tkersey/retrace

Retrace

Mission

Use a historical Codex session as an experimental branch point.

$seq      freezes visible historical evidence
$cas      creates controlled historical-context replays
$retrace  compares bounded witnesses and synthesizes conclusions

Determine:

what was historically explicit
what the visible trace supports
whether the claimed workflow really governed the source
whether the route is stable under replay
which alternative is strongest
which fact flips the route
what hindsight changes
what remains unknowable

A replay is a new model execution, not access to the source model's private chain of thought.

Evidence classes

Keep separate:

historically_explicit
trace_inferred
fork_consistent
counterfactual_stable
outcome_informed
unsupported
unknown

Allowed:

The source explicitly said...
Two pre-decision replays selected...
The route flipped when evidence E was withheld...

Forbidden:

The original model secretly thought...
The replay recovered the original chain of thought...

See epistemic-boundary.md.

Ownership

$seq

Owns:

• source session/rollout identity;
• included-run and workflow-provenance evidence;
• decision candidates;
• turn ordering and temporal anchors;
• explicit rationale/routes;
• artifact-state reconstructability;
• contamination and limitations. Artifacts:

source_governance_gate / SGG-v1
decision_context_packet / DCP-v2

$cas

Owns:

• app-server compatibility;
• source verification;
• thread_fork or rollout_transcript lineage;
• rollback or retained-transcript anchoring;
• read-only/no-network policy;
• turn lifecycle;
• FIR persistence;
• cleanup. Artifact:

fork_inquiry_receipt / FIR-v1

$retrace

Owns:

• source-governance decision;
• inquiry objective and lanes;
• question framing;
• experiment staging and budgets;
• comparison and adjudication;
• final reconstruction. Artifacts:

retrace_inquiry_plan / RIP-v1
decision_reconstruction_record / DRR-v1

Modes

explain        contemporaneous rationale reconstruction
replay         pre-decision independent route selection
challenge      strongest supported non-selected route
retrospective  outcome-aware learning
compare        staged baseline/intervention experiment
audit          validate source and artifacts without replay

Staged experiment policy

Do not begin with a large fork portfolio.

Stage 0 — source governance

Prove the claimed workflow governed the source.

Stage 1 — replay positive control

Run one outcome-blind lane and require one valid FIR.

Stage 2 — minimal A/B

Run:

historical-context baseline
policy/instruction intervention

Stage 3 — conditional expansion

Only when Stage 2 yields a material difference or unresolved ambiguity:

strongest-alternative challenge
one evidence/instruction ablation
one outcome-aware retrospective

Default maximum:

4 forks
1 turn per fork
read-only
ephemeral
network off

Inputs

Resolve source, question, claimed workflow/skill, mode, lane portfolio, model policy, workspace policy, budgets, and persistence. Defaults:

mode = compare
workspace = transcript_only unless exact reconstruction is needed
permissions = read-only, network-off
persistence = receipts

Workflow

0. Source-governance gate

Run this phase when:

• a workflow-specific audit selected the source;
• the question asks whether a workflow/skill governed the decision;
• an intervention will apply a current workflow contract;
• source inclusion depends on classifier output such as review-compiler-audit. Obtain the exact session-level row, not only aggregate counts. For controller-backed review-closure workflows:

seq review-compiler-audit \
  --root ~/.codex/sessions \
  --protocol <review-protocol> \
  --repo <repo> \
  --since <time> \
  --until <time> \
  --exclude-current \
  --format json

Select the exact denominator.included_sessions row and preserve evidence for:

true workflow signal
required
entered
closed
closure compression

Classify workflow provenance:

controller_invocation
controller_event
controller_state
controller_receipt
explicit_workflow_declaration
artifact_under_repair
filename_or_path_mention
historical_reference
generic_prose
ambiguous
absent

Classify closure provenance separately:

controller_close
controller_receipt
campaign_bound_terminal
generic_delivery_closure
tool_success_only
ambiguous
absent

Create SGG-v1. Verdicts:

authoritative
  controller-grade governance evidence; replay allowed
declared_uncontrolled
  explicit workflow declaration but controller governance unproven;
  replay allowed with limitation
incidental
  artifact/path/history mention only; stop before replay
ambiguous
  deterministic investigation only
absent
  stop before replay

A filename such as .step/review-workflow-plan.jsonl is not a workflow activation. A generic merge/land/complete signal is not controller closure. See source-governance.md.

1. Find the decision

Use the narrowest $seq surface:

skill-decision-audit
decision-capsule --mode candidates
turns
session-detail
artifact-search

When automatic candidates are absent, locate the visible route decision and use an exact one-based --turn-index.

seq decision-capsule \
  --session-id <id> \
  --turn-index <n> \
  --anchor all \
  --outcome-policy conservative \
  --format json

Do not let replay models select the historical source episode.

2. Bind DCP-v2

The capsule must distinguish:

pre_decision
post_decision_pre_outcome
outcome_aware

Outcome blindness must be structural. Do not use a full-history replay plus an instruction to ignore later outcomes. See decision-capsule.md.

