Seven Billion Tokens.
One Field Incident.
Tokens are not the claim. The claim is that a field fault became a design ruling, a machine-readable constraint, and a reproducible block while the same ledger kept the cost, code scale, and limitations visible.
Seven billion tokens is not a productivity trophy. It is a trace of an engineering system spending context to keep agents inside a contract: business intent enters the harness, the inner loop implements and verifies, the living SDD keeps what was learned, and field evidence decides what is true.
A token count only matters when it is tied to repositories, code volume, evidence, and failure boundaries. Without that map it is vanity accounting. With it, the number shows where the system paid for uncertainty, iteration, context reloads, and correction.
Investigating Things
Go to the thing itself: session stores, tests, code, radio behavior, and the places where the first measurement failed.
Start with the bill, then demote it
The bill is useful because it can be checked. It is not useful because it is large.
Recorded volume
7,007,437,567tokens across seven harnesses, 41,257 turns, and 882 sessions.
Cache-read share
95.2%6,670,776,092 tokens were context being reloaded, not new prose being produced.
Fresh work
4.8%336,661,475 fresh input plus output; model output was 30,061,161 tokens.
Priced sample
$9.30 / BPi recorded $15.11 for 1,625,562,962 tokens.
| Harness | Tokens | Turns | Sessions | Source |
|---|---|---|---|---|
| claude-code | 3,129,411,481 | 21,648 | 194 | ~/.claude/projects/**/*.jsonl |
| pi | 1,625,562,962 | 8,719 | 60 | usage.totalTokens plus cost |
| kimi-code | 1,201,837,537 | 9,769 | 229 | usage.record |
| codex | 547,796,667 | 717 | 116 | JSONL cross-checked with threads.tokens_used |
| mulerun | 471,558,325 | 181 | 181 | session.tokens_* |
| cursor | 27,079,705 | 217 | 96 | non-zero bubbleId:* records |
| opencode | 4,190,890 | 6 | 6 | opencode.db |
Seven billion is a floor
The honest total has two columns: recorded and estimated. They must not be merged.
iFlow
37,202 assistant turns carry usage fields, but all report zero. The transcript body estimates to 234,247,399 tokens, range 187M-281M, and stays outside the recorded total.
Hermes
292 sessions, 9,121 messages and 4,336 tool calls are recorded, but every token_count is zero.
Cursor/Qoder
Cursor records 27,079,705 tokens but no working directory. Qoder is a strict duplicate subset and would double-count 1,500,804 tokens.
Retention
Claude Code starts at 2026-08-05, while older repository work predates that retained window.
With estimates included, the floor becomes 7,241,684,966 tokens, 87,580 interactions and 1,359 sessions. It is still a floor because Hermes contributes interactions but no recorded usage.
The first number was wrong by 1.6B
致知在格物, or in plain engineering: knowledge extends by investigating things, not by trusting labels.
The first pass reported 5,380,941,148 tokens. It under-counted because I searched only the stores I already knew. It over-counted because I added fields whose semantics were not siblings.
Codex correction
cached_input_tokens is a subset of input_tokens, not an additive field. The corrected transcript total is 547,796,667; the independent database sum is 546,858,767.
Pi correction
totalTokens already equals input + output + cacheRead + cacheWrite. Summing all exposed fields inflated the value by roughly its own size.
A number not reconciled against a second source is a hypothesis about a number. Disagreement between sources is not noise; it is the finding.
One cable already on the bench
The starting intent was not "build an AI project". It was whether the FT-710 could be remotely operated over the USB cable already plugged into it.
Field observation
The FT-710 exposes an FT4222 SPI bridge that yields real spectrum data at roughly 21 fps.
Media path
The same cable carries bidirectional Opus audio at 48 kHz and full CAT control.
Product boundary
No SCU-LAN10, no Hamlib dependency for this path, and no second device between operator and radio.
FT-710 reached 439 automated tests: Ran 439 tests in 10.675s / OK. Source baseline is v1.8.1; packaged Windows installer is v1.8.0.
