Seven Billion Tokens, Dissected.
One business intent, seven months, eleven agent harnesses — and a published number that turned out to be wrong in three directions at once.
For anyone who wants to do business with AI. This is the dissection of one real case: the same operator, the same amateur-radio ecosystem, seven months (2026-03 → 2026-09, plus a 2025-10 → 2026-04 control period), seven repositories, eleven agent harnesses that record usage, and a dozen models. It answers one through-line: how a business intent becomes a delivered product, how many round-trips the process took and why it kept getting smoother, and what the numbers behind that story can actually support. Every figure carries a source and a cutoff; fact and inference are labelled separately; the ledger tables and the measurement boundaries live in the companion volumes so this page can stay a narrative.
Five questions this piece answers
- FactCan one ledger be inflated and deflated at the same time?
- FactIs the number on the ledger an asset?
- FactIs the efficiency curve real?
- InferenceWhat does the model decide, and what does the method decide?
- ThesisWhich question should come first?
The ledger that shrank
Before any argument about method, one lesson about measurement — learned the hard way.
On 2026-09-05 this site published a token census: 7,007,437,567 recorded tokens. One week later the same measurement was re-run and returned 7,404,583,808 — 397,146,241 more. The interesting part is not the growth. It is that the old number was wrong in three separate ways at the same time.
| Direction | What happened | Evidence |
|---|---|---|
| Overstated | claude-code's log writes one usage record per chunk of the same API response, so summing every line inflates the total by roughly 2.7×. Deduplicating by message.id leaves 818,520,401 — and brings the whole ledger to 5,695,506,675. | 6,600 distinct API responses |
| Understated | Hermes was published as "not recorded". It records fine — in a sessions table holding 485,213,348 tokens. The earlier check queried messages.token_count, which is all zeros. | 292 sessions, 2026-05-15 → 07-24 |
| Drifting | claude-code deleted its own older logs at 2026-09-07T13:13:57Z, so the same harness measured 679,899,086 tokens less than a week before. Then a second run of the identical script, twelve minutes later, differed by 51,427,037 — because the session performing the measurement was itself writing to the logs being measured. | Observer effect, reproduced twice |
The counter-check matters as much as the finding. codex has been dormant since late August, and its figure came back bit-for-bit identical at 547,796,667 — measured twice, by two different code paths. A stable measurement and an unstable one are both information; only the stable one deserves to be called a total.
All figures on this page are a snapshot taken at 2026-09-12T02:40:59Z across eleven harnesses. The full per-harness table, each harness's own measurement window, and the rules for what may and may not be compared are in the companion volume The Ledger and Its Measurement Boundaries. Recorded and estimated ledgers must not be merged: iFlow's roughly 247 million tokens are estimated from character counts because its usage fields are zero, and stay outside every total here.
Evidence must be snapshotted, never accumulated. A number without a cutoff time is not a measurement — it is a rumour with decimal places.
Intent never starts with "AI"
Every product in this ecosystem began as a sentence in business language — and the sentence still stands with the word "AI" removed.
Four business intents, none of them mentions AI
AI was brought in later as a means of execution. It was never the purpose.
MRRC · 2026-03-06 · ba66892
Hams needed a universal remote station interface — drive any Hamlib-capable radio from a browser, without rewriting the control logic for each radio. Save effort.
SunMRRC / SunsdrMobile · 2026-06-21 · 38cb85e
SunSDR2 DX users needed to bypass the desktop software and get synchronized control, spectrum, audio, and deterministic transmit safety. Safety.
MRRC FT-710 · 2026-07-06 · 9403e2e
The FT-710 already had a USB cable plugged into it — use that cable instead of the SCU-LAN10 accessory for remote operation. (Proven later: the in-cable FT4222 bridge yields ~21 fps of real spectrum, 48 kHz bidirectional Opus audio, and full CAT control.) Save money.
MRRC-FT8 · 2026-08-03
FT8 is not a single CAT command — it is a complete workflow of 15-second cycle synchronization, decode selection, a QSO state machine, and controlled transmit timing. Save time.
First rule: if removing the word "AI" kills your business need, it was never a business need — it was a technology demo.
Six months with nowhere for knowledge to land
The models were not weak. What was weak: knowledge had no carrier.
