Loading page
What Scar has actually cost, measured from real sessions rather than modelled from file sizes — and the return it would have to have produced to justify that. Spend is counted. Savings are never claimed. Every figure is available for the whole lifetime and per Scar version, because a lifetime average cannot show a trend.
Every session the ledger has ever priced, and it keeps accumulating. This is the bill. The version selector above holds the same counters partitioned by release, which is the only way to see whether the bill is getting cheaper per unit of work.
Kernel loads have been priced since 2026-08-16. Sessions before that date paid the same standing cost unmetered, so the kernel share of any total spanning them is understated.
The ledger prices 36.2% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
18,888
tokens of rework, per intervention that helped
4,721,187 spent, carried by 250.0 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 18,888 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 250.0 helps inside the priced window, not the 623 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 93 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
The same measured counters, partitioned by the version they were spent under — not a second estimate of them. Read the rates across rows, never the totals: a version that ran for three weeks spent more than one that ran for three days, and that difference is not efficiency. A version boundary is declared by hand, because nothing about a knowledge base derives one.
Lower is better: it is the payload a single scar_recall puts in a context window, which is the half of the bill a payload change can move.
| version | spent | recalls | tok/recall | wasted | break-even |
|---|---|---|---|---|---|
| 0.01 | 174,051 | 100 | 1,116 | 4% | 39,588 |
| 0.02 | 80,925 | 16 | 1,095 | 25% | 40,463 |
| 0.03 | 53,200 | 16 | 1,344 | 21% | 11,007 |
| 0.04 | 95,870 | 25 | 1,292 | 9% | 40,920 |
| 0.05 | 129,612 | 46 | 1,229 | 23% | 9,898 |
| 0.06 | 101,017 | 36 | 1,301 | 12% | 15,950 |
| 0.07 | 54,582 | 18 | 1,438 | 9% | 27,291 |
| 0.08 | 55,314 | 24 | 1,275 | 8% | 9,907 |
| 0.09 | 107,411 | 52 | 1,085 | 25% | 20,789 |
| 0.1 | 47,449 | 19 | 1,200 | 39% | 15,816 |
| 0.11 | 94,463 | 17 | 1,206 | 16% | 26,989 |
| 0.12 | 85,006 | 47 | 1,219 | 8% | 42,431 |
| 0.13 | 224,336 | 83 | 1,324 | 21% | 38,763 |
| 0.14 | 93,203 | 25 | 1,400 | 41% | 32,895 |
| 0.15 | 86,705 | 27 | 1,510 | 14% | 10,510 |
| 0.16 | 229,872 | 63 | 1,514 | 32% | 18,513 |
| 0.17 | 170,881 | 23 | 1,384 | 55% | 31,042 |
| 0.18 | 233,061 | 75 | 1,472 | 33% | 17,868 |
| 0.19 | 357,860 | 98 | 1,563 | 34% | 19,110 |
| 0.20 | 193,829 | 51 | 1,811 | 25% | 14,106 |
| 0.21 | 177,711 | 67 | 1,644 | 36% | 19,930 |
| 0.22 | 58,526 | 18 | 1,730 | 14% | 21,009 |
| 0.23 | 241,332 | 101 | 1,602 | 23% | 17,705 |
| 0.24 | 433,133 | 193 | 1,596 | 25% | 15,694 |
| 0.25 | 705,307 | 263 | 1,305 | 43% | 10,479 |
| 0.26open | 436,531 | 149 | 1,294 | 45% | 12,708 |
Version 0.26. Counters are still moving; every recall and kernel load from here is charged to it. 9% of all spend the ledger has ever priced.
recall noise filter collecting (233); decisive margin on relevance (234); hypothesis firings carry investigation age (237); worker escalation v1 trained (239)
The ledger prices 47.4% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
12,708
tokens of rework, per intervention that helped
436,531 spent, carried by 34.4 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 12,708 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 34.4 helps inside the priced window, not the 55 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 33 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.25. Closed and frozen — these figures cannot move again, which is what makes them comparable. 15% of all spend the ledger has ever priced.
per-delivery attribution; outcome gate re-arms per document and names every id; repeats are not recalls
The ledger prices 53.3% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
10,479
tokens of rework, per intervention that helped
705,307 spent, carried by 67.3 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 10,479 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 67.3 helps inside the priced window, not the 103 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 41 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.24. Closed and frozen — these figures cannot move again, which is what makes them comparable. 9% of all spend the ledger has ever priced.
unread: an honest non-judgement outcome, inert in rank, split out of unattributed
The ledger prices 42.2% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
15,694
tokens of rework, per intervention that helped
433,133 spent, carried by 27.6 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 15,694 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 27.6 helps inside the priced window, not the 59 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 33 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.23. Closed and frozen — these figures cannot move again, which is what makes them comparable. 5% of all spend the ledger has ever priced.
