The local Copilot ledger behind Commons Collective, grouped by the model identifier actually recorded. Calculations reuse a pinned version of Ethical Tech CoLab's usage-calc.
A snapshot, not a bill or a live usage meter. Includes records before (UTC). Start of the audit-refresh request. Includes earlier audit production and all scoped project work before this cutoff; excludes this refresh and subsequent activity. Other machines and unlogged activity are not established by these records.
10-minute sitting residual; not actual prompting or review time
Author review (workload estimate)
3.46 h
49,444 baseline word exposures; section-version reading estimate, not logged labor. Checking/editing unquantified.
Inferred engaged time
13.26 h
Model-active union plus human-side residual
Output-channel tokens
1,538,231
Priced output channel; not unique accepted work
USD equivalent / not a bill
$888.84
Upstream charge conversion, not an invoice
Summed model request time
11.67 h
Additive work across requests; not elapsed wall time
Start date is the earliest request-interval start inferred from recorded completion time and duration. Last update is this snapshot's capture time, not the last model response. Days covers inclusive UTC dates in the recorded activity span; it does not count human workdays. The last included response was 2026-09-20T19:36:13.463Z.
Model time and interaction-time proxy
The interaction residual is not a timesheet or a productivity estimate. Actual prompting time and actual review time are not measured. The pinned framework groups requests into sittings according to gaps between completion timestamps, starts each sitting at its earliest request start, and unions overlapping sittings. The non-model remainder is an inference; it can include tool waits, interruptions, or unattended automation. It can miss human reading after the last request and human work concurrent with model activity.
13.26 h inferred engaged time = 10.50 h model-active elapsed time + 2.76 h inferred human-side residual, at the 10-minute cutoff.
Idle cutoff
Merged sittings
Engaged union
Model-active union
Inferred human-side residual
2 min
91
12.47 h
10.50 h
1.97 h
5 min
25
13.12 h
10.50 h
2.62 h
10 min (headline)
12
13.26 h
10.50 h
2.76 h
30 min
8
14.26 h
10.50 h
3.75 h
The model-active union stays fixed across cutoffs. Hours are rounded independently; the exact millisecond identity is checked before display. The human residual changes with the engagement assumption; this sensitivity is not a confidence interval or a bound on actual labor. The complete recorded wall-clock span is 50.15 h, including long gaps, and must not be counted as human time. These are pooled intervals for this one project's observed machine, not a sum of per-agent or per-model human hours.
No prompt or response content is needed for this inference. The same pinned usage-calc interval functions compute the sitting and model unions. Date reporting intentionally uses UTC for reproducibility rather than the upstream dashboard's machine-local date display.
By recorded model
Model IDs are taken from the ledger, not inferred from agent names. The identifier does not independently establish the model weights, revision, or hidden provider routing.
Model ID
Requests
Input
Cache read
Cache write
Output
USD equivalent
Summed request time
gpt-6-astra
1,153
3,531
285,416,442
12,933,339
1,532,681
$888.77
11.66 h
gpt-5.4-mini
11
35,200
172,032
0
5,550
$0.06
0.01 h
Token values come from price-bearing channel details. Cached traffic is repeated context, not unique authored material. Recorded reasoning metadata totals 811,105 tokens and is not added again to the 300,098,775 priced-channel tokens.
Displayed model costs round independently to cents, so adding displayed rows can differ by a cent from the rounded total. The downloadable data retains each unrounded decimal and exact charge units.
Main assistant and delegated work
Model ID
Role
Requests
Priced-channel tokens
USD equivalent
Summed request time
gpt-5.4-mini
Delegated agents
11
212,782
$0.06
0.01 h
gpt-6-astra
Main assistant
672
273,351,196
$817.60
8.83 h
gpt-6-astra
Delegated agents
481
26,534,797
$71.17
2.83 h
These are alternative breakdowns of the same total, not additional charges. Agent identities and task text are withheld. Summed request time is 42,016,957 ms (11.67 hours); the overlap-aware union is 37,811,777 ms (10.50 hours. Neither is measured human labor or GPU time. Displayed USD and hours are rounded; reconciliation uses integer charge units and milliseconds.
Channels and observed rates
Recorded channel
Tokens
USD equivalent (unrounded)
Input channel
38,731
$0.07506
Cache read
285,588,474
$511.5862094
Cache write
12,933,339
$279.967775
Output channel
1,538,231
$97.207575
Inspect all observed model/channel rates
Model ID
Channel
USD equivalent / million tokens
Tokens at rate
Requests with rate
gpt-5.4-mini
Input channel
$0.75
35,200
11
gpt-5.4-mini
Cache read
$0.08
172,032
11
gpt-5.4-mini
Cache write
$0.00
0
11
gpt-5.4-mini
Output channel
$4.50
5,550
11
gpt-6-astra
Input channel
$10.00
2,196
732
gpt-6-astra
Input channel
$20.00
1,335
421
gpt-6-astra
Cache read
$1.00
59,259,577
732
gpt-6-astra
Cache read
$2.00
226,156,865
421
gpt-6-astra
Cache write
$12.50
3,469,256
732
gpt-6-astra
Cache write
$25.00
9,464,083
421
gpt-6-astra
Output channel
$50.00
710,739
732
gpt-6-astra
Output channel
$75.00
821,942
421
Rates can vary within one model ID. No reason for the variation is inferred. A request can appear in several channel rows; do not add their request counts.
Recorded token charges converted using usage-calc's AI-credit assumption; a list-price equivalent, not subscription spending or an invoice. The conversion assumes 1,000,000,000 nano-AIU per credit and USD 0.01 per credit. 88883661940000 recorded nano-AIU therefore converts to USD 888.8366194 before display rounding. No invoice, subscription allowance, or charged premium-request total was measured. Tool-service, CI, storage, and other non-model charges are not included.
Daily recorded usage
UTC date
Requests
Priced-channel tokens
USD equivalent
2026-09-18
392
62,824,811
$143.83
2026-09-19
574
148,568,070
$441.38
2026-09-20
198
88,705,894
$303.63
UTC date of the recorded usage event, not inferred human workdays. Request-active wall time is not summed across models or agents.
Scope, reconciliation, and privacy
Included event timestamps
2026-09-18T17:27:23.164Z to 2026-09-20T19:36:13.463Z
Request-interval span
2026-09-18T17:27:17.907Z to 2026-09-20T19:36:13.463Z; 180,535,556 ms
1,163 rows differ from the priced token-channel vector. This can reflect different channel definitions; it is not a count of failed requests.
The collector uses an exact working-directory match and a read-only SQLite transaction. It passes only pre-cutoff usage metadata to the upstream strict parser in memory. Missing, unsupported, or unreconciled critical values stop capture; the tool does not estimate them from prompt length or substitute flat token columns.
No prompt text, responses, turn labels, paths, machine names, raw session/agent IDs, or individual request rows are published. Prompt and response tables are not queried. Model grouping, roles, channels, days, rates, and charge units reconcile before publication.
Provenance and integrity references
Upstream commit: d688a913d298cec5cca722d73c328f756b3dd685; version 0.1.0. The store, metrics, and intervals modules are unmodified; a local import shim and collection/presentation adapters are documented separately.
Time inference: 10-minute default completion-gap cutoff. The displayed model/human decomposition reconciles in milliseconds, not by adding rounded hours.
This is a one-machine, selected-record snapshot, not proof that all project activity or other machines were logged.
The fixed cutoff excludes this refresh and later work, while earlier audit production is included. The snapshot does not refresh from private telemetry in GitHub Pages.
Model identifiers are reported exactly as recorded, not mapped to unverified public product names.
Requests are ledger events, not human messages; cached input is repeated traffic, not unique authored content.
Reasoning metadata is not added to the priced token-channel total; zero does not establish absence of reasoning.