From 2025-10-20 to 2026-04-16, the same person ran 185 iFlow sessions and 37,202 turns across six model backends (qwen3-coder-plus, glm-5 — 10,420 turns in a single week — kimi-k2.5, minimax-m2.5). The output was real: MRRC's ancestor was born in this period. But nothing about the work settled anywhere.
Every one of the 37,202 assistant turns records zero usage, and no engineering artifact survived the era: no contract, no document system, no gate. When a session ended, what it had learned ended with it.
That is the problem every mechanism in Volume III exists to solve.
From round-trips to right-first-time
"It got smoother" is a feeling; quantified, it is evidence. Three metrics, all reproducible from git history and local logs.
Human turns per delivered commit
How many conversational round-trips a human must put in for each commit that lands.
| Project | Exploration period | Smooth period | Change |
|---|---|---|---|
| ft8 | W31: 17.1 turns/commit (7 commits burned 120 turns + 530M tokens) | W32–34: 0.3–0.7 turns/commit | ~30× |
| ft710 | W29–30: 4.6–4.9 turns/commit | W33–34: 0.7 turns/commit | ~6× |
| website | W34: 1.6 turns/commit | W35–36: 0.4 turns/commit | ~4× |
| mrrc_modern | — | last week: 0.73 turns/commit (30 human turns over 41 commits) | held |
| MRRC | — | last week: 6.00 turns/commit (12 human turns over 2 commits) | reversed |
Three independent curves fall in the same direction — ft8's reads 17.1 → 0.3 turns per commit — and the low number held for a fourth week on mrrc_modern. What "<1" means: a one-sentence requirement, the agent autonomously runs through multiple commits, and the human only supplies intent and judgment. MRRC's last week shows the other side of the same rule — two commits that were still settling their direction cost 6.00 turns each.
Tokens per commit
The meter reads uncertainty, not productivity.
| Project | Exploration period | Maintenance period | Gap |
|---|---|---|---|
| ft710 | 15.2 M/commit (W30) | 0.95 M/commit (W34) | 16× |
| ft8 | 75.5 M/commit (W31) | 0.9 M/commit (W36) | 84× |
| sunsdr | 1.9 M/commit (W25) | 0.4 M/commit (W27) | 5× |
| mrrc_modern | — | last week: 5.95 M/commit (244,037,840 over 41 commits) | rising |
| MRRC | — | last week: 47.07 M/commit (94,144,081 over 2 commits) | rising |
High-cost weeks are exploration with the direction undecided; low-cost weeks are delivery after the pattern was established. Last week is the proof rather than the exception: mrrc_modern's unit cost climbed back from 0.95 M to 5.95 M precisely because it took on four new feature releases in five days, while MRRC's two direction-setting commits cost 47.07 M each. Token peaks always land where uncertainty is highest — the figure is an uncertainty meter, not a productivity trophy.
Not a monotonic decline — an arc
Rework concentrates in exploration and integration weeks; maintenance weeks sit in single digits.
ft710's fix-commit share fell from 36% in its opening week to 6–12% in maintenance weeks; ft8's fell from 29% to 17% — achieved during its two highest-volume weeks (147 / 82 commits). The outlier is on record too: one ft710 commit in W32 burned 410M tokens as an exploration spike (analysis doc, 2026-09-06).
The counter-evidence is just as real, and it arrived this week. MRRC hit 50% fix share in its big integration week (W32, 798M tokens); modern hit 31% in its big refactor week (W35); sunsdr's early hardware-tuning weeks hit 51%. Last week mrrc_modern closed 41 commits of which 17 were fixes — a 41% fix share, its highest since the refactor. New requirements always bring new rework. The difference is who does the rework: humans intervened only 0.73 turns per commit, so the agent completed nearly all of it autonomously.
The workflow takes shape
Contract red first, agent delivers to green.
Release cadence kept accelerating: this week mrrc_modern shipped v1.14.0 through v1.14.3 — four releases in five days, with the version's own changelog corrected the same day it was published. Earlier: modern shipped 6 versions in 13 days (3 in a single day); ft8 shipped 3 versions in its first 3 days; MRRC once shipped 9 versions in two days.