Six months without a harness
The same person, radios and field problems produced a different kind of output before there was somewhere for knowledge to land.
iFlow covers 2025-10-20 to 2026-04-16: 185 sessions, 37,202 assistant turns, and six model backends. Useful work came out of it, including MRRC ancestors and PSK Reporter analysis, but there was no AGENTS.md, no versioned SDD, no machine-readable constraint registry, and no gate.
Wang Yangming's investigating the bamboo story is useful here as a warning about method: enormous sincerity applied to a procedure with no feedback path does not converge. That is the shape of pre-harness work.
Extending Knowledge
A field observation becomes a contract, a contract becomes a gate, and many gates become a reusable ontology.
Harness outside, execution inside
This is the mechanism page reduced to the parts this incident uses.
Harness
Business, technical and product constraints define why the work exists, which interfaces and safety boundaries matter, and what evidence can change a claim.
Loop
Echo -> Delta -> Product contains the inner loop: Specify -> Implement -> Test -> Review -> Deploy / Observe -> Update SDD.
Living SDD
The SDD links intent to code, tests, deployment, known issues and reusable decisions. Runtime prompts load the contract, not a stale copy.
The inner loop is the part executed by agents with real repositories and gates. Intent, architecture boundaries and the judgment of sufficient evidence stay human.
A field fault became an executable rule
This is 物格而后知至: things investigated, then knowledge arrives, and then the knowledge has to block a bad edit.
Incident
On the FT-710, DN; is not a DNR query. It steps the active VFO down by about 20 Hz per call, so polling it caused live frequency drift.
Ruling
The incident became AD-014: DNR level is intentionally not polled.
$ 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 chain is field incident -> SDD ruling -> machine-readable constraint -> non-zero pre-edit block in mrrc_ft710. Every link is an artifact and the last link is reproducible by command.
Token volume does not follow LOC
The useful question is not "how much AI was used?" It is where the loop had to keep reloading context to resolve uncertainty.
| Repository / bucket | Tokens | Tracked source lines | Tokens / line | Engineering reading |
|---|---|---|---|---|
| ft8 | 1,875,258,566 | 69,972 | 26,800 | Small codebase, high risk density: DSP, slot timing and TX guards. |
| mrrc_ft710 | 1,275,611,389 | 728,632 | 1,751 | Large tracked native/mobile tree; vertical slice produced the reusable contract. |
| mrrc_modern | 1,260,804,213 | 744,195 | 1,694 | Platform extraction and three radio profiles. |
| MRRC | 1,150,197,250 | 150,345 | 7,650 | Universal family; longest history, partly truncated by retention. |
| website | 819,343,926 | 35,816 | 22,876 | Evidence reconciliation, bilingual structure and cross-site consistency. |
| sunsdr | 64,490,898 | 38,361 | 1,681 | Station server and client website; much protocol learning predates retained logs. |
| pskreporter | 54,409,498 | not available | n/a | No standalone local checkout for reproducible line count. |
| efhw-knowledge | 45,949,584 | 22,425 | 2,049 | Reading, synthesis and ontology extraction. |
| wfview | 23,459,822 | 172,509 | 136 | Reference tree with small local agent spend. |
| unattributed bucket | 410,832,716 | n/a | n/a | Parent-directory sessions plus non-radio work. |
| working-directory total | 6,980,357,862 | 1,962,255 counted | mixed | radio-ecosystem repositories alone: 6,491,655,826 |
Zhu Xi's phrase sudden thorough comprehension is a useful analogy for this accumulation: one thing today, another tomorrow, then consolidation. The ontology came after repeated evidence across product boundaries, not before.
Tokens do not measure productivity and source lines do not measure difficulty. Their ratio is a diagnostic for uncertainty, context reload, evidence alignment and safety-boundary work.
Unity of Knowing and Acting
Knowing that does not act is not yet knowing; acting without evidence is not knowledge either.
Shipping is evidence work
The suites were run, not read, on 2026-09-05.
| Family | Command | Result |
|---|---|---|
| FT-710 | venv/bin/python -m unittest discover -s tests | Ran 439 tests / OK |
| Modern | venv/bin/python -m unittest discover -s tests | Ran 682 tests / FAILED (errors=1); Windows-only launcher import on macOS. |
| FT8 | venv/bin/python -m pytest tests -q --collect-only | 937 tests collected, collected not passed. |
Deployment rules also came from defects: quoted remote heredocs, explicit remote deletes because tar -x never deletes, lean per-site backups, and rollback hints scoped to one site directory.
