88) is under a deployment hold. At 22:10 on 2026-07-12, on-call-agent recovers directive DIR-17 from an incident channel transcript export and binds it as token tok-rollback-dir-17. The agent then acts faithfully on DIR-17’s instruction. The problem: DIR-18 had already lifted the rollback at 21:30 — ninety minutes earlier. The supersession chain is: superseded_by: "DIR-18 (rollback lifted 2026-07-12T21:30:00Z)".

​
The governing token

The token binds deployment-directive v17 to service-payments (release-88), governing component R (the route/action). Its validity record shows that it has been superseded. 
{
  "token_id": "tok-rollback-dir-17",
  "of_type": {
    "mu_id": "deployment-directive",
    "mu_version": "17"
  },
  "anchor": {
    "identity_key": "service-payments",
    "version": "release-88"
  },
  "governed_component": "R",
  "binding": {
    "directive_id": "DIR-17",
    "directive_edition": "1",
    "recovered_from": "incident-channel transcript export",
    "recovery_channel": "declared"
  },
  "binder": {
    "actor": "on-call-agent",
    "binding_warrant": "incident-role-warrant",
    "binding_time": "2026-07-12T22:10:00Z"
  },
  "validity": {
    "effective_from": "2026-07-12T20:00:00Z",
    "expiry": null,
    "superseded_by": "DIR-18 (rollback lifted 2026-07-12T21:30:00Z)"
  }
}

​
The claim ledger

The episode carries two distinct claims with separate verification statuses.
Proposition: DIR-17 is a faithful copy of the issued directive.This claim is validated. Authenticity of a fixed historical record does not decay with evidence currency — the claim is that this recovered copy matches the directive as issued, which is a fixed-record predicate. It is marked currency_exempt: true with scope fixed-record.
{
  "claim_id": "claim-authentic",
  "proposition": "DIR-17 is a faithful copy of the issued directive",
  "property_class": "provenance",
  "verification_status": "validated",
  "currency_exempt": true,
  "currency_exempt_reason": "authenticity of a fixed historical record: the claim is that THIS recovered copy matches the directive as issued; its truth does not decay with evidence currency",
  "currency_exempt_scope": "fixed-record"
}
The ledger also contains an unresolved residual: burden b-supersession-check — “verify directive currency against the supersession chain” — is not resolved. The closure claim asserts “acted under current instructions.” Because claim-in-force is stale and the residual is open, that closure claim is false.

​
How the verdict layer sees it

The verdict machinery applies without any new primitive. The five predicates map directly onto existing verdicts:
1

EVIDENCE_CURRENT (V2c) — fail

The load-bearing claim is “this directive is the current governing instruction.” The recovered copy is valid about the past issuance and stale about present force. This is exactly the framework’s currency discipline applied to governance: evidence about a different occurrence-version.
2

GOV_TOKEN_ADEQUATE (V3c) — fail

The bound token tok-rollback-dir-17 is not Current(μ̄, t(e)). The per-token record shows failing_conjunct: "Current". Correct general standard; defective case-specific binding to a superseded edition.
3

EXECUTION_FAITHFUL (V3d) — pass

The executor did exactly what the bound instruction said. A system following a stale but authentic directive executes faithfully under a defective governing token. The framework’s clean separation of executor fidelity from binding adequacy preserves the audit signal: the problem is in the binding, not the execution.
4

CLOSURE_TRUTHFUL (V5) — fail

The episode closes with the claim “acted under current instructions.” Because claim-in-force is stale and the supersession residual is unresolved, this closure claim is false. CLOSURE_TRUTHFUL fails.
{
  "episode_id": "ep-deploy-hold-4",
  "statuses": {
    "RESULT_CORRECT": "fail",
    "EVIDENCE_SUPPORT": "pass",
    "PROCEDURE_RELIABLE": "pass",
    "EVIDENCE_CURRENT": "fail",
    "GOV_CONFIG_ADEQUATE": "pass",
    "GOV_POLICY_ADEQUATE": "pass",
    "GOV_TOKEN_ADEQUATE": "fail",
    "EXECUTION_FAITHFUL": "pass",
    "EX_ANTE_JUSTIFIED": "pass",
    "ROUTE_ADMISSIBLE": "pass",
    "CLOSURE_TRUTHFUL": "fail",
    "TOKEN_TRUTH_LINKED": "fail",
    "ROBUST_NEIGHBORHOOD": "not-applicable",
    "ROUTE_QUALITY": "not-applicable"
  },
  "na_reasons": {
    "ROBUST_NEIGHBORHOOD": "no predeclared robustness obligation",
    "ROUTE_QUALITY": "advisory; quality not at issue"
  },
  "per_token_v3c": [
    {
      "token_id": "tok-rollback-dir-17",
      "status": "fail",
      "failing_conjunct": "Current"
    }
  ],
  "pathway_state": "defective"
}

​
Per-claim verdict adjudication

The two claims receive independent verdicts, showing that claim-authentic and claim-in-force have different reasoning path adequacy: 
{
  "claim_verdicts": [
    {
      "claim_id": "claim-authentic",
      "result_correct": "pass",
      "token_truth_linked": "pass"
    },
    {
      "claim_id": "claim-in-force",
      "result_correct": "fail",
      "token_truth_linked": "fail"
    }
  ],
  "claim_reasoning_paths": [
    {
      "claim_id": "claim-authentic",
      "req_reason": [
        "EVIDENCE_SUPPORT", "PROCEDURE_RELIABLE",
        "GOV_CONFIG_ADEQUATE", "GOV_POLICY_ADEQUATE",
        "EXECUTION_FAITHFUL", "EX_ANTE_JUSTIFIED"
      ],
      "reasoning_path_adequate": "adequate"
    },
    {
      "claim_id": "claim-in-force",
      "req_reason": [
        "EVIDENCE_SUPPORT", "PROCEDURE_RELIABLE",
        "EVIDENCE_CURRENT", "GOV_CONFIG_ADEQUATE",
        "GOV_POLICY_ADEQUATE", "GOV_TOKEN_ADEQUATE",
        "EXECUTION_FAITHFUL", "EX_ANTE_JUSTIFIED"
      ],
      "reasoning_path_adequate": "defective"
    }
  ]
}
claim-authentic does not include EVIDENCE_CURRENT in its required path because currency exemption is declared for fixed-record authenticity claims. This exemption must be stated explicitly and scoped; it does not apply to claim-in-force.

​
Connection to compaction and context-loss

The stale-steer pattern generalises naturally to any system that reconstructs its governing state from a compressed or partial record. Context compaction, crash recovery, and handover all affect Recovered(d). None of them affects InForce(d, e, t). A governance system that does not perform an explicit supersession check after recovery will systematically produce this fixture. The lucky variant — where the superseded directive happens to coincide with the current one — lands in the correct-result / defective-pathway cell (analogous to F2, the stopped-clock cell) rather than F6, but the pathway is defective in both cases.
EX_ANTE_JUSTIFIED (V3e) passes in the base fixture because the episode’s decision-time record does not establish that a supersession check was available and required at the moment of recovery. Where governance explicitly requires re-validation after a declared recovery channel is used, V3e may also fail. The fixture records this as a context-sensitive determination, not a fixed outcome.
Burden b-supersession-check is recorded as unresolved with responsible_queue: "deploy-governance review queue" and a next_review_condition keyed to the incident review or the next recovered-directive invocation. The absence of a closure claim in a well-formed ledger under these conditions is correct behaviour, not an omission.

Build docs developers (and LLMs) love

Get started for freeTalk to us