Since 2026-W35 the website repo shows a stable "contract test goes red first, feature commit turns it green" pattern: 5 test→feat pairs, with a median time-to-green of 19 minutes. This article's own series is being built the same way.
Write the contract first and let AI deliver to green — an institutionalized capability that only formed late, and the organizational shape of "smooth".
What the harness and the SDD did at each stage
The method's growth is precisely datable in git history, and its hardest evidence is a chain that starts with a field incident.
How the method grew
Document era, then contract era: knowledge stops needing to be remembered and appears when it should.
| Date | Event | What changed |
|---|---|---|
| 2026-05-02 / 05-10 | MRRC introduces the SDD framework (58aa675); the FDE practice guide follows (88f519f) | Document era — knowledge has a carrier, but it goes stale and gets skipped |
| 2026-07-06 | ft710 rewrites its SDD from day one (87297d1) | Contract first |
| 2026-07-20 | sdd-guardian is born (2498ec2): constraint registry + runtime gates — constraints injected at session start (SessionStart → prime), edits intercepted before they land on disk (PreToolUse → hook) | Contract era — knowledge needs no remembering |
| 2026-08-02 | ft8's first substantive commit (d4a7a32) brings AGENTS.md, a 15-chapter SDD, the guardian and contract tests in one shot | The method is mature enough to inherit wholesale — a new project has gates on day one |
| 2026-08-26 | modern's constraint registry grows to 21 rules (2bc3d30) | As of 2026-09-12 the three gated repos hold 52 machine-readable constraints (ft710 17 / modern 21 / ft8 14), each with severity, scope, match patterns and a citation to its origin |
| 2026-09-05 → 09-12 | The registry count does not change | An honest plateau. A week of heavy delivery added no new constraint rows — either the gates held, or a class of incident is going unregistered. Both readings are on the table; the ledger cannot yet tell them apart. |
The division of labour: the outer harness defines business, technical and product constraints — why the work exists, where the boundaries are, what evidence can change a conclusion. The inner loop — Specify → Implement → Test → Review → Observe → Update SDD — is delegated to agents wholesale. The living SDD lets design intent flow downward and field evidence flow back up, with explicit lifecycle states and no automatic promotion.
The hardest evidence that the method works
Not one of the 52 constraints was designed cleverly in advance; every one traces back to a concrete incident.
FT-710 frequency drift · 13-day loop
DN; is not a DNR query but a VFO step command — polling it every 2 seconds drifted a live radio's frequency by ~20 Hz per call. Field incident (fixed 07-07, 72fd6f0) → design ruling AD-014 written into the SDD (07-08, bb128ff) → machine-readable constraint cat-no-dn (07-20, 2498ec2) → pre-edit interception, reproducible (exit code 2).
Deploy backups fill the disk · 7-day loop
08-29: server /var/tmp at 100% (about 300 MB per backup) → 09-05: sister fixes land in four repositories on the same day — and, following the thread, the hidden defect that portal backups had never worked (an unquoted remote heredoc, 067f565) → the rules are promoted to the four deployment iron rules.
$ python3 .agents/skills/sdd-guardian/harness/sdd_context.py check probe_b4_tmp.py
SDD-GUARDIAN: blocking violations found:
[BLOCK] cat-no-dn (AD-014; SDD V1.2 freq-drift incident) probe_b4_tmp.py:2: return c.query("DN;")
$ echo $?
2The incident conversion rate — the time from incident to permanent immunity — is the true measure of an organization's learning ability. This ecosystem runs at 13 days and 7 days. A registry that stopped growing is not automatically good news: it is a question, and the answer is in whether new incident rows appear when the next incident does.
Two ledgers that never vouch for each other
Test suites were run, not read, on census day 2026-09-12.
FT-710: 439 passed. mrrc_modern: 724 passed, up from 682 a week earlier. ft8: 935 passed, 1 skipped, 1 xfailed — previously published as "937 collected", and collected is not passed; the wording had to change with the evidence. Note also that all three product repositories moved to their own virtualenvs (Python 3.13): running the suites under the system Python 3.9 now fails at collection time on X | None annotations. That is an environment limit, not a product defect, and the boundary is written down rather than quietly dropped.