Kernel priced at the prompt, not the window; read trigger re-arms on a task boundary; gates infer their own verdicts; nudge ceiling becomes 4/hour
The ledger prices 34.4% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
17,705
tokens of rework, per intervention that helped
241,332 spent, carried by 13.6 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 17,705 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 13.6 helps inside the priced window, not the 36 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 18 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.22. Closed and frozen — these figures cannot move again, which is what makes them comparable. 1% of all spend the ledger has ever priced.
Storage and measurement rebuild: index v4 stable document keys, a store seam, SQLite durable store with JSON as derived read snapshot, incremental indexing on content hashes, auto-rebuild on stale rea
The ledger prices 25.7% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
21,009
tokens of rework, per intervention that helped
58,526 spent, carried by 2.8 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 21,009 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 2.8 helps inside the priced window, not the 8 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 6 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.21. Closed and frozen — these figures cannot move again, which is what makes them comparable. 4% of all spend the ledger has ever priced.
adaptive-k admission ladder + gate economics (decision 160) open as the after; verification scar 16 (search-shape negatives); parity-defect-reporting lesson
The ledger prices 29.2% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
19,930
tokens of rework, per intervention that helped
177,711 spent, carried by 8.9 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 19,930 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 8.9 helps inside the priced window, not the 33 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 14 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.20. Closed and frozen — these figures cannot move again, which is what makes them comparable. 4% of all spend the ledger has ever priced.
verification scar 15 (a cheap check answers a nearby question); /brain-update skill re-grounded against current scripts, version-cut step added; before_defect forcing gate refused pending a measured r
The ledger prices 31.6% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
14,106
tokens of rework, per intervention that helped
193,829 spent, carried by 13.7 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 14,106 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 13.7 helps inside the priced window, not the 32 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 13 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.19. Closed and frozen — these figures cannot move again, which is what makes them comparable. 8% of all spend the ledger has ever priced.
Outcomes by rank position (130); economics --json flush fix; halt gate installed everywhere
The ledger prices 36.7% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
19,110
tokens of rework, per intervention that helped
357,860 spent, carried by 18.7 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 19,110 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 18.7 helps inside the priced window, not the 47 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. 20 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.18. Closed and frozen — these figures cannot move again, which is what makes them comparable. 5% of all spend the ledger has ever priced.
Outcome coverage computed per document (the 114% figure); headroom guard self-heals MCP payload protection; hardening pass on brain/ — 7 fixed, 2 deferred
The ledger prices 34.7% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
17,868
tokens of rework, per intervention that helped
233,061 spent, carried by 13.0 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 17,868 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 13.0 helps inside the priced window, not the 37 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 13 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.17. Closed and frozen — these figures cannot move again, which is what makes them comparable. 4% of all spend the ledger has ever priced.
Dashboard rebuilt for hierarchy, semantic colour and responsiveness; scanner split by release (forward-only); outcomes read from the release record; harden pass added static-lint, a ui test runner, er
The ledger prices 30.7% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
31,042
tokens of rework, per intervention that helped
170,881 spent, carried by 5.5 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 31,042 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 5.5 helps inside the priced window, not the 17 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 7 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.16. Closed and frozen — these figures cannot move again, which is what makes them comparable. 5% of all spend the ledger has ever priced.
verify --scope lesson; rule->mechanism and rule->gate->project joins; governance report
The ledger prices 29.6% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
18,513
tokens of rework, per intervention that helped
229,872 spent, carried by 12.4 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 18,513 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 12.4 helps inside the priced window, not the 34 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 18 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.15. Closed and frozen — these figures cannot move again, which is what makes them comparable. 2% of all spend the ledger has ever priced.
Direct-run guards on all 8 self-invoking modules, kernel A/B holdout (off by default), selftest coverage for destructive-git + capture, join-signature narrowed off spread syntax, const-read signature
The ledger prices 32.9% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
10,510
tokens of rework, per intervention that helped
86,705 spent, carried by 8.3 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 10,510 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 8.3 helps inside the priced window, not the 18 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 8 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.14. Closed and frozen — these figures cannot move again, which is what makes them comparable. 2% of all spend the ledger has ever priced.