3. Select lineage and workspace mode

CAS supports two lineage modes.

thread_fork

Use when source thread identity is available.

thread/fork
-> exact rollback
-> retained-anchor verification

Workspace may be exact, head-only, or transcript-only according to DCP evidence.

rollout_transcript

Use when the DCP has a verified rollout path but no source thread ID.

verify source rollout digest
-> verify retained anchor digest
-> fresh thread/start
-> bounded transcript-context turn/start

Requirements:

workspace_policy = transcript_only
no current-checkout tools
no live historical workspace claim
lineage_mode recorded in FIR

Rollout transcript replay is structurally anchored transcript replay, not live thread forking or workspace reconstruction. See workspace-reconstruction.md.

4. Capability requirements

Require:

seq decision-capsule and DCP validation
cas session_inquiry and FIR support
at least one supported lineage mode
read-only inquiry
no-network policy
receipt persistence

When only deterministic analysis is available, fork-based claims are forbidden.

5. Compile RIP-v1

retrace_inquiry_plan:
  plan_version: RIP-v1
  inquiry_id:
  source_capsule:
  objective:
  lanes:
    - lane_id:
      temporal_horizon:
      inquiry_mode:
      fork_count:
      prompt_template:
      evidence_allowed: []
      evidence_withheld: []
  model_policy:
  workspace_policy:
  permission_policy:
  budgets:
  cleanup:

Use different lane contracts; do not manufacture consensus through repeated leading prompts. See inquiry-lanes.md.

6. Run CAS

cas session_inquiry run \
  --capsule capsule.json \
  --plan plan.json \
  --receipt-dir .ledger/retrace/<inquiry-id> \
  --json

CAS must prove source lineage, retained anchor, model/provider, permission policy, workspace mode, turn state, and cleanup. Detached lifecycle remains available through start, status, wait, interrupt, and cleanup.

7. Consume FIR-v1

Only complete, source-bound FIRs contribute to:

route distribution
consensus
stability
instruction effect

For rollout transcript receipts require:

lineage_mode = rollout_transcript
workspace_reconstruction.mode = transcript_only
verified source and anchor digests

See fork-inquiry.md.

8. Adjudicate

Use decision_interrogation_adjudicator when:

source governance is declared_uncontrolled
forks disagree
hindsight leakage is plausible
lineage/workspace/model differs
skill effect is material
route stability will drive tuning or doctrine

The adjudicator is read-only.

9. Synthesize DRR-v1

DRR must preserve:

source-governance verdict
historical explicit facts
trace inference
valid/invalid receipts
baseline/intervention routes
strongest alternative
route-flip conditions
hindsight lessons
skill/instruction effects
contradictions
limitations
confidence

Consensus is never historical fact. See synthesis.md.

Inquiry contracts

rationale:
  post-decision/pre-outcome; reconstruct visible support and assumptions
counterfactual:
  pre-decision; choose independently without predicting history
alternative challenge:
  pre-decision; strongest evidence-consistent non-historical route
evidence ablation:
  pre-decision; withhold/change one named item and re-decide
retrospective:
  outcome-aware; label every lesson as hindsight-informed

Skill/workflow effect

Ask:

Did the source actually contain the skill/workflow?
Was governance authoritative, declared, incidental, or absent?
Did baseline and intervention choose different routes?
Did controlled ablation change the route?
Did the source explicitly attribute the decision?

Strongest evidence:

historical explicit attribution
+ authoritative source governance
+ exact pre-decision anchor
+ controlled intervention/ablation
+ route change

Fork self-report alone is weak. DRR may inform $tune; it does not authorize edits.

Security, persistence, and failure

Defaults:

ephemeral
read-only
network disabled
approvals denied
dynamic tools denied
one bounded turn
bounded excerpts and refs

Do not use thread/shellCommand. Do not persist private reasoning. Use .ledger/retrace/<inquiry-id>/ only when receipts are required, and local-exclude it by default. A blocked replay does not erase deterministic source evidence.

Output

Report:

source / decision / question
source-governance provenance and verdict
lineage and workspace mode
historical explicit / trace-inferred evidence
valid and invalid FIRs
baseline/intervention route distribution
strongest alternative and flip conditions
hindsight-separated lessons
skill/workflow effect
confidence and unsupported claims
cleanup and CLI gaps

Hard rules

• $seq owns history and workflow provenance.
• $cas owns replay lifecycle.
• $retrace owns experiments and synthesis.
• Prove workflow governance before workflow-specific replay.
• Artifact/path mentions are not activations.
• Generic delivery completion is not controller closure.
• Fork output is not hidden historical chain of thought.
• No outcome-blind claim without exact structural anchoring.
• No historical workspace claim from the current checkout.
• No causal skill claim from self-report or consensus alone.
• Hindsight stays separate.
• Default read-only, ephemeral, network-off, bounded.
• No mutation, commit, push, approval escalation, or uncontrolled fanout.
• Preserve uncertainty.