Ledger A · Product maturity
- design target
- simulation
- automated test
- bench verified
- field verified
- released / operational
- known issue (the honest rung)
Ledger B · Process evidence
- registry entry exists
- spec / plan trail
- enforced inside sessions
- interception reproduced
The two ledgers never intersect: a pretty process ledger never vouches for a product claim. The open boundaries stay open in public: FT710Mobile has an unresolved P0 PTT safety issue, and the EFHW tuner's claim status stays at design-target until PCB fabrication and bench verification.
You can delegate the doing; you can never delegate the "does it count". Responsibility cannot be delegated.
The model is not the long-term variable
Seven months, six eras, and one week that made the whole question smaller. The almanac lives in its own volume.
The main force changed inside a week
Measured from local logs, not vendor announcements.
On 2026-09-05 this ecosystem's claude-code workload was dominated by one value-tier model. Seven days later the same harness is dominated by a different one: qwen3.8-max-0902 at 3,883 calls, with the previous leader still present but no longer driving. On the pi side, a model that did not appear in the last census at all — glm-5.3-flash — is already at 745 calls, and a vision variant has appeared in the logs.
Inside the previous "flash hegemony" there was constant challenger rotation, and the rotation did not stop — it accelerated. The model market turns over every three to four weeks. Any method that hard-codes a model name is buying a depreciating asset; the selection mechanism, not the choice, is the durable part.
The measured division of labour still holds — flagships buy direction, value models buy execution, long-context variants buy the specials, harnesses buy the way of working — and the full era-by-era record, the task × model × harness matching table, the public benchmarks (independent measurements and vendor claims kept in separate layers) and the routing-label trap are all in the companion volume The Model × Harness Almanac.
Where the efficiency actually came from
Two variables moved at once for seven months. Here is the evidence that separates them.
Stronger models, or a thicker method?
Both variables changed over the seven months; four pieces of evidence separate them.
1 · Same model, ±method: an order of magnitude
Flash exploring ft8 in W32 cost 75.5 M per commit; the same model in maintenance cost 0.9 M per commit. The model did not change — the contract thickness did.
2 · Strong models, no method: still zero
iFlow-era glm-5 produced 10,420 turns in a single week, yet the 37,202 turns left zero sediment. With method absent, model strength is beside the point.
3 · The method's birthday is precisely visible
The contract package grew in ft710 on 07-20 and was inherited wholesale by ft8 on 08-02 — and ft8's intervention turns fell from 17.1 straight to 0.3.
4 · Model upgrades widen the delegation radius
The task granularity one turn can reliably deliver grew from "change this function" to "deliver this feature against this spec and pass the contract tests". Models decide what can be delegated; method decides whether what was delegated comes back.
The attribution ratio cannot be separated precisely from this data, but the direction is clear: model capability sets the ceiling, and engineering method decides how much of that ceiling you actually capture. Of the two variables, only one is under your control.
Two tracks
The engineering track is yours; the model track is the market's. The compound interest is on yours.
The engineering-capability track (you control it — the compound interest is here):
Snapshot before you quote
Any figure you intend to act on gets a cutoff timestamp and a written window. This article's ledger is stamped
2026-09-12T02:40:59Z, and every harness carries its own window, because one of them deletes its own history.Acceptance before implementation
Write the acceptance conditions before the first line of code — the standard of "does it count" precedes implementation and is the number-one killer of round-trips.
Machine-readable constraints, pre-write gates
The fix is only the first link; ruling → machine-readable constraint → gate is the closed loop (this ecosystem: 13 days / 7 days).
Red first, then green
Contract tests first, agents deliver to green — the median loop in this ecosystem is 19 minutes, and the registry's plateau is watched rather than assumed.
The model-capability track (the market gives it — the strategy is not locking in):
Pick tiers by task, not by faith
Flagships for greenfield and blitz work, the value tier for daily work, long-context variants for specials — in measurement, no model wins every task, and this week the main force changed.
Re-evaluate the market every 3–4 weeks
The timeline above changed main force six times; the selection mechanism — fast trials plus a weekly review — matters more than any single choice.
Beware model names
Routing labels are not model entities, and vendor claims are not independent benchmarks. Your own logs — which task, how many tokens, how many round-trips, what result — are the only benchmark that never lies to you.