Harness-plan pass: prospective benchmark (70% vs 96% post-hoc), signature mutation-testing half, distilled-005 retired, first two superseded_by expiries, INDEX.md moved to gitignored derived state + n
The ledger prices 40.5% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
32,895
tokens of rework, per intervention that helped
93,203 spent, carried by 2.8 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 32,895 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 2.8 helps inside the priced window, not the 6 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 8 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.13. Closed and frozen — these figures cannot move again, which is what makes them comparable. 5% of all spend the ledger has ever priced.
Split the dismissal channel: signature-scan firings now report 'dismiss' against the signature, not 'noise' against search rank; rank.mjs caps the legacy penalty at the retrieval count
The ledger prices 30.0% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
38,763
tokens of rework, per intervention that helped
224,336 spent, carried by 5.8 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 38,763 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 5.8 helps inside the priced window, not the 23 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 23 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.12. Closed and frozen — these figures cannot move again, which is what makes them comparable. 2% of all spend the ledger has ever priced.
or.use/or.wst become measurable end to end: 0.11 straddled the orientation boundary and now carries a derived footnote saying its columns are a floor, not a measurement. Also six lessons, four newly c
The ledger prices 22.2% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
42,431
tokens of rework, per intervention that helped
85,006 spent, carried by 2.0 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 42,431 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 2.0 helps inside the priced window, not the 9 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 25 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.11. Closed and frozen — these figures cannot move again, which is what makes them comparable. 2% of all spend the ledger has ever priced.
Fixed the signature-scan whole-file-fallback mislabeling (3rd recurrence — the message claimed 'matching what you just wrote' even on a fallback scan), generalized the refusal-gate prompt for reachabi
The ledger prices 47.7% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
26,989
tokens of rework, per intervention that helped
94,463 spent, carried by 3.5 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 26,989 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 3.5 helps inside the priced window, not the 4 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 2 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.1. Closed and frozen — these figures cannot move again, which is what makes them comparable. 1% of all spend the ledger has ever priced.
Closed the seventh recall-gate over-fire recurrence, shipped before-defect marking and gate-lint, and gave verification a 13th scar for self-authored prose.
The ledger prices 41.4% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
15,816
tokens of rework, per intervention that helped
47,449 spent, carried by 3.0 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 15,816 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 3.0 helps inside the priced window, not the 8 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 6 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.09. Closed and frozen — these figures cannot move again, which is what makes them comparable. 2% of all spend the ledger has ever priced.
the scanner stops reporting clean on files it never opened; outcome reporting goes per-document
The ledger prices 48.3% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
20,789
tokens of rework, per intervention that helped
107,411 spent, carried by 5.2 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 20,789 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 5.2 helps inside the priced window, not the 13 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 17 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.08. Closed and frozen — these figures cannot move again, which is what makes them comparable. 1% of all spend the ledger has ever priced.
Hook hot path cached (208ms->9ms), kernel pass 2 retired, mutation targets bound to their write calls, always-loaded duplication removed
The ledger prices 34.5% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
9,907
tokens of rework, per intervention that helped
55,314 spent, carried by 5.6 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 9,907 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 5.6 helps inside the priced window, not the 7 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 7 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.07. Closed and frozen — these figures cannot move again, which is what makes them comparable. 1% of all spend the ledger has ever priced.
write-back reach is measured: the gate parses only structured ids, brain_learn refuses phantoms, and reach is recorded for every armed session
The ledger prices 46.5% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
27,291
tokens of rework, per intervention that helped
54,582 spent, carried by 2.0 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 27,291 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 2.0 helps inside the priced window, not the 5 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 4 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.06. Closed and frozen — these figures cannot move again, which is what makes them comparable. 2% of all spend the ledger has ever priced.
exp/structural-memory: oriented recall + stale-index lease — engine basis changed
The ledger prices 44.0% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
15,950
tokens of rework, per intervention that helped
101,017 spent, carried by 6.3 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 15,950 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 6.3 helps inside the priced window, not the 11 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 12 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.05. Closed and frozen — these figures cannot move again, which is what makes them comparable. 3% of all spend the ledger has ever priced.
gates lead, recall feeds them: enforcement-first positioning, comment-aware signature scan, self-authored recall marking
The ledger prices 44.4% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
9,898
tokens of rework, per intervention that helped
129,612 spent, carried by 13.1 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 9,898 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 13.1 helps inside the priced window, not the 19 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 10 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.04. Closed and frozen — these figures cannot move again, which is what makes them comparable. 2% of all spend the ledger has ever priced.
