Spectral LedgerCondensed phase spectrum

Audited release telemetry / frozen snapshot

What the research pipeline actually ran.

This ledger separates native agents from isolated review sessions and external reviewer calls. Missing billing data stays missing; it is never converted into a fictional zero.

Released

Frozen at release completion

commit 503ead5
34

Codex sessions

10 native agent threads + 24 isolated review runs

78

Model invocations

All recorded Codex, AGY, and Claude executions

3 h 31 m 12 s

Approved runtime

Approval to release-completion timestamp

1,813,129,602

Tracked Codex tokens

Gross logged total, dominated by cached context

Not recorded

Monetary spend

No billing ledger was retained for this run

Execution accounting

34 sessions. 78 model calls.

“Agent” can mean different things. The native Codex team contained 10 threads; the broader Codex execution record contains 34 sessions. Adding the 44 external review calls yields 78 recorded invocations.

Execution groupRunsRoleDisposition
Native Codex agent threads10Authoring, synthesis, implementation, and auditsRelease record
Isolated Codex review sessions24Early schema-bound reviewer attemptsSuperseded by Claude reviews
AGY review rounds20Scientific peer review of five papers and synthesisCanonical paper-review record
Claude review rounds24Haskell, Lean, and website code reviewCanonical code-review record
Recorded model invocations78Snapshot total across all four groups

Model registry

Models and effort levels

Counts are recorded executions, not peak concurrency. Early Claude artifacts did not retain an exact model alias, so the registry reports Opus/high only where it was explicitly pinned.

SystemModelEffortRunsCoverage
Codex native teamgpt-5.6-solxhigh, then Ultra10Native authoring and audit threads
Codex isolated reviewgpt-5.5xhigh24Superseded reviewer attempts
AGYGemini 3.1 ProHigh20Paper and synthesis review rounds
Claude CLIClaudeOpus / high where explicitly pinned24Code and website review rounds; some early artifacts omit the exact alias

Token accounting

Gross traffic, with cache made visible

Codex token events record repeated context as input. That is why the gross figure is large: 97.75% of input was cached. AGY and Claude token usage was not retained and is outside this total.

Gross Codex total
1,813,129,602
Input
1,808,004,007
Cached input
1,767,338,496
Uncached input
40,665,511
Output
5,125,595
Reasoning output
1,954,724 subset of output
Native Codex sessions
1,791,555,257
Isolated Codex sessions
21,574,345

Release loop

Code → review → fix → verify → deploy

Each stage advances only after its release evidence exists. Review and verification remain separate: a reviewer can find an issue, while a build gate confirms that the applied correction survives the corpus.

  1. 01

    Code

    Five papers plus synthesis, six Haskell packages, and a Lean library representation produced.

    complete
  2. 02

    Review

    20 AGY scientific rounds and 24 Claude code-review rounds recorded.

    complete
  3. 03

    Fix

    Valid mathematical, interface, accessibility, and release findings resolved.

    complete
  4. 04

    Verify

    Paper, Haskell, Lean, static export, SEO, and responsive gates passed.

    complete
  5. 05

    Deploy

    Public repository and zero-client-JavaScript website released.

    complete

Coverage and provenance

What this snapshot can support

Counts were reconstructed from Codex session ledgers, committed AGY and Claude review-round files, and Git timestamps. The snapshot ends at the release-completion record, commit 503ead5.

Approved window
→
Request to release
3 h 38 m 21 s
External token coverage
Not recorded
Monetary spend
Not recorded; no estimate substituted