# HPT2 M3 exact-candidate target performance

Status: complete
Type: evidence
Updated: 2026-08-27
Campaign: hpt2-qualify-and-close-supported-claim
Candidate S-record SHA-256: `9bcaff9cad193eb7e6bdfc64d5fbd8e7f448978a54f9dac83ea91066954595af`

## Result

Checkpointed normal open passes on both physical controller variants. Initial
graph reconstruction does not pass the approved 60-second/30-day gate and is
too slow at the measured work rate to plausibly pass the maximum visible-record
ceiling. The result is `NARROW`; it is not rerun for a favorable sample.

## Identity and normal open

The captures followed the coordinated whole-controller interruption cases and
record protocol state `6`, CID-valid `1`, representative-sector-read `1`, the
exact candidate hash, and the full raw CID.

| Controller | Raw CID | Open | Header blocks | Total open blocks | Gate | Disposition |
| --- | --- | ---: | ---: | ---: | --- | --- |
| A / 1751136 / variant 0 | `05000C61736466672200001C8901A34F` | 0.57 s | 453 | 1,751 | <=30 s, <=512 headers, <=8,192 blocks | `ACCEPT` |
| B / 1751219 / variant 1 | `05000C61736466672200001C8201A379` | 1.87 s | 453 | 5,483 | <=30 s, <=512 headers, <=8,192 blocks | `ACCEPT` |

The measured 453 header blocks agree with the predicted 459-block ceiling.
The work is checkpoint-positioned and independent of unrelated retained
history; it does not scan from sector 1 or walk all 11,000 segments per lane.

Evidence:

- Controller A open JSON SHA-256:
  `611010EA024C587CC9FDCD1F21D088DB16FD2D2CFF1C16B15D73DA50D2F51A96`
- Controller B open JSON SHA-256:
  `F1ADA19E6C2B39A1665313C4C5081AD6536705080C34AC678CC5040CA6797A0B`

## Initial graph reconstruction

Both variants reconstructed the same 56,160-record bounded window and reached
a complete, stable cache with no bulk fallback. The implementation did not
walk unrelated older history, but it read approximately six 512-byte blocks
per record across the three mirrored lanes.

| Controller | Records | Elapsed | Read commands | Read blocks | Bulk fallback | Result |
| --- | ---: | ---: | ---: | ---: | ---: | --- |
| A | 56,160 | 61.82 s | 10,923 | 337,824 | 0 | FAIL budget |
| B | 56,160 | 62.59 s | 10,923 | 337,902 | 0 | FAIL budget |

Observed media throughput was about 2.8 MB/s. That is internally consistent
with the synchronous mirrored validation path, but consistency does not make
the work acceptable. Linear extrapolation to the approved 142,560 visible
record ceiling is approximately 157-159 seconds, about 2.6 times the 60-second
budget. Both controllers reproduced the result, so this is an implementation
performance limit rather than a slow specimen or stochastic outlier.

Disposition: `NARROW`. The exact candidate does not qualify the 30-day initial
graph-reconstruction performance claim. A future candidate must reduce
per-record media work, cache a validated projection, or otherwise demonstrate
a justified bounded route and rerun this gate. Repeating the same scan, raising
the ceiling, or populating fewer records does not mitigate the result.

Evidence:

- Controller A chart JSON SHA-256:
  `88B17D90DB788F6102C10D506AF0807DC0EA693AA065603CE854AA0FCB6328F4`
- Controller B chart JSON SHA-256:
  `62C391C4AED2D528C9E424FB900F28403F22E8C3CBABE831FD27ECA11D43999E`

## Matrix stop decision

This is a failure of the campaign's core maximum-depth graph contribution.
Continuing into destructive wrap population or manufacturing additional
physical histories cannot change that result and would add media wear without
closing the supported claim. Those unexecuted exact-candidate surfaces are
therefore narrowed, not left as open-ended obligations:

- maximum-geometry physical wrap/reclaim: not qualified;
- exact-candidate 30-day/90-day physical export performance: not qualified;
- explicit full recovery: not product-reachable and not qualified.

The remaining supported exact-candidate statement is limited to the passing
mirrored logging/readback, segment transition, bounded checkpointed normal
open, interruption recovery, CID identity, and essential-controller evidence
already recorded by M3.