L1 gains a lifecycle: lessons retire via superseded_by instead of accumulating, and the corpus reports its own overlap, kernel allocation, and payload integrity
The ledger prices 36.1% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
40,920
tokens of rework, per intervention that helped
95,870 spent, carried by 2.3 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 40,920 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 2.3 helps inside the priced window, not the 7 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 7 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.03. Closed and frozen — these figures cannot move again, which is what makes them comparable. 1% of all spend the ledger has ever priced.
one definition of a mutating turn: gates cover shell-routed writes, and read fan-out covers sessions that never write
The ledger prices 40.0% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
11,007
tokens of rework, per intervention that helped
53,200 spent, carried by 4.8 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 11,007 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 4.8 helps inside the priced window, not the 9 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 7 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.02. Closed and frozen — these figures cannot move again, which is what makes them comparable. 2% of all spend the ledger has ever priced.
versioned economics: spend and outcomes now accrue per release, not only lifetime
The ledger prices 50.0% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
40,463
tokens of rework, per intervention that helped
80,925 spent, carried by 2.0 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 40,463 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 2.0 helps inside the priced window, not the 3 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 7 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Version 0.01. Closed and frozen — these figures cannot move again, which is what makes them comparable. 4% of all spend the ledger has ever priced.
This version predates version tracking. Its spend is exact — every token was counted as it was served — but its outcomes are the lifetime feedback counters, because nothing recorded them per version at the time. It holds the whole of Scar’s history before versioning existed, so it matches the Lifetime tab exactly until a later version accrues. They diverge from there.
Kernel loads have been priced since 2026-08-16. Sessions before that date paid the same standing cost unmetered, so the kernel share of any total spanning them is understated.
The ledger prices 19.5% of recorded recalls. Outcome counters predate the token ledger, so most recalls in Scar’s history have outcomes attached but no price. Every absolute figure here is therefore a floor on true spend, not an estimate of it — and the ratios below describe the measured slice only.
A document’s tokens are allocated across its recalls and each recall is sorted by the outcome an agent reported for it. Every share below traces to a counter written at the point of use — none of it is a judgement made at render time.
Why unattributed: not decomposable for this version — it predates per-delivery attribution.
39,588
tokens of rework, per intervention that helped
174,051 spent, carried by 4.4 confirmed helps inside the priced window. For Scar to have cost nothing on net, each of those must have avoided at least 39,588 tokens of rework. Whether it did is a judgement, and this page deliberately leaves it to you rather than making it for you.
The kernel’s standing cost is charged against those interventions too. It is paid on every session, including the ones where nothing fired, and a break-even that excused it would understate the bill.
Everything above is measured. Everything below is not.
Deliberately multiplied by the 4.4 helps inside the priced window, not the 19 recorded over Scar’s lifetime. Setting lifetime outcomes against partial spend is how this page would flatter itself — the first run of the report behind it did exactly that and printed a 141× return. Nothing on this card is stored, and nothing here feeds ranking.
By tokens delivered into real contexts, in this window. A document high on this list that has never earned anything is the clearest promotion-or-deletion signal Scar produces — and a document that drops off it between versions is the clearest sign a payload change worked. Read cost and outcome together: the same “earned” label covers a rule at 49 helps against 7 noise and a lesson at 3 against 28, and only the counts separate them. This is analysis, not a ranking signal — cost never feeds retrieval, for the reason under the table. This window closed before the per-document counters existed, so its help/noise and per-call columns are blank rather than zero. A frozen release is not recomputed — that is what keeps its figures from drifting. 17 private-visibility rows are excluded here — their ids name clients.
Measured here, never ranked on. Scar prices its own knowledge layer and separately learns from demonstrated usefulness; the two are analysed together and wired apart. Retrieval ranking reads outcome reports and has no access to cost — otherwise a verbose document would accumulate ranking weight for being long, which is a property of its prose and not of its worth. The ledger reads feedback; feedback never reads the ledger.
Everything above prices what Scar cost. This is the other half: how much of it runs without being read. A lesson is wired when it names a grep-able signature that signature-scan executes against code as it is written; the rest is advisory and fires only if an agent happens to read the line. The ceiling here is not 100% — some lessons have no grep-able form and never will, so this is reported and never gated.
That last figure is the only one on this page whose author is not the document’s own author. Each of these lessons is cited by id in a comment of the code it changed, written by whoever made the change — so unlike a confidence label or a recurrence count, it cannot be asserted by the lesson itself. A build check refuses any citation that does not resolve, which is what keeps it from decaying into a grep. 45 of the 65 shown are advisory — evidence for lessons no mechanism can produce evidence for. 15 private-visibility ids are withheld — they name clients.
Reported, never ranked. The agent that writes a citation is the agent that used the lesson, so the channel is inflatable in principle — for the same reason nothing in the ledger above feeds retrieval ranking either.