The ledger also goes stale
The system has to check its own evidence, not just product code.
Landing page drift
V1.0 · 180+ tests -> v1.8.1 · 439 tests; v1.10.1 · 593 tests -> v1.12.1 · 682 tests; 40 tests -> 937 collected.
Evidence ledger drift
v1.12.0 · 633 tests -> v1.12.1 · 682 tests.
Constraint metadata drift
Registry versions lag the SDDs: V1.7 vs v1.8.0, V2.27 vs V2.30, V1.0 vs V1.8.
Version ambiguity
FT-710 source is v1.8.1 while the Windows package is v1.8.0. One version field would be wrong.
Precision is what survived. A dated evidence record and a current-state claim are different artifacts even when they contain the same number.
What was never delegated
Agents carry execution. Humans retain intent, boundaries and judgment.
Non-compressible gates
Hardware fabrication and bench bring-up, RF safety, live station verification, and publication labels cannot be bought with more tokens.
Open defect
FT710Mobile keeps an unresolved P0 PTT safety issue open. 439 server-side tests do not close a client safety claim.
Product maturity
- design target
- simulation result
- automated test
- bench verified
- field verified
- released / operational
- known issue
Process constraint
- B1 registry entry
- B2 spec / plan trail
- B3 enforced in session
- B4 block verified
A B4 process badge never promotes a product claim. This is why knowing and not acting is not enough, but acting is not knowing either.
What transfers outside radios
The method is domain-independent: constrain, execute, verify, record, and keep product maturity separate from process evidence.
Define success first
Write acceptance conditions and explicit non-goals before implementation.
Make constraints machine-readable
Each rule needs severity, scope, patterns and a citation to the decision or incident that produced it.
Load the live contract
Reference the SDD and slice it at runtime. Do not copy it into a stale prompt.
Gate before write
In
mrrc_ft710, a pre-edit gate prevents the defect; a later review depends on someone noticing.Turn incidents into rules
The frequency drift bug became AD-014, then
cat-no-dn, then a blocked edit.Reconcile every number
The first census was wrong by 1.6B while another database on the same disk held the correction.
Keep two ladders
Process evidence and product maturity are different claims and must stay separate.
Every number has a boundary
Census date for all rows: 2026-09-05 unless stated.
| Class | Evidence |
|---|---|
| Token ledger | claude-code 3,129,411,481 / 21,648 / 194; pi 1,625,562,962 / 8,719 / 60; kimi-code 1,201,837,537 / 9,769 / 229; codex 547,796,667 / 717 / 116; mulerun 471,558,325 / 181; cursor 27,079,705; opencode 4,190,890. Total 7,007,437,567 / 41,257 / 882. |
| Excluded or estimated | iflow estimate 234,247,399 tokens, range 187M-281M; all usage fields recorded zero. Hermes: 292 sessions, 9,121 messages, 4,336 tool calls, no token usage. Qoder excluded as duplicate subset. |
| Tests and constraints | FT-710 439 tests OK; Modern 682 tests with one macOS environment error; FT8 937 tests collected. Constraint registries: 17, 21 and 14 rules; total 52. |
| Lineage | mrrc_modern and mrrc_ft710 share initial commit 9403e2e. mrrc_ft8 introduced AGENTS.md, SDD and guardian skill in first commit d4a7a32. Registry growth includes 2bc3d30. |
| Method limits | Attribution uses each tool's recorded working directory. Source-line counts come from local git ls-files at each git root, with common text/source extensions and generated dependency directories excluded. Model identifiers are as recorded by each harness. |
| Classical sources | Great Learning: "致知在格物，物格而后知至". Zhu Xi: "一旦豁然贯通焉". Wang Yangming / Chuanxi lu: investigating the bamboo and "知而不行，只是未知". These are analogies and framings, not evidence that the method works. |
| Related pages | Agentic Engineering; Engineering Mechanism; product families: MRRC Universal, MRRC FT-710, MRRC Modern, SunMRRC, SunsdrMobile, MRRC-FT8, EFHW. |