The checklists, templates and the case-derived failure library that turn these seven rules into something you can copy on Monday are in the companion volume.
Ask where it settles first
Seven months of lessons fold into one sentence.
First ask "where does my judgment settle", then ask "what can AI do for me".
AI use without a settling ground is consumption — 37,202 turns of conversation can leave nothing behind. AI use with a settling ground is asset accumulation — every incident, every constraint, every spec lowers the cost of the next delegation. And the ledger is the same: a number you snapshot and date is an asset; a number you remember and quote is weather. Models change generation every quarter; the contract thickens every day. The former is weather; the latter is climate. What you can build is the climate.
Separate judgment from execution
The distillation layer. Four dimensions explain everything the ledger showed — and they travel to any business, radio or not.
Four dimensions, four disciplines
The distillation: what the case means once the ledger is closed.
Agentic engineering is not "using AI to write code". It separates judgment from execution: judgment settles into contracts, execution is delegated to agents, and evidence decides whether a contract survives. That one sentence explains why the same operator left evaporated chat logs in the control period and seven repositories with 52 machine-readable constraints in the measured one — the difference was never the model; it was whether judgment had a carrier that outlives the conversation.
1 · Engineering history — the abstraction ladder rises again
Machine code → assembly → high-level languages → frameworks → cloud: every generation delegates one more layer of "how" to the machine. Agentic delegates the whole implementation layer. What remains for the person is three acts: define the intent, mark the boundaries, exercise the verdict. The workload moved accordingly — 12 design specs and 10 implementation plans were written before the code; ft8's first substantive commit brought AGENTS.md, a 15-chapter SDD and contract tests before any business code. Definition now precedes implementation.
2 · Epistemology — the carrier decides the knowledge's lifespan
Chat logs evaporate with the session; documents expire and get skipped; machine-readable constraints with runtime gates do not need to be remembered — they appear when needed. 52 constraints are 52 judgments made permanent, and they survived six harnesses because the carrier is the repository, not a tool. Incident-to-immunity time is the true measure of organizational learning: this ecosystem ran 13 days and 7 days.
3 · Economics — tokens are the price of uncertainty, and cache reads are the rent
Token spend follows uncertainty density, not line count, and the peaks land exactly where direction is undecided. In the three harnesses that report a complete split, cache reads account for 93.7% of all recorded tokens — the standing cost of re-sending context that has not been made machine-sliceable. Spending less is not "use AI less"; it is reducing the uncertainty itself, and shrinking the context that must be re-read.
4 · Responsibility — two ledgers, three non-delegable acts
Product maturity and process evidence are separate books that never vouch for each other: on census day, 439 passing FT-710 tests could not close one open P0 PTT safety issue in the client. Three acts are never delegable — hardware bring-up and bench debugging, RF safety on real radios, and the verdict on whether evidence is sufficient. Not because AI is too weak, but because responsibility cannot be delegated.
1 · Definition before implementation — acceptance criteria, non-goals and boundaries written before the first line of code; intent phrased in business language that stands without the word AI. 2 · Every incident becomes a rule — repair is only the first link; verdict → constraint → pre-write interception closes the loop. 3 · Watch where uncertainty spends — token peaks are a direction radar; audit every important number against two sources, and date the snapshot. 4 · Keep the two ledgers apart — process badges never promote product claims; "do" can be delegated, "did it count" never leaves your hands.
The three volumes that left this page
This article is the causal chain. The queries live next door, each independently citable.
The Ledger and Its Measurement Boundaries
The full eleven-harness census with every measurement window, the dual-caliber explanation, the retention and observer-effect evidence, and the rules for what may and may not be compared. If you only want to check a number, start here.
The Model × Harness Almanac
Six model eras, the task × model × harness matching table, public benchmarks with independent measurements and vendor claims held in separate layers, and the routing-label trap.
The Practitioner's Playbook
The four disciplines expanded into three checklists, three copyable templates, and a failure library where every entry traces to a real incident in this ecosystem.
Related pages: Agentic Engineering; Engineering Mechanism. Product families: MRRC, MRRC Modern, MRRC-FT8, SunMRRC, MRRC-FT8, EFHW.