Summed request duration is model work; union time removes overlaps using usage-calc's completion-timestamp interpretation. Neither measures GPU-hours or human attention.
Inferred human time is engaged sitting-union time minus model-active union time at a declared idle threshold. Tool waits, interruptions and unattended automation can inflate it; reading after the last request and human work concurrent with model activity can be missed.
Idle-threshold sensitivity is not a confidence interval. These single-project sittings cannot establish actual presence, other-project work, or a person's total workday.
Recorded charge units reconcile; their USD interpretation follows the upstream conversion assumption and is not an actual bill.
No energy, water, carbon, human labor, or output value is measured by this audit.
Non-model tool services, CI, storage, network charges, and any hidden upstream routing are not measured.
Raw records remain private. Readers can reconcile the published aggregates, but independent confirmation of the underlying ledger requires authorized local access.
Author review, rereading and corrections
Review is substantive human work, not merely prompting. Reading, checking claims, taking responsibility for the content, making corrections and reviewing those corrections are distinct activities. API timing does not measure them.
Git revisions establish the review workload, not whether reading happened. The baseline assumes one contextual reading pass per version of each level-two source section. It counts the latest text once, plus the previous text of each changed or removed section. Unchanged sections, whitespace-only edits and section renumbering add no pass. This is a declared review policy, not a measured minimum of actual human time.
The author reports rereading some sections two or three times. Those reads may already be represented by section revisions, so additional passes default to zero. Add them only for reading beyond the version-based baseline. Only the main report is selected by default.
Current selected text
22,130 words
Superseded section text
27,314 prior-version words
Baseline word exposures
49,444
Current text, one pass
1.55 h
Selected baseline reading
3.46 h
Baseline plus extra rereads
3.46 h
Additional review work
Not quantified; not included in the reading estimate
Estimated author-review effort
3.46 h
Required-review workload assuming one contextual pass per committed section version. Extra passes apply only to current text and must exclude rereads already represented by revisions. This is not observed labor or proof that review occurred.
At 300 to 175 words/minute: 2.75 to 4.71 h. Sensitivity to reading assumptions, not a confidence interval or bound on actual labor.
Revision baseline = current words + superseded/removed section words. Reading minutes = (baseline word exposures + current words × extra-reread fraction × extra passes) ÷ words per minute. Add only separately entered, non-overlapping checking/editing minutes.
Reference: Brysbaert (2019), review and meta-analysis of reading rate. The 238 words/minute reference and 175-300 sensitivity range concern adult English silent nonfiction reading, not technical review or fact-checking productivity.
Counts cover selected Markdown source prose, headings, tables and captions through the audit's material commit, not today's rendered pages. Bibliography, citation markers, image pixels and code blocks are excluded. Generated-only financial tables, charts, layout, code, external sources, uncommitted drafts and later changes are not counted. Mechanical wording edits are included when section text changes. A section title change is recorded as removal/addition, retaining one pass per version rather than a whole-document reread per commit.
Traceable section-revision evidence
Frozen material: 6ed823c. Public Git history capture: 2026-09-20T21:22:26.268Z. The evidence table includes all optional documents; the estimate above uses only checked documents. Quoted historical section titles retain their original wording, including the former project name.
Title and preamble (changed): 90 prior words counted; 103 next-version words.
989090c (2026-09-20): 2,299 current words, 1 added sections, 0 changed/removed sections.
No superseded section words added.
5e270c0 (2026-09-20): 2,478 current words, 3 added sections, 0 changed/removed sections.
No superseded section words added.
Startup funding analysis: 1 committed revisions
989090c (2026-09-20): 1,457 current words, 8 added sections, 0 changed/removed sections.
No superseded section words added.
Do not add this estimate to the interaction proxy or model elapsed time. Review can happen during model calls, inside the non-model residual, or outside recorded sittings. The quantities overlap and have different scopes. Actual prompting and actual review time remain unmeasured without an author time log. A word-based estimate does not certify that review occurred, that the content is correct, or that responsibility has been